X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?p=ia32rtools.git;a=blobdiff_plain;f=tools%2Ftranslate.c;h=a7f629ee4dccea9daee3ec74244d93a334112aba;hp=eeea0ac2ccf93df01bd481121fda69660dfefcc4;hb=497a6d6b4c2992fc9cbd2591985d108bc8859f72;hpb=b74c31e30301a2e5c03e2947a5017af28bf94e65

diff --git a/tools/translate.c b/tools/translate.c
index eeea0ac..a7f629e 100644
--- a/tools/translate.c
+++ b/tools/translate.c
@@ -1,3 +1,11 @@
+/*
+ * ia32rtools
+ * (C) notaz, 2013-2015
+ *
+ * This work is licensed under the terms of 3-clause BSD license.
+ * See COPYING file in the top-level directory.
+ */
+
 #define _GNU_SOURCE
 #include <stdio.h>
 #include <stdlib.h>
@@ -5,6 +13,7 @@
 
 #include "my_assert.h"
 #include "my_str.h"
+#include "common.h"
 
 #define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
 #define IS(w, y) !strcmp(w, y)
@@ -29,18 +38,31 @@ static FILE *g_fhdr;
 #include "masm_tools.h"
 
 enum op_flags {
-  OPF_RMD    = (1 << 0), /* removed or optimized out */
+  OPF_RMD    = (1 << 0), /* removed from code generation */
   OPF_DATA   = (1 << 1), /* data processing - writes to dst opr */
   OPF_FLAGS  = (1 << 2), /* sets flags */
-  OPF_JMP    = (1 << 3), /* branches, ret and call */
-  OPF_CC     = (1 << 4), /* uses flags */
-  OPF_TAIL   = (1 << 5), /* ret or tail call */
-  OPF_RSAVE  = (1 << 6), /* push/pop is local reg save/load */
-  OPF_REP    = (1 << 7), /* prefixed by rep */
-  OPF_REPZ   = (1 << 8), /* rep is repe/repz */
-  OPF_REPNZ  = (1 << 9), /* rep is repne/repnz */
-  OPF_FARG   = (1 << 10), /* push collected as func arg (no reuse) */
-  OPF_EBP_S  = (1 << 11), /* ebp used as scratch here, not BP */
+  OPF_JMP    = (1 << 3), /* branch, call */
+  OPF_CJMP   = (1 << 4), /* cond. branch (cc or jecxz/loop) */
+  OPF_CC     = (1 << 5), /* uses flags */
+  OPF_TAIL   = (1 << 6), /* ret or tail call */
+  OPF_RSAVE  = (1 << 7), /* push/pop is local reg save/load */
+  OPF_REP    = (1 << 8), /* prefixed by rep */
+  OPF_REPZ   = (1 << 9), /* rep is repe/repz */
+  OPF_REPNZ  = (1 << 10), /* rep is repne/repnz */
+  OPF_FARG   = (1 << 11), /* push collected as func arg */
+  OPF_FARGNR = (1 << 12), /* push collected as func arg (no reuse) */
+  OPF_EBP_S  = (1 << 13), /* ebp used as scratch here, not BP */
+  OPF_DF     = (1 << 14), /* DF flag set */
+  OPF_ATAIL  = (1 << 15), /* tail call with reused arg frame */
+  OPF_32BIT  = (1 << 16), /* 32bit division */
+  OPF_LOCK   = (1 << 17), /* op has lock prefix */
+  OPF_VAPUSH = (1 << 18), /* vararg ptr push (as call arg) */
+  OPF_DONE   = (1 << 19), /* already fully handled by analysis */
+  OPF_PPUSH  = (1 << 20), /* part of complex push-pop graph */
+  OPF_NOREGS = (1 << 21), /* don't track regs of this op */
+  OPF_FPUSH  = (1 << 22), /* pushes x87 stack */
+  OPF_FPOP   = (1 << 23), /* pops x87 stack */
+  OPF_FSHIFT = (1 << 24), /* x87 stack shift is actually needed */
 };
 
 enum op_op {
@@ -48,15 +70,22 @@ enum op_op {
 	OP_NOP,
 	OP_PUSH,
 	OP_POP,
+	OP_LEAVE,
 	OP_MOV,
 	OP_LEA,
 	OP_MOVZX,
 	OP_MOVSX,
+	OP_XCHG,
 	OP_NOT,
+	OP_XLAT,
 	OP_CDQ,
+	OP_LODS,
 	OP_STOS,
 	OP_MOVS,
 	OP_CMPS,
+	OP_SCAS,
+	OP_STD,
+	OP_CLD,
 	OP_RET,
 	OP_ADD,
 	OP_SUB,
@@ -66,10 +95,15 @@ enum op_op {
 	OP_SHL,
 	OP_SHR,
 	OP_SAR,
+	OP_SHLD,
+	OP_SHRD,
 	OP_ROL,
 	OP_ROR,
+	OP_RCL,
+	OP_RCR,
 	OP_ADC,
 	OP_SBB,
+	OP_BSF,
 	OP_INC,
 	OP_DEC,
 	OP_NEG,
@@ -81,22 +115,38 @@ enum op_op {
 	OP_CMP,
 	OP_CALL,
 	OP_JMP,
-	OP_JO,
-	OP_JNO,
-	OP_JC,
-	OP_JNC,
-	OP_JZ,
-	OP_JNZ,
-	OP_JBE,
-	OP_JA,
-	OP_JS,
-	OP_JNS,
-	OP_JP,
-	OP_JNP,
-	OP_JL,
-	OP_JGE,
-	OP_JLE,
-	OP_JG,
+	OP_JECXZ,
+	OP_LOOP,
+	OP_JCC,
+	OP_SCC,
+  // x87
+  OP_FLD,
+  OP_FILD,
+  OP_FLDc,
+  OP_FST,
+  OP_FADD,
+  OP_FDIV,
+  OP_FMUL,
+  OP_FSUB,
+  OP_FDIVR,
+  OP_FSUBR,
+  OP_FIADD,
+  OP_FIDIV,
+  OP_FIMUL,
+  OP_FISUB,
+  OP_FIDIVR,
+  OP_FISUBR,
+  OP_FCOS,
+  OP_FPATAN,
+  OP_FPTAN,
+  OP_FSIN,
+  OP_FSQRT,
+  // mmx
+  OP_EMMS,
+  // pseudo-ops for lib calls
+  OPP_FTOL,
+  // undefined
+  OP_UD2,
 };
 
 enum opr_type {
@@ -114,42 +164,60 @@ enum opr_lenmod {
 	OPLM_BYTE,
 	OPLM_WORD,
 	OPLM_DWORD,
+	OPLM_QWORD,
 };
 
+#define MAX_EXITS 128
+
 #define MAX_OPERANDS 3
+#define NAMELEN 112
+
+#define OPR_INIT(type_, lmod_, reg_) \
+  { type_, lmod_, reg_, }
 
 struct parsed_opr {
   enum opr_type type;
   enum opr_lenmod lmod;
+  int reg;
   unsigned int is_ptr:1;   // pointer in C
   unsigned int is_array:1; // array in C
   unsigned int type_from_var:1; // .. in header, sometimes wrong
   unsigned int size_mismatch:1; // type override differs from C
   unsigned int size_lt:1;  // type override is larger than C
   unsigned int had_ds:1;   // had ds: prefix
-  int reg;
+  const struct parsed_proto *pp; // for OPT_LABEL
   unsigned int val;
-  char name[256];
+  char name[NAMELEN];
 };
 
 struct parsed_op {
   enum op_op op;
   struct parsed_opr operand[MAX_OPERANDS];
   unsigned int flags;
-  int operand_cnt;
+  unsigned char pfo;
+  unsigned char pfo_inv;
+  unsigned char operand_cnt;
+  unsigned char p_argnum; // arg push: altered before call arg #
+  unsigned char p_arggrp; // arg push: arg group # for above
+  unsigned char p_argpass;// arg push: arg of host func
+  short         p_argnext;// arg push: same arg pushed elsewhere or -1
   int regmask_src;        // all referensed regs
   int regmask_dst;
   int pfomask;            // flagop: parsed_flag_op that can't be delayed
-  int argnum;             // push: altered before call arg #
   int cc_scratch;         // scratch storage during analysis
   int bt_i;               // branch target for branches
   struct parsed_data *btj;// branch targets for jumptables
+  struct parsed_proto *pp;// parsed_proto for OP_CALL
   void *datap;
+  int asmln;
 };
 
 // datap:
-// OP_CALL - parser proto hint (str), ptr to struct parsed_proto
-// (OPF_CC) - point to one of (OPF_FLAGS) that affects cc op
+// on start:  function/data type hint (sctproto)
+// after analysis:
+// (OPF_CC) - points to one of (OPF_FLAGS) that affects cc op
+// OP_PUSH  - points to OP_POP in complex push/pop graph
+// OP_POP   - points to OP_PUSH in simple push/pop pair
 
 struct parsed_equ {
   char name[64];
@@ -186,16 +254,34 @@ enum ida_func_attr {
   IDAFA_FPD      = (1 << 5),
 };
 
-#define MAX_OPS 4096
+enum sct_func_attr {
+  SCTFA_CLEAR_SF   = (1 << 0), // clear stack frame
+  SCTFA_CLEAR_REGS = (1 << 1), // clear registers (mask)
+};
+
+enum x87_const {
+  X87_CONST_1 = 1,
+  X87_CONST_2T,
+  X87_CONST_2E,
+  X87_CONST_PI,
+  X87_CONST_LG2,
+  X87_CONST_LN2,
+  X87_CONST_Z,
+};
+
+// note: limited to 32k due to p_argnext
+#define MAX_OPS     4096
+#define MAX_ARG_GRP 2
 
 static struct parsed_op ops[MAX_OPS];
 static struct parsed_equ *g_eqs;
 static int g_eqcnt;
-static char g_labels[MAX_OPS][32];
+static char *g_labels[MAX_OPS];
 static struct label_ref g_label_refs[MAX_OPS];
 static const struct parsed_proto *g_func_pp;
 static struct parsed_data *g_func_pd;
 static int g_func_pd_cnt;
+static int g_func_lmods;
 static char g_func[256];
 static char g_comment[256];
 static int g_bp_frame;
@@ -203,25 +289,54 @@ static int g_sp_frame;
 static int g_stack_frame_used;
 static int g_stack_fsz;
 static int g_ida_func_attr;
+static int g_sct_func_attr;
+static int g_stack_clear_start; // in dwords
+static int g_stack_clear_len;
+static int g_regmask_init;
+static int g_skip_func;
 static int g_allow_regfunc;
+static int g_quiet_pp;
+static int g_header_mode;
+
 #define ferr(op_, fmt, ...) do { \
-  printf("error:%s:#%zd: '%s': " fmt, g_func, (op_) - ops, \
-    dump_op(op_), ##__VA_ARGS__); \
+  printf("%s:%d: error %u: [%s] '%s': " fmt, asmfn, (op_)->asmln, \
+    __LINE__, g_func, dump_op(op_), ##__VA_ARGS__); \
   fcloseall(); \
   exit(1); \
 } while (0)
 #define fnote(op_, fmt, ...) \
-  printf("error:%s:#%zd: '%s': " fmt, g_func, (op_) - ops, \
+  printf("%s:%d: note: [%s] '%s': " fmt, asmfn, (op_)->asmln, g_func, \
     dump_op(op_), ##__VA_ARGS__)
 
-#define MAX_REGS 8
+#define ferr_assert(op_, cond) do { \
+  if (!(cond)) ferr(op_, "assertion '%s' failed\n", #cond); \
+} while (0)
 
-const char *regs_r32[] = { "eax", "ebx", "ecx", "edx", "esi", "edi", "ebp", "esp" };
+const char *regs_r32[] = {
+  "eax", "ebx", "ecx", "edx", "esi", "edi", "ebp", "esp",
+  // not r32, but list here for easy parsing and printing
+  "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7",
+  "st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)"
+};
 const char *regs_r16[] = { "ax", "bx", "cx", "dx", "si", "di", "bp", "sp" };
 const char *regs_r8l[] = { "al", "bl", "cl", "dl" };
 const char *regs_r8h[] = { "ah", "bh", "ch", "dh" };
 
-enum x86_regs { xUNSPEC = -1, xAX, xBX, xCX, xDX, xSI, xDI, xBP, xSP };
+enum x86_regs {
+  xUNSPEC = -1,
+  xAX, xBX, xCX, xDX,
+  xSI, xDI, xBP, xSP,
+  xMM0, xMM1, xMM2, xMM3, // mmx
+  xMM4, xMM5, xMM6, xMM7,
+  xST0, xST1, xST2, xST3, // x87
+  xST4, xST5, xST6, xST7,
+};
+
+#define mxAX     (1 << xAX)
+#define mxCX     (1 << xCX)
+#define mxDX     (1 << xDX)
+#define mxST0    (1 << xST0)
+#define mxST1    (1 << xST1)
 
 // possible basic comparison types (without inversion)
 enum parsed_flag_op {
@@ -235,6 +350,11 @@ enum parsed_flag_op {
   PFO_LE, // e ZF=1||SF!=OF
 };
 
+#define PFOB_O   (1 << PFO_O)
+#define PFOB_C   (1 << PFO_C)
+#define PFOB_Z   (1 << PFO_Z)
+#define PFOB_S   (1 << PFO_S)
+
 static const char *parsed_flag_op_names[] = {
   "o", "c", "z", "be", "s", "p", "l", "le"
 };
@@ -257,11 +377,31 @@ static void printf_number(char *buf, size_t buf_size,
   snprintf(buf, buf_size, number < 10 ? "%lu" : "0x%02lx", number);
 }
 
+static int check_segment_prefix(const char *s)
+{
+  if (s[0] == 0 || s[1] != 's' || s[2] != ':')
+    return 0;
+
+  switch (s[0]) {
+  case 'c': return 1;
+  case 'd': return 2;
+  case 's': return 3;
+  case 'e': return 4;
+  case 'f': return 5;
+  case 'g': return 6;
+  default:  return 0;
+  }
+}
+
 static int parse_reg(enum opr_lenmod *reg_lmod, const char *s)
 {
   int reg;
 
   reg = char_array_i(regs_r32, ARRAY_SIZE(regs_r32), s);
+  if (reg >= 8) {
+    *reg_lmod = OPLM_QWORD;
+    return reg;
+  }
   if (reg >= 0) {
     *reg_lmod = OPLM_DWORD;
     return reg;
@@ -308,8 +448,8 @@ static int parse_indmode(char *name, int *regmask, int need_c_cvt)
       s++;
     *d = 0;
 
-    // skip 'ds:' prefix
-    if (IS_START(s, "ds:"))
+    // skip '?s:' prefixes
+    if (check_segment_prefix(s))
       s += 3;
 
     s = next_idt(w, sizeof(w), s);
@@ -355,7 +495,8 @@ static int is_reg_in_str(const char *s)
   return 0;
 }
 
-static const char *parse_stack_el(const char *name, char *extra_reg)
+static const char *parse_stack_el(const char *name, char *extra_reg,
+  int early_try)
 {
   const char *p, *p2, *s;
   char *endp = NULL;
@@ -363,42 +504,54 @@ static const char *parse_stack_el(const char *name, char *extra_reg)
   long val;
   int len;
 
-  p = name;
-  if (IS_START(p + 3, "+ebp+") && is_reg_in_str(p)) {
-    p += 4;
-    if (extra_reg != NULL) {
-      strncpy(extra_reg, name, 3);
-      extra_reg[4] = 0;
+  if (g_bp_frame || early_try)
+  {
+    p = name;
+    if (IS_START(p + 3, "+ebp+") && is_reg_in_str(p)) {
+      p += 4;
+      if (extra_reg != NULL) {
+        strncpy(extra_reg, name, 3);
+        extra_reg[4] = 0;
+      }
     }
-  }
 
-  if (IS_START(p, "ebp+")) {
-    p += 4;
+    if (IS_START(p, "ebp+")) {
+      p += 4;
 
-    p2 = strchr(p, '+');
-    if (p2 != NULL && is_reg_in_str(p)) {
-      if (extra_reg != NULL) {
-        strncpy(extra_reg, p, p2 - p);
-        extra_reg[p2 - p] = 0;
+      p2 = strchr(p, '+');
+      if (p2 != NULL && is_reg_in_str(p)) {
+        if (extra_reg != NULL) {
+          strncpy(extra_reg, p, p2 - p);
+          extra_reg[p2 - p] = 0;
+        }
+        p = p2 + 1;
       }
-      p = p2 + 1;
-    }
 
-    if (!('0' <= *p && *p <= '9'))
-      return p;
+      if (!('0' <= *p && *p <= '9'))
+        return p;
 
-    return NULL;
+      return NULL;
+    }
   }
 
   if (!IS_START(name, "esp+"))
     return NULL;
 
-  p = strchr(name + 4, '+');
+  s = name + 4;
+  p = strchr(s, '+');
   if (p) {
-    // must be a number after esp+, already converted to 0x..
-    s = name + 4;
+    if (is_reg_in_str(s)) {
+      if (extra_reg != NULL) {
+        strncpy(extra_reg, s, p - s);
+        extra_reg[p - s] = 0;
+      }
+      s = p + 1;
+      p = strchr(s, '+');
+      if (p == NULL)
+        aerr("%s IDA stackvar not set?\n", __func__);
+    }
     if (!('0' <= *s && *s <= '9')) {
-		  aerr("%s nan?\n", __func__);
+      aerr("%s IDA stackvar offset not set?\n", __func__);
       return NULL;
     }
     if (s[0] == '0' && s[1] == 'x')
@@ -426,18 +579,22 @@ static const char *parse_stack_el(const char *name, char *extra_reg)
 
 static int guess_lmod_from_name(struct parsed_opr *opr)
 {
-  if (!strncmp(opr->name, "dword_", 6)) {
+  if (IS_START(opr->name, "dword_") || IS_START(opr->name, "off_")) {
     opr->lmod = OPLM_DWORD;
     return 1;
   }
-  if (!strncmp(opr->name, "word_", 5)) {
+  if (IS_START(opr->name, "word_")) {
     opr->lmod = OPLM_WORD;
     return 1;
   }
-  if (!strncmp(opr->name, "byte_", 5)) {
+  if (IS_START(opr->name, "byte_")) {
     opr->lmod = OPLM_BYTE;
     return 1;
   }
+  if (IS_START(opr->name, "qword_")) {
+    opr->lmod = OPLM_QWORD;
+    return 1;
+  }
   return 0;
 }
 
@@ -445,18 +602,21 @@ static int guess_lmod_from_c_type(enum opr_lenmod *lmod,
   const struct parsed_type *c_type)
 {
   static const char *dword_types[] = {
-    "int", "_DWORD", "UINT_PTR", "DWORD",
-    "WPARAM", "LPARAM", "UINT",
-    "HIMC",
+    "uint32_t", "int", "_DWORD", "UINT_PTR", "DWORD",
+    "WPARAM", "LPARAM", "UINT", "__int32",
+    "LONG", "HIMC", "BOOL", "size_t",
+    "float",
   };
   static const char *word_types[] = {
-    "uint16_t", "int16_t", "_WORD",
+    "uint16_t", "int16_t", "_WORD", "WORD",
     "unsigned __int16", "__int16",
   };
   static const char *byte_types[] = {
     "uint8_t", "int8_t", "char",
     "unsigned __int8", "__int8", "BYTE", "_BYTE",
-    "_UNKNOWN",
+    "CHAR", "_UNKNOWN",
+    // structures.. deal the same as with _UNKNOWN for now
+    "CRITICAL_SECTION",
   };
   const char *n;
   int i;
@@ -492,6 +652,24 @@ static int guess_lmod_from_c_type(enum opr_lenmod *lmod,
   return 0;
 }
 
+static char *default_cast_to(char *buf, size_t buf_size,
+  struct parsed_opr *opr)
+{
+  buf[0] = 0;
+
+  if (!opr->is_ptr || strchr(opr->name, '['))
+    return buf;
+  if (opr->pp == NULL || opr->pp->type.name == NULL
+    || opr->pp->is_fptr)
+  {
+    snprintf(buf, buf_size, "%s", "(void *)");
+    return buf;
+  }
+
+  snprintf(buf, buf_size, "(%s)", opr->pp->type.name);
+  return buf;
+}
+
 static enum opr_type lmod_from_directive(const char *d)
 {
   if (IS(d, "dd"))
@@ -521,7 +699,7 @@ static int parse_operand(struct parsed_opr *opr,
   int *regmask, int *regmask_indirect,
   char words[16][256], int wordc, int w, unsigned int op_flags)
 {
-  const struct parsed_proto *pp;
+  const struct parsed_proto *pp = NULL;
   enum opr_lenmod tmplmod;
   unsigned long number;
   char buf[256];
@@ -563,7 +741,10 @@ static int parse_operand(struct parsed_opr *opr,
 
     if (label != NULL) {
       opr->type = OPT_LABEL;
-      if (IS_START(label, "ds:")) {
+      ret = check_segment_prefix(label);
+      if (ret != 0) {
+        if (ret >= 5)
+          aerr("fs/gs used\n");
         opr->had_ds = 1;
         label += 3;
       }
@@ -580,6 +761,8 @@ static int parse_operand(struct parsed_opr *opr,
         opr->lmod = OPLM_WORD;
       else if (IS(words[w], "byte"))
         opr->lmod = OPLM_BYTE;
+      else if (IS(words[w], "qword"))
+        opr->lmod = OPLM_QWORD;
       else
         aerr("type parsing failed\n");
       w += 2;
@@ -590,8 +773,10 @@ static int parse_operand(struct parsed_opr *opr,
   if (wordc_in == 2) {
     if (IS(words[w], "offset")) {
       opr->type = OPT_OFFSET;
+      opr->lmod = OPLM_DWORD;
       strcpy(opr->name, words[w + 1]);
-      return wordc;
+      pp = proto_parse(g_fhdr, opr->name, 1);
+      goto do_label;
     }
     if (IS(words[w], "(offset")) {
       p = strchr(words[w + 1], ')');
@@ -607,7 +792,10 @@ static int parse_operand(struct parsed_opr *opr,
   if (wordc_in != 1)
     aerr("parse_operand 1 word expected\n");
 
-  if (IS_START(words[w], "ds:")) {
+  ret = check_segment_prefix(words[w]);
+  if (ret != 0) {
+    if (ret >= 5)
+      aerr("fs/gs used\n");
     opr->had_ds = 1;
     memmove(words[w], words[w] + 3, strlen(words[w]) - 2);
   }
@@ -620,12 +808,16 @@ static int parse_operand(struct parsed_opr *opr,
       aerr("[] parse failure\n");
 
     parse_indmode(opr->name, regmask_indirect, 1);
-    if (opr->lmod == OPLM_UNSPEC && parse_stack_el(opr->name, NULL)) {
+    if (opr->lmod == OPLM_UNSPEC && parse_stack_el(opr->name, NULL, 1))
+    {
       // might be an equ
       struct parsed_equ *eq =
-        equ_find(NULL, parse_stack_el(opr->name, NULL), &i);
+        equ_find(NULL, parse_stack_el(opr->name, NULL, 1), &i);
       if (eq)
         opr->lmod = eq->lmod;
+
+      // might be unaligned access
+      g_func_lmods |= 1 << OPLM_BYTE;
     }
     return wordc;
   }
@@ -657,7 +849,7 @@ static int parse_operand(struct parsed_opr *opr,
 
   // most likely var in data segment
   opr->type = OPT_LABEL;
-  pp = proto_parse(g_fhdr, opr->name, 0);
+  pp = proto_parse(g_fhdr, opr->name, g_quiet_pp);
 
 do_label:
   if (pp != NULL) {
@@ -684,6 +876,7 @@ do_label:
     }
     opr->is_array = pp->type.is_array;
   }
+  opr->pp = pp;
 
   if (opr->lmod == OPLM_UNSPEC)
     guess_lmod_from_name(opr);
@@ -699,24 +892,35 @@ static const struct {
   { "repz",   OPF_REP|OPF_REPZ },
   { "repne",  OPF_REP|OPF_REPNZ },
   { "repnz",  OPF_REP|OPF_REPNZ },
+  { "lock",   OPF_LOCK }, // ignored for now..
 };
 
+#define OPF_CJMP_CC (OPF_JMP|OPF_CJMP|OPF_CC)
+
 static const struct {
   const char *name;
   enum op_op op;
-  unsigned int minopr;
-  unsigned int maxopr;
+  unsigned short minopr;
+  unsigned short maxopr;
   unsigned int flags;
+  unsigned char pfo;
+  unsigned char pfo_inv;
 } op_table[] = {
   { "nop",  OP_NOP,    0, 0, 0 },
   { "push", OP_PUSH,   1, 1, 0 },
   { "pop",  OP_POP,    1, 1, OPF_DATA },
+  { "leave",OP_LEAVE,  0, 0, OPF_DATA },
   { "mov" , OP_MOV,    2, 2, OPF_DATA },
   { "lea",  OP_LEA,    2, 2, OPF_DATA },
   { "movzx",OP_MOVZX,  2, 2, OPF_DATA },
   { "movsx",OP_MOVSX,  2, 2, OPF_DATA },
+  { "xchg", OP_XCHG,   2, 2, OPF_DATA },
   { "not",  OP_NOT,    1, 1, OPF_DATA },
+  { "xlat", OP_XLAT,   0, 0, OPF_DATA },
   { "cdq",  OP_CDQ,    0, 0, OPF_DATA },
+  { "lodsb",OP_LODS,   0, 0, OPF_DATA },
+  { "lodsw",OP_LODS,   0, 0, OPF_DATA },
+  { "lodsd",OP_LODS,   0, 0, OPF_DATA },
   { "stosb",OP_STOS,   0, 0, OPF_DATA },
   { "stosw",OP_STOS,   0, 0, OPF_DATA },
   { "stosd",OP_STOS,   0, 0, OPF_DATA },
@@ -726,6 +930,11 @@ static const struct {
   { "cmpsb",OP_CMPS,   0, 0, OPF_DATA|OPF_FLAGS },
   { "cmpsw",OP_CMPS,   0, 0, OPF_DATA|OPF_FLAGS },
   { "cmpsd",OP_CMPS,   0, 0, OPF_DATA|OPF_FLAGS },
+  { "scasb",OP_SCAS,   0, 0, OPF_DATA|OPF_FLAGS },
+  { "scasw",OP_SCAS,   0, 0, OPF_DATA|OPF_FLAGS },
+  { "scasd",OP_SCAS,   0, 0, OPF_DATA|OPF_FLAGS },
+  { "std",  OP_STD,    0, 0, OPF_DATA }, // special flag
+  { "cld",  OP_CLD,    0, 0, OPF_DATA },
   { "add",  OP_ADD,    2, 2, OPF_DATA|OPF_FLAGS },
   { "sub",  OP_SUB,    2, 2, OPF_DATA|OPF_FLAGS },
   { "and",  OP_AND,    2, 2, OPF_DATA|OPF_FLAGS },
@@ -735,10 +944,15 @@ static const struct {
   { "shr",  OP_SHR,    2, 2, OPF_DATA|OPF_FLAGS },
   { "sal",  OP_SHL,    2, 2, OPF_DATA|OPF_FLAGS },
   { "sar",  OP_SAR,    2, 2, OPF_DATA|OPF_FLAGS },
+  { "shld", OP_SHLD,   3, 3, OPF_DATA|OPF_FLAGS },
+  { "shrd", OP_SHRD,   3, 3, OPF_DATA|OPF_FLAGS },
   { "rol",  OP_ROL,    2, 2, OPF_DATA|OPF_FLAGS },
   { "ror",  OP_ROR,    2, 2, OPF_DATA|OPF_FLAGS },
-  { "adc",  OP_ADC,    2, 2, OPF_DATA|OPF_FLAGS|OPF_CC },
-  { "sbb",  OP_SBB,    2, 2, OPF_DATA|OPF_FLAGS|OPF_CC },
+  { "rcl",  OP_RCL,    2, 2, OPF_DATA|OPF_FLAGS|OPF_CC, PFO_C },
+  { "rcr",  OP_RCR,    2, 2, OPF_DATA|OPF_FLAGS|OPF_CC, PFO_C },
+  { "adc",  OP_ADC,    2, 2, OPF_DATA|OPF_FLAGS|OPF_CC, PFO_C },
+  { "sbb",  OP_SBB,    2, 2, OPF_DATA|OPF_FLAGS|OPF_CC, PFO_C },
+  { "bsf",  OP_BSF,    2, 2, OPF_DATA|OPF_FLAGS },
   { "inc",  OP_INC,    1, 1, OPF_DATA|OPF_FLAGS },
   { "dec",  OP_DEC,    1, 1, OPF_DATA|OPF_FLAGS },
   { "neg",  OP_NEG,    1, 1, OPF_DATA|OPF_FLAGS },
@@ -751,63 +965,103 @@ static const struct {
   { "retn", OP_RET,    0, 1, OPF_TAIL },
   { "call", OP_CALL,   1, 1, OPF_JMP|OPF_DATA|OPF_FLAGS },
   { "jmp",  OP_JMP,    1, 1, OPF_JMP },
-  { "jo",   OP_JO,     1, 1, OPF_JMP|OPF_CC }, // 70 OF=1
-  { "jno",  OP_JNO,    1, 1, OPF_JMP|OPF_CC }, // 71 OF=0
-  { "jc",   OP_JC,     1, 1, OPF_JMP|OPF_CC }, // 72 CF=1
-  { "jb",   OP_JC,     1, 1, OPF_JMP|OPF_CC }, // 72
-  { "jnc",  OP_JNC,    1, 1, OPF_JMP|OPF_CC }, // 73 CF=0
-  { "jnb",  OP_JNC,    1, 1, OPF_JMP|OPF_CC }, // 73
-  { "jae",  OP_JNC,    1, 1, OPF_JMP|OPF_CC }, // 73
-  { "jz",   OP_JZ,     1, 1, OPF_JMP|OPF_CC }, // 74 ZF=1
-  { "je",   OP_JZ,     1, 1, OPF_JMP|OPF_CC }, // 74
-  { "jnz",  OP_JNZ,    1, 1, OPF_JMP|OPF_CC }, // 75 ZF=0
-  { "jne",  OP_JNZ,    1, 1, OPF_JMP|OPF_CC }, // 75
-  { "jbe",  OP_JBE,    1, 1, OPF_JMP|OPF_CC }, // 76 CF=1 || ZF=1
-  { "jna",  OP_JBE,    1, 1, OPF_JMP|OPF_CC }, // 76
-  { "ja",   OP_JA,     1, 1, OPF_JMP|OPF_CC }, // 77 CF=0 && ZF=0
-  { "jnbe", OP_JA,     1, 1, OPF_JMP|OPF_CC }, // 77
-  { "js",   OP_JS,     1, 1, OPF_JMP|OPF_CC }, // 78 SF=1
-  { "jns",  OP_JNS,    1, 1, OPF_JMP|OPF_CC }, // 79 SF=0
-  { "jp",   OP_JP,     1, 1, OPF_JMP|OPF_CC }, // 7a PF=1
-  { "jpe",  OP_JP,     1, 1, OPF_JMP|OPF_CC }, // 7a
-  { "jnp",  OP_JNP,    1, 1, OPF_JMP|OPF_CC }, // 7b PF=0
-  { "jpo",  OP_JNP,    1, 1, OPF_JMP|OPF_CC }, // 7b
-  { "jl",   OP_JL,     1, 1, OPF_JMP|OPF_CC }, // 7c SF!=OF
-  { "jnge", OP_JL,     1, 1, OPF_JMP|OPF_CC }, // 7c
-  { "jge",  OP_JGE,    1, 1, OPF_JMP|OPF_CC }, // 7d SF=OF
-  { "jnl",  OP_JGE,    1, 1, OPF_JMP|OPF_CC }, // 7d
-  { "jle",  OP_JLE,    1, 1, OPF_JMP|OPF_CC }, // 7e ZF=1 || SF!=OF
-  { "jng",  OP_JLE,    1, 1, OPF_JMP|OPF_CC }, // 7e
-  { "jg",   OP_JG,     1, 1, OPF_JMP|OPF_CC }, // 7f ZF=0 && SF=OF
-  { "jnle", OP_JG,     1, 1, OPF_JMP|OPF_CC }, // 7f
-  { "seto",   OP_JO,   1, 1, OPF_DATA|OPF_CC },
-  { "setno",  OP_JNO,  1, 1, OPF_DATA|OPF_CC },
-  { "setc",   OP_JC,   1, 1, OPF_DATA|OPF_CC },
-  { "setb",   OP_JC,   1, 1, OPF_DATA|OPF_CC },
-  { "setnc",  OP_JNC,  1, 1, OPF_DATA|OPF_CC },
-  { "setae",  OP_JNC,  1, 1, OPF_DATA|OPF_CC },
-  { "setz",   OP_JZ,   1, 1, OPF_DATA|OPF_CC },
-  { "sete",   OP_JZ,   1, 1, OPF_DATA|OPF_CC },
-  { "setnz",  OP_JNZ,  1, 1, OPF_DATA|OPF_CC },
-  { "setne",  OP_JNZ,  1, 1, OPF_DATA|OPF_CC },
-  { "setbe",  OP_JBE,  1, 1, OPF_DATA|OPF_CC },
-  { "setna",  OP_JBE,  1, 1, OPF_DATA|OPF_CC },
-  { "seta",   OP_JA,   1, 1, OPF_DATA|OPF_CC },
-  { "setnbe", OP_JA,   1, 1, OPF_DATA|OPF_CC },
-  { "sets",   OP_JS,   1, 1, OPF_DATA|OPF_CC },
-  { "setns",  OP_JNS,  1, 1, OPF_DATA|OPF_CC },
-  { "setp",   OP_JP,   1, 1, OPF_DATA|OPF_CC },
-  { "setpe",  OP_JP,   1, 1, OPF_DATA|OPF_CC },
-  { "setnp",  OP_JNP,  1, 1, OPF_DATA|OPF_CC },
-  { "setpo",  OP_JNP,  1, 1, OPF_DATA|OPF_CC },
-  { "setl",   OP_JL,   1, 1, OPF_DATA|OPF_CC },
-  { "setnge", OP_JL,   1, 1, OPF_DATA|OPF_CC },
-  { "setge",  OP_JGE,  1, 1, OPF_DATA|OPF_CC },
-  { "setnl",  OP_JGE,  1, 1, OPF_DATA|OPF_CC },
-  { "setle",  OP_JLE,  1, 1, OPF_DATA|OPF_CC },
-  { "setng",  OP_JLE,  1, 1, OPF_DATA|OPF_CC },
-  { "setg",   OP_JG,   1, 1, OPF_DATA|OPF_CC },
-  { "setnle", OP_JG,   1, 1, OPF_DATA|OPF_CC },
+  { "jecxz",OP_JECXZ,  1, 1, OPF_JMP|OPF_CJMP },
+  { "loop", OP_LOOP,   1, 1, OPF_JMP|OPF_CJMP|OPF_DATA },
+  { "jo",   OP_JCC,    1, 1, OPF_CJMP_CC, PFO_O,  0 }, // 70 OF=1
+  { "jno",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_O,  1 }, // 71 OF=0
+  { "jc",   OP_JCC,    1, 1, OPF_CJMP_CC, PFO_C,  0 }, // 72 CF=1
+  { "jb",   OP_JCC,    1, 1, OPF_CJMP_CC, PFO_C,  0 }, // 72
+  { "jnc",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_C,  1 }, // 73 CF=0
+  { "jnb",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_C,  1 }, // 73
+  { "jae",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_C,  1 }, // 73
+  { "jz",   OP_JCC,    1, 1, OPF_CJMP_CC, PFO_Z,  0 }, // 74 ZF=1
+  { "je",   OP_JCC,    1, 1, OPF_CJMP_CC, PFO_Z,  0 }, // 74
+  { "jnz",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_Z,  1 }, // 75 ZF=0
+  { "jne",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_Z,  1 }, // 75
+  { "jbe",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_BE, 0 }, // 76 CF=1||ZF=1
+  { "jna",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_BE, 0 }, // 76
+  { "ja",   OP_JCC,    1, 1, OPF_CJMP_CC, PFO_BE, 1 }, // 77 CF=0&&ZF=0
+  { "jnbe", OP_JCC,    1, 1, OPF_CJMP_CC, PFO_BE, 1 }, // 77
+  { "js",   OP_JCC,    1, 1, OPF_CJMP_CC, PFO_S,  0 }, // 78 SF=1
+  { "jns",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_S,  1 }, // 79 SF=0
+  { "jp",   OP_JCC,    1, 1, OPF_CJMP_CC, PFO_P,  0 }, // 7a PF=1
+  { "jpe",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_P,  0 }, // 7a
+  { "jnp",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_P,  1 }, // 7b PF=0
+  { "jpo",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_P,  1 }, // 7b
+  { "jl",   OP_JCC,    1, 1, OPF_CJMP_CC, PFO_L,  0 }, // 7c SF!=OF
+  { "jnge", OP_JCC,    1, 1, OPF_CJMP_CC, PFO_L,  0 }, // 7c
+  { "jge",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_L,  1 }, // 7d SF=OF
+  { "jnl",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_L,  1 }, // 7d
+  { "jle",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_LE, 0 }, // 7e ZF=1||SF!=OF
+  { "jng",  OP_JCC,    1, 1, OPF_CJMP_CC, PFO_LE, 0 }, // 7e
+  { "jg",   OP_JCC,    1, 1, OPF_CJMP_CC, PFO_LE, 1 }, // 7f ZF=0&&SF=OF
+  { "jnle", OP_JCC,    1, 1, OPF_CJMP_CC, PFO_LE, 1 }, // 7f
+  { "seto",   OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_O,  0 },
+  { "setno",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_O,  1 },
+  { "setc",   OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_C,  0 },
+  { "setb",   OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_C,  0 },
+  { "setnc",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_C,  1 },
+  { "setae",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_C,  1 },
+  { "setnb",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_C,  1 },
+  { "setz",   OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_Z,  0 },
+  { "sete",   OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_Z,  0 },
+  { "setnz",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_Z,  1 },
+  { "setne",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_Z,  1 },
+  { "setbe",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_BE, 0 },
+  { "setna",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_BE, 0 },
+  { "seta",   OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_BE, 1 },
+  { "setnbe", OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_BE, 1 },
+  { "sets",   OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_S,  0 },
+  { "setns",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_S,  1 },
+  { "setp",   OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_P,  0 },
+  { "setpe",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_P,  0 },
+  { "setnp",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_P,  1 },
+  { "setpo",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_P,  1 },
+  { "setl",   OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_L,  0 },
+  { "setnge", OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_L,  0 },
+  { "setge",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_L,  1 },
+  { "setnl",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_L,  1 },
+  { "setle",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_LE, 0 },
+  { "setng",  OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_LE, 0 },
+  { "setg",   OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_LE, 1 },
+  { "setnle", OP_SCC,  1, 1, OPF_DATA|OPF_CC, PFO_LE, 1 },
+  // x87
+  { "fld",    OP_FLD,    1, 1, OPF_FPUSH },
+  { "fild",   OP_FILD,   1, 1, OPF_FPUSH },
+  { "fld1",   OP_FLDc,   0, 0, OPF_FPUSH },
+  { "fldz",   OP_FLDc,   0, 0, OPF_FPUSH },
+  { "fstp",   OP_FST,    1, 1, OPF_FPOP },
+  { "fst",    OP_FST,    1, 1, 0 },
+  { "fadd",   OP_FADD,   0, 2, 0 },
+  { "faddp",  OP_FADD,   0, 2, OPF_FPOP },
+  { "fdiv",   OP_FDIV,   0, 2, 0 },
+  { "fdivp",  OP_FDIV,   0, 2, OPF_FPOP },
+  { "fmul",   OP_FMUL,   0, 2, 0 },
+  { "fmulp",  OP_FMUL,   0, 2, OPF_FPOP },
+  { "fsub",   OP_FSUB,   0, 2, 0 },
+  { "fsubp",  OP_FSUB,   0, 2, OPF_FPOP },
+  { "fdivr",  OP_FDIVR,  0, 2, 0 },
+  { "fdivrp", OP_FDIVR,  0, 2, OPF_FPOP },
+  { "fsubr",  OP_FSUBR,  0, 2, 0 },
+  { "fsubrp", OP_FSUBR,  0, 2, OPF_FPOP },
+  { "fiadd",  OP_FIADD,  1, 1, 0 },
+  { "fidiv",  OP_FIDIV,  1, 1, 0 },
+  { "fimul",  OP_FIMUL,  1, 1, 0 },
+  { "fisub",  OP_FISUB,  1, 1, 0 },
+  { "fidivr", OP_FIDIVR, 1, 1, 0 },
+  { "fisubr", OP_FISUBR, 1, 1, 0 },
+  { "fcos",   OP_FCOS,   0, 0, 0 },
+  { "fpatan", OP_FPATAN, 0, 0, OPF_FPOP },
+  { "fptan",  OP_FPTAN,  0, 0, OPF_FPUSH },
+  { "fsin",   OP_FSIN,   0, 0, 0 },
+  { "fsqrt",  OP_FSQRT,  0, 0, 0 },
+  // mmx
+  { "emms",   OP_EMMS,   0, 0, OPF_DATA },
+  { "movq",   OP_MOV,    2, 2, OPF_DATA },
+  // pseudo-ops for lib calls
+  { "_ftol",  OPP_FTOL },
+  // must be last
+  { "ud2",    OP_UD2 },
 };
 
 static void parse_op(struct parsed_op *op, char words[16][256], int wordc)
@@ -819,7 +1073,7 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc)
   int op_w = 0;
   int opr = 0;
   int w = 0;
-  int i;
+  int i, j;
 
   for (i = 0; i < ARRAY_SIZE(pref_table); i++) {
     if (IS(words[w], pref_table[i].name)) {
@@ -840,29 +1094,39 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc)
       break;
   }
 
-  if (i == ARRAY_SIZE(op_table))
-    aerr("unhandled op: '%s'\n", words[0]);
+  if (i == ARRAY_SIZE(op_table)) {
+    if (!g_skip_func)
+      aerr("unhandled op: '%s'\n", words[0]);
+    i--; // OP_UD2
+  }
   w++;
 
   op->op = op_table[i].op;
   op->flags = op_table[i].flags | prefix_flags;
+  op->pfo = op_table[i].pfo;
+  op->pfo_inv = op_table[i].pfo_inv;
   op->regmask_src = op->regmask_dst = 0;
+  op->asmln = asmln;
+
+  if (op->op == OP_UD2)
+    return;
+
+  for (opr = 0; opr < op_table[i].maxopr; opr++) {
+    if (opr >= op_table[i].minopr && w >= wordc)
+      break;
 
-  for (opr = 0; opr < op_table[i].minopr; opr++) {
     regmask = regmask_ind = 0;
     w = parse_operand(&op->operand[opr], &regmask, &regmask_ind,
       words, wordc, w, op->flags);
 
     if (opr == 0 && (op->flags & OPF_DATA))
       op->regmask_dst = regmask;
-    // for now, mark dst as src too
-    op->regmask_src |= regmask | regmask_ind;
-  }
+    else
+      op->regmask_src |= regmask;
+    op->regmask_src |= regmask_ind;
 
-  for (; w < wordc && opr < op_table[i].maxopr; opr++) {
-    w = parse_operand(&op->operand[opr],
-      &op->regmask_src, &op->regmask_src,
-      words, wordc, w, op->flags);
+    if (op->operand[opr].lmod != OPLM_UNSPEC)
+      g_func_lmods |= 1 << op->operand[opr].lmod;
   }
 
   if (w < wordc)
@@ -874,55 +1138,129 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc)
   if (!strncmp(op_table[i].name, "set", 3))
     op->operand[0].lmod = OPLM_BYTE;
 
-  // ops with implicit argumets
   switch (op->op) {
+  // first operand is not dst
+  case OP_CMP:
+  case OP_TEST:
+    op->regmask_src |= op->regmask_dst;
+    op->regmask_dst = 0;
+    break;
+
+  // first operand is src too
+  case OP_NOT:
+  case OP_ADD:
+  case OP_AND:
+  case OP_OR:
+  case OP_RCL:
+  case OP_RCR:
+  case OP_ADC:
+  case OP_INC:
+  case OP_DEC:
+  case OP_NEG:
+  // more below..
+    op->regmask_src |= op->regmask_dst;
+    break;
+
+  // special
+  case OP_XCHG:
+    op->regmask_src |= op->regmask_dst;
+    op->regmask_dst |= op->regmask_src;
+    goto check_align;
+
+  case OP_SUB:
+  case OP_SBB:
+  case OP_XOR:
+    if (op->operand[0].type == OPT_REG && op->operand[1].type == OPT_REG
+     && op->operand[0].lmod == op->operand[1].lmod
+     && op->operand[0].reg == op->operand[1].reg
+     && IS(op->operand[0].name, op->operand[1].name)) // ! ah, al..
+    {
+      op->regmask_src = 0;
+    }
+    else
+      op->regmask_src |= op->regmask_dst;
+    break;
+
+  // ops with implicit argumets
+  case OP_XLAT:
+    op->operand_cnt = 2;
+    setup_reg_opr(&op->operand[0], xAX, OPLM_BYTE, &op->regmask_src);
+    op->regmask_dst = op->regmask_src;
+    setup_reg_opr(&op->operand[1], xBX, OPLM_DWORD, &op->regmask_src);
+    break;
+
   case OP_CDQ:
     op->operand_cnt = 2;
     setup_reg_opr(&op->operand[0], xDX, OPLM_DWORD, &op->regmask_dst);
     setup_reg_opr(&op->operand[1], xAX, OPLM_DWORD, &op->regmask_src);
     break;
 
+  case OP_LODS:
   case OP_STOS:
-    if (op->operand_cnt != 0)
-      break;
-    if      (IS(words[op_w], "stosb"))
+  case OP_SCAS:
+    if      (words[op_w][4] == 'b')
       lmod = OPLM_BYTE;
-    else if (IS(words[op_w], "stosw"))
+    else if (words[op_w][4] == 'w')
       lmod = OPLM_WORD;
-    else if (IS(words[op_w], "stosd"))
+    else if (words[op_w][4] == 'd')
       lmod = OPLM_DWORD;
-    op->operand_cnt = 3;
-    setup_reg_opr(&op->operand[0], xDI, lmod, &op->regmask_src);
-    setup_reg_opr(&op->operand[1], xCX, OPLM_DWORD, &op->regmask_src);
+    j = 0;
+    op->regmask_src = 0;
+    setup_reg_opr(&op->operand[j++], op->op == OP_LODS ? xSI : xDI,
+      OPLM_DWORD, &op->regmask_src);
     op->regmask_dst = op->regmask_src;
-    setup_reg_opr(&op->operand[2], xAX, OPLM_DWORD, &op->regmask_src);
+    setup_reg_opr(&op->operand[j++], xAX, lmod,
+      op->op == OP_LODS ? &op->regmask_dst : &op->regmask_src);
+    if (op->flags & OPF_REP) {
+      setup_reg_opr(&op->operand[j++], xCX, OPLM_DWORD, &op->regmask_src);
+      op->regmask_dst |= 1 << xCX;
+    }
+    op->operand_cnt = j;
     break;
 
   case OP_MOVS:
   case OP_CMPS:
-    if (op->operand_cnt != 0)
-      break;
     if      (words[op_w][4] == 'b')
       lmod = OPLM_BYTE;
     else if (words[op_w][4] == 'w')
       lmod = OPLM_WORD;
     else if (words[op_w][4] == 'd')
       lmod = OPLM_DWORD;
-    op->operand_cnt = 3;
-    setup_reg_opr(&op->operand[0], xDI, lmod, &op->regmask_src);
-    setup_reg_opr(&op->operand[1], xSI, OPLM_DWORD, &op->regmask_src);
-    setup_reg_opr(&op->operand[2], xCX, OPLM_DWORD, &op->regmask_src);
+    j = 0;
+    op->regmask_src = 0;
+    // note: lmod is not correct, don't have where to place it
+    setup_reg_opr(&op->operand[j++], xDI, lmod, &op->regmask_src);
+    setup_reg_opr(&op->operand[j++], xSI, OPLM_DWORD, &op->regmask_src);
+    if (op->flags & OPF_REP)
+      setup_reg_opr(&op->operand[j++], xCX, OPLM_DWORD, &op->regmask_src);
+    op->operand_cnt = j;
     op->regmask_dst = op->regmask_src;
     break;
 
+  case OP_LOOP:
+    op->regmask_dst = 1 << xCX;
+    // fallthrough
+  case OP_JECXZ:
+    op->operand_cnt = 2;
+    op->regmask_src = 1 << xCX;
+    op->operand[1].type = OPT_REG;
+    op->operand[1].reg = xCX;
+    op->operand[1].lmod = OPLM_DWORD;
+    break;
+
   case OP_IMUL:
+    if (op->operand_cnt == 2) {
+      if (op->operand[0].type != OPT_REG)
+        aerr("reg expected\n");
+      op->regmask_src |= 1 << op->operand[0].reg;
+    }
     if (op->operand_cnt != 1)
       break;
     // fallthrough
   case OP_MUL:
     // singleop mul
+    op->regmask_src |= op->regmask_dst;
     op->regmask_dst = (1 << xDX) | (1 << xAX);
-    op->regmask_src |= (1 << xAX);
     if (op->operand[0].lmod == OPLM_UNSPEC)
       op->operand[0].lmod = OPLM_DWORD;
     break;
@@ -931,9 +1269,8 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc)
   case OP_IDIV:
     // we could set up operands for edx:eax, but there is no real need to
     // (see is_opr_modified())
-    regmask = (1 << xDX) | (1 << xAX);
-    op->regmask_dst = regmask;
-    op->regmask_src |= regmask;
+    op->regmask_src |= op->regmask_dst;
+    op->regmask_dst = (1 << xDX) | (1 << xAX);
     if (op->operand[0].lmod == OPLM_UNSPEC)
       op->operand[0].lmod = OPLM_DWORD;
     break;
@@ -943,11 +1280,21 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc)
   case OP_SAR:
   case OP_ROL:
   case OP_ROR:
+    op->regmask_src |= op->regmask_dst;
     if (op->operand[1].lmod == OPLM_UNSPEC)
       op->operand[1].lmod = OPLM_BYTE;
     break;
 
+  case OP_SHLD:
+  case OP_SHRD:
+    op->regmask_src |= op->regmask_dst;
+    if (op->operand[2].lmod == OPLM_UNSPEC)
+      op->operand[2].lmod = OPLM_BYTE;
+    break;
+
   case OP_PUSH:
+    op->regmask_src |= op->regmask_dst;
+    op->regmask_dst = 0;
     if (op->operand[0].lmod == OPLM_UNSPEC
         && (op->operand[0].type == OPT_CONST
          || op->operand[0].type == OPT_OFFSET
@@ -957,10 +1304,14 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc)
 
   // alignment
   case OP_MOV:
+  check_align:
     if (op->operand[0].type == OPT_REG && op->operand[1].type == OPT_REG
-     && op->operand[0].reg == xDI && op->operand[1].reg == xDI)
+     && op->operand[0].lmod == op->operand[1].lmod
+     && op->operand[0].reg == op->operand[1].reg
+     && IS(op->operand[0].name, op->operand[1].name)) // ! ah, al..
     {
-      op->flags |= OPF_RMD;
+      op->flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+      op->regmask_src = op->regmask_dst = 0;
     }
     break;
 
@@ -971,7 +1322,7 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc)
       char buf[16];
       snprintf(buf, sizeof(buf), "%s+0", op->operand[0].name);
       if (IS(buf, op->operand[1].name))
-        op->flags |= OPF_RMD;
+        op->flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
     }
     break;
 
@@ -980,17 +1331,109 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc)
     op->regmask_dst = 0;
     break;
 
+  case OP_LEAVE:
+    op->regmask_dst = (1 << xBP) | (1 << xSP);
+    op->regmask_src =  1 << xBP;
+    break;
+
+  case OP_FLD:
+  case OP_FILD:
+    op->regmask_dst |= mxST0;
+    break;
+
+  case OP_FLDc:
+    op->regmask_dst |= mxST0;
+    if      (IS(words[op_w] + 3, "1"))
+      op->operand[0].val = X87_CONST_1;
+    else if (IS(words[op_w] + 3, "z"))
+      op->operand[0].val = X87_CONST_Z;
+    else
+      aerr("TODO\n");
+    break;
+
+  case OP_FST:
+    op->regmask_src |= mxST0;
+    break;
+
+  case OP_FADD:
+  case OP_FDIV:
+  case OP_FMUL:
+  case OP_FSUB:
+  case OP_FDIVR:
+  case OP_FSUBR:
+    op->regmask_src |= mxST0;
+    if (op->operand_cnt == 2)
+      op->regmask_src |= op->regmask_dst;
+    else if (op->operand_cnt == 1) {
+      memcpy(&op->operand[1], &op->operand[0], sizeof(op->operand[1]));
+      op->operand[0].type = OPT_REG;
+      op->operand[0].lmod = OPLM_QWORD;
+      op->operand[0].reg = xST0;
+      op->regmask_dst |= mxST0;
+    }
+    else
+      // IDA doesn't use this
+      aerr("no operands?\n");
+    break;
+
+  case OP_FIADD:
+  case OP_FIDIV:
+  case OP_FIMUL:
+  case OP_FISUB:
+  case OP_FIDIVR:
+  case OP_FISUBR:
+  case OP_FCOS:
+  case OP_FSIN:
+  case OP_FSQRT:
+    op->regmask_src |= mxST0;
+    op->regmask_dst |= mxST0;
+    break;
+
+  case OP_FPATAN:
+    op->regmask_src |= mxST0 | mxST1;
+    op->regmask_dst |= mxST0;
+    break;
+
+  case OP_FPTAN:
+    aerr("TODO\n");
+    break;
+
   default:
     break;
   }
+
+  if (op->operand[0].type == OPT_REG
+   && op->operand[1].type == OPT_CONST)
+  {
+    struct parsed_opr *op1 = &op->operand[1];
+    if ((op->op == OP_AND && op1->val == 0)
+     || (op->op == OP_OR
+      && (op1->val == ~0
+       || (op->operand[0].lmod == OPLM_WORD && op1->val == 0xffff)
+       || (op->operand[0].lmod == OPLM_BYTE && op1->val == 0xff))))
+    {
+      op->regmask_src = 0;
+    }
+  }
 }
 
-static const char *op_name(enum op_op op)
+static const char *op_name(struct parsed_op *po)
 {
+  static char buf[16];
+  char *p;
   int i;
 
+  if (po->op == OP_JCC || po->op == OP_SCC) {
+    p = buf;
+    *p++ = (po->op == OP_JCC) ? 'j' : 's';
+    if (po->pfo_inv)
+      *p++ = 'n';
+    strcpy(p, parsed_flag_op_names[po->pfo]);
+    return buf;
+  }
+
   for (i = 0; i < ARRAY_SIZE(op_table); i++)
-    if (op_table[i].op == op)
+    if (op_table[i].op == po->op)
       return op_table[i].name;
 
   return "???";
@@ -1006,7 +1449,7 @@ static const char *dump_op(struct parsed_op *po)
   if (po == NULL)
     return "???";
 
-  snprintf(out, sizeof(out), "%s", op_name(po->op));
+  snprintf(out, sizeof(out), "%s", op_name(po));
   for (i = 0; i < po->operand_cnt; i++) {
     p += strlen(p);
     if (i > 0)
@@ -1023,6 +1466,8 @@ static const char *lmod_type_u(struct parsed_op *po,
   enum opr_lenmod lmod)
 {
   switch (lmod) {
+  case OPLM_QWORD:
+    return "u64";
   case OPLM_DWORD:
     return "u32";
   case OPLM_WORD:
@@ -1039,6 +1484,8 @@ static const char *lmod_cast_u(struct parsed_op *po,
   enum opr_lenmod lmod)
 {
   switch (lmod) {
+  case OPLM_QWORD:
+    return "";
   case OPLM_DWORD:
     return "";
   case OPLM_WORD:
@@ -1055,6 +1502,8 @@ static const char *lmod_cast_u_ptr(struct parsed_op *po,
   enum opr_lenmod lmod)
 {
   switch (lmod) {
+  case OPLM_QWORD:
+    return "*(u64 *)";
   case OPLM_DWORD:
     return "*(u32 *)";
   case OPLM_WORD:
@@ -1071,6 +1520,8 @@ static const char *lmod_cast_s(struct parsed_op *po,
   enum opr_lenmod lmod)
 {
   switch (lmod) {
+  case OPLM_QWORD:
+    return "(s64)";
   case OPLM_DWORD:
     return "(s32)";
   case OPLM_WORD:
@@ -1094,6 +1545,8 @@ static const char *lmod_cast(struct parsed_op *po,
 static int lmod_bytes(struct parsed_op *po, enum opr_lenmod lmod)
 {
   switch (lmod) {
+  case OPLM_QWORD:
+    return 8;
   case OPLM_DWORD:
     return 4;
   case OPLM_WORD:
@@ -1124,15 +1577,50 @@ static unsigned int opr_const(struct parsed_op *po, int opr_num)
 
 static const char *opr_reg_p(struct parsed_op *po, struct parsed_opr *popr)
 {
-  if ((unsigned int)popr->reg >= MAX_REGS)
+  if ((unsigned int)popr->reg >= ARRAY_SIZE(regs_r32))
     ferr(po, "invalid reg: %d\n", popr->reg);
   return regs_r32[popr->reg];
 }
 
+static int check_simple_cast(const char *cast, int *bits, int *is_signed)
+{
+  if      (IS_START(cast, "(s8)") || IS_START(cast, "(u8)"))
+    *bits = 8;
+  else if (IS_START(cast, "(s16)") || IS_START(cast, "(u16)"))
+    *bits = 16;
+  else if (IS_START(cast, "(s32)") || IS_START(cast, "(u32)"))
+    *bits = 32;
+  else if (IS_START(cast, "(s64)") || IS_START(cast, "(u64)"))
+    *bits = 64;
+  else
+    return -1;
+
+  *is_signed = cast[1] == 's' ? 1 : 0;
+  return 0;
+}
+
+static int check_deref_cast(const char *cast, int *bits)
+{
+  if      (IS_START(cast, "*(u8 *)"))
+    *bits = 8;
+  else if (IS_START(cast, "*(u16 *)"))
+    *bits = 16;
+  else if (IS_START(cast, "*(u32 *)"))
+    *bits = 32;
+  else if (IS_START(cast, "*(u64 *)"))
+    *bits = 64;
+  else
+    return -1;
+
+  return 0;
+}
+
 // cast1 is the "final" cast
 static const char *simplify_cast(const char *cast1, const char *cast2)
 {
   static char buf[256];
+  int bits1, bits2;
+  int s1, s2;
 
   if (cast1[0] == 0)
     return cast2;
@@ -1140,14 +1628,22 @@ static const char *simplify_cast(const char *cast1, const char *cast2)
     return cast1;
   if (IS(cast1, cast2))
     return cast1;
-  if (IS(cast1, "(s8)") && IS(cast2, "(u8)"))
-    return cast1;
-  if (IS(cast1, "(s16)") && IS(cast2, "(u16)"))
-    return cast1;
-  if (IS(cast1, "(u8)") && IS_START(cast2, "*(u8 *)"))
-    return cast2;
-  if (IS(cast1, "(u16)") && IS_START(cast2, "*(u16 *)"))
-    return cast2;
+
+  if (check_simple_cast(cast1, &bits1, &s1) == 0
+    && check_simple_cast(cast2, &bits2, &s2) == 0)
+  {
+    if (bits1 <= bits2)
+      return cast1;
+  }
+  if (check_simple_cast(cast1, &bits1, &s1) == 0
+    && check_deref_cast(cast2, &bits2) == 0)
+  {
+    if (bits1 == bits2) {
+      snprintf(buf, sizeof(buf), "*(%c%d *)", s1 ? 's' : 'u', bits1);
+      return buf;
+    }
+  }
+
   if (strchr(cast1, '*') && IS_START(cast2, "(u32)"))
     return cast1;
 
@@ -1211,14 +1707,14 @@ static struct parsed_equ *equ_find(struct parsed_op *po, const char *name,
 static int is_stack_access(struct parsed_op *po,
   const struct parsed_opr *popr)
 {
-  return (parse_stack_el(popr->name, NULL)
+  return (parse_stack_el(popr->name, NULL, 0)
     || (g_bp_frame && !(po->flags & OPF_EBP_S)
         && IS_START(popr->name, "ebp")));
 }
 
 static void parse_stack_access(struct parsed_op *po,
   const char *name, char *ofs_reg, int *offset_out,
-  int *stack_ra_out, const char **bp_arg_out)
+  int *stack_ra_out, const char **bp_arg_out, int is_lea)
 {
   const char *bp_arg = "";
   const char *p = NULL;
@@ -1242,7 +1738,7 @@ static void parse_stack_access(struct parsed_op *po,
       ferr(po, "ebp- parse of '%s' failed\n", name);
   }
   else {
-    bp_arg = parse_stack_el(name, ofs_reg);
+    bp_arg = parse_stack_el(name, ofs_reg, 0);
     snprintf(g_comment, sizeof(g_comment), "%s", bp_arg);
     eq = equ_find(po, bp_arg, &offset);
     if (eq == NULL)
@@ -1253,8 +1749,12 @@ static void parse_stack_access(struct parsed_op *po,
   if (!strncmp(name, "ebp", 3))
     stack_ra = 4;
 
-  if (ofs_reg[0] == 0 && stack_ra <= offset && offset < stack_ra + 4)
+  // yes it sometimes LEAs ra for compares..
+  if (!is_lea && ofs_reg[0] == 0
+    && stack_ra <= offset && offset < stack_ra + 4)
+  {
     ferr(po, "reference to ra? %d %d\n", offset, stack_ra);
+  }
 
   *offset_out = offset;
   *stack_ra_out = stack_ra;
@@ -1262,7 +1762,7 @@ static void parse_stack_access(struct parsed_op *po,
     *bp_arg_out = bp_arg;
 }
 
-static void stack_frame_access(struct parsed_op *po,
+static int stack_frame_access(struct parsed_op *po,
   struct parsed_opr *popr, char *buf, size_t buf_size,
   const char *name, const char *cast, int is_src, int is_lea)
 {
@@ -1274,13 +1774,15 @@ static void stack_frame_access(struct parsed_op *po,
   int unaligned = 0;
   int stack_ra = 0;
   int offset = 0;
+  int retval = -1;
   int sf_ofs;
   int lim;
 
   if (po->flags & OPF_EBP_S)
     ferr(po, "stack_frame_access while ebp is scratch\n");
 
-  parse_stack_access(po, name, ofs_reg, &offset, &stack_ra, &bp_arg);
+  parse_stack_access(po, name, ofs_reg, &offset,
+    &stack_ra, &bp_arg, is_lea);
 
   if (offset > stack_ra)
   {
@@ -1294,7 +1796,7 @@ static void stack_frame_access(struct parsed_op *po,
         if (cast[0] == 0)
           cast = "(u32)";
         snprintf(buf, buf_size, "%sap", cast);
-        return;
+        return -1;
       }
       ferr(po, "offset %d (%s,%d) doesn't map to any arg\n",
         offset, bp_arg, arg_i);
@@ -1313,6 +1815,7 @@ static void stack_frame_access(struct parsed_op *po,
       ferr(po, "arg %d not in prototype?\n", arg_i);
 
     popr->is_ptr = g_func_pp->arg[i].type.is_ptr;
+    retval = i;
 
     switch (popr->lmod)
     {
@@ -1384,13 +1887,16 @@ static void stack_frame_access(struct parsed_op *po,
     // common problem
     guess_lmod_from_c_type(&tmp_lmod, &g_func_pp->arg[i].type);
     if (tmp_lmod != OPLM_DWORD
-      && (unaligned || (!is_src && tmp_lmod < popr->lmod)))
+      && (unaligned || (!is_src && lmod_bytes(po, tmp_lmod)
+                         < lmod_bytes(po, popr->lmod) + (offset & 3))))
     {
       ferr(po, "bp_arg arg%d/w offset %d and type '%s' is too small\n",
         i + 1, offset, g_func_pp->arg[i].type.name);
     }
-    if (popr->is_ptr && popr->lmod != OPLM_DWORD)
-      ferr(po, "bp_arg arg%d: non-dword ptr access\n", i + 1);
+    // can't check this because msvc likes to reuse
+    // arg space for scratch..
+    //if (popr->is_ptr && popr->lmod != OPLM_DWORD)
+    //  ferr(po, "bp_arg arg%d: non-dword ptr access\n", i + 1);
   }
   else
   {
@@ -1441,16 +1947,28 @@ static void stack_frame_access(struct parsed_op *po,
       snprintf(buf, buf_size, "%ssf.d[%d]", prefix, sf_ofs / 4);
       break;
 
+    case OPLM_QWORD:
+      ferr_assert(po, !(sf_ofs & 7));
+      ferr_assert(po, ofs_reg[0] == 0);
+      // float callers set is_lea
+      ferr_assert(po, is_lea);
+      snprintf(buf, buf_size, "%ssf.q[%d]", prefix, sf_ofs / 8);
+      break;
+
     default:
       ferr(po, "bp_stack bad lmod: %d\n", popr->lmod);
     }
   }
+
+  return retval;
 }
 
 static void check_func_pp(struct parsed_op *po,
   const struct parsed_proto *pp, const char *pfx)
 {
+  enum opr_lenmod tmp_lmod;
   char buf[256];
+  int ret, i;
 
   if (pp->argc_reg != 0) {
     if (/*!g_allow_regfunc &&*/ !pp->is_fastcall) {
@@ -1461,9 +1979,22 @@ static void check_func_pp(struct parsed_op *po,
       ferr(po, "%s: %d reg arg(s) with %d stack arg(s)\n",
         pfx, pp->argc_reg, pp->argc_stack);
   }
+
+  // fptrs must use 32bit args, callsite might have no information and
+  // lack a cast to smaller types, which results in incorrectly masked
+  // args passed (callee may assume masked args, it does on ARM)
+  if (!pp->is_osinc) {
+    for (i = 0; i < pp->argc; i++) {
+      ret = guess_lmod_from_c_type(&tmp_lmod, &pp->arg[i].type);
+      if (ret && tmp_lmod != OPLM_DWORD)
+        ferr(po, "reference to %s with arg%d '%s'\n", pp->name,
+          i + 1, pp->arg[i].type.name);
+    }
+  }
 }
 
-static void check_label_read_ref(struct parsed_op *po, const char *name)
+static const char *check_label_read_ref(struct parsed_op *po,
+  const char *name)
 {
   const struct parsed_proto *pp;
 
@@ -1473,14 +2004,17 @@ static void check_label_read_ref(struct parsed_op *po, const char *name)
 
   if (pp->is_func)
     check_func_pp(po, pp, "ref");
+
+  return pp->name;
 }
 
 static char *out_src_opr(char *buf, size_t buf_size,
-	struct parsed_op *po, struct parsed_opr *popr, const char *cast,
+  struct parsed_op *po, struct parsed_opr *popr, const char *cast,
   int is_lea)
 {
   char tmp1[256], tmp2[256];
   char expr[256];
+  const char *name;
   char *p;
   int ret;
 
@@ -1493,6 +2027,9 @@ static char *out_src_opr(char *buf, size_t buf_size,
       ferr(po, "lea from reg?\n");
 
     switch (popr->lmod) {
+    case OPLM_QWORD:
+      snprintf(buf, buf_size, "%s%s.q", cast, opr_reg_p(po, popr));
+      break;
     case OPLM_DWORD:
       snprintf(buf, buf_size, "%s%s", cast, opr_reg_p(po, popr));
       break;
@@ -1548,29 +2085,28 @@ static char *out_src_opr(char *buf, size_t buf_size,
     break;
 
   case OPT_LABEL:
-    check_label_read_ref(po, popr->name);
+    name = check_label_read_ref(po, popr->name);
     if (cast[0] == 0 && popr->is_ptr)
       cast = "(u32)";
 
     if (is_lea)
-      snprintf(buf, buf_size, "(u32)&%s", popr->name);
+      snprintf(buf, buf_size, "(u32)&%s", name);
     else if (popr->size_lt)
       snprintf(buf, buf_size, "%s%s%s%s", cast,
         lmod_cast_u_ptr(po, popr->lmod),
-        popr->is_array ? "" : "&",
-        popr->name);
+        popr->is_array ? "" : "&", name);
     else
-      snprintf(buf, buf_size, "%s%s%s", cast, popr->name,
+      snprintf(buf, buf_size, "%s%s%s", cast, name,
         popr->is_array ? "[0]" : "");
     break;
 
   case OPT_OFFSET:
-    check_label_read_ref(po, popr->name);
+    name = check_label_read_ref(po, popr->name);
     if (cast[0] == 0)
       cast = "(u32)";
     if (is_lea)
       ferr(po, "lea an offset?\n");
-    snprintf(buf, buf_size, "%s&%s", cast, popr->name);
+    snprintf(buf, buf_size, "%s&%s", cast, name);
     break;
 
   case OPT_CONST:
@@ -1599,6 +2135,9 @@ static char *out_dst_opr(char *buf, size_t buf_size,
   switch (popr->type) {
   case OPT_REG:
     switch (popr->lmod) {
+    case OPLM_QWORD:
+      snprintf(buf, buf_size, "%s.q", opr_reg_p(po, popr));
+      break;
     case OPLM_DWORD:
       snprintf(buf, buf_size, "%s", opr_reg_p(po, popr));
       break;
@@ -1650,62 +2189,50 @@ static char *out_src_opr_u32(char *buf, size_t buf_size,
   return out_src_opr(buf, buf_size, po, popr, NULL, 0);
 }
 
-static enum parsed_flag_op split_cond(struct parsed_op *po,
-  enum op_op op, int *is_inv)
+static char *out_src_opr_float(char *buf, size_t buf_size,
+  struct parsed_op *po, struct parsed_opr *popr)
 {
-  *is_inv = 0;
-
-  switch (op) {
-  case OP_JO:
-    return PFO_O;
-  case OP_JC:
-    return PFO_C;
-  case OP_JZ:
-    return PFO_Z;
-  case OP_JBE:
-    return PFO_BE;
-  case OP_JS:
-    return PFO_S;
-  case OP_JP:
-    return PFO_P;
-  case OP_JL:
-    return PFO_L;
-  case OP_JLE:
-    return PFO_LE;
-
-  case OP_JNO:
-    *is_inv = 1;
-    return PFO_O;
-  case OP_JNC:
-    *is_inv = 1;
-    return PFO_C;
-  case OP_JNZ:
-    *is_inv = 1;
-    return PFO_Z;
-  case OP_JA:
-    *is_inv = 1;
-    return PFO_BE;
-  case OP_JNS:
-    *is_inv = 1;
-    return PFO_S;
-  case OP_JNP:
-    *is_inv = 1;
-    return PFO_P;
-  case OP_JGE:
-    *is_inv = 1;
-    return PFO_L;
-  case OP_JG:
-    *is_inv = 1;
-    return PFO_LE;
+  const char *cast = NULL;
+  char tmp[256];
 
-  case OP_ADC:
-  case OP_SBB:
-    return PFO_C;
+  switch (popr->type) {
+  case OPT_REG:
+    if (popr->reg < xST0 || popr->reg > xST7)
+      ferr(po, "bad reg: %d\n", popr->reg);
+
+    snprintf(buf, buf_size, "f_st%d", popr->reg - xST0);
+    break;
+
+  case OPT_REGMEM:
+  case OPT_LABEL:
+  case OPT_OFFSET:
+    switch (popr->lmod) {
+    case OPLM_QWORD:
+      cast = "double";
+      break;
+    case OPLM_DWORD:
+      cast = "float";
+      break;
+    default:
+      ferr(po, "unhandled lmod: %d\n", popr->lmod);
+      break;
+    }
+    out_src_opr(tmp, sizeof(tmp), po, popr, "", 1);
+    snprintf(buf, buf_size, "*((%s *)%s)", cast, tmp);
+    break;
 
   default:
-    ferr(po, "split_cond: bad op %d\n", op);
-    return -1;
+    ferr(po, "invalid float type: %d\n", popr->type);
   }
+
+  return buf;
+}
+
+static char *out_dst_opr_float(char *buf, size_t buf_size,
+  struct parsed_op *po, struct parsed_opr *popr)
+{
+  // same?
+  return out_src_opr_float(buf, buf_size, po, popr);
 }
 
 static void out_test_for_cc(char *buf, size_t buf_size,
@@ -1747,7 +2274,7 @@ static void out_cmp_for_cc(char *buf, size_t buf_size,
   char buf1[256], buf2[256];
   enum opr_lenmod lmod;
 
-  if (po->operand[0].lmod != po->operand[1].lmod)
+  if (po->op != OP_DEC && po->operand[0].lmod != po->operand[1].lmod)
     ferr(po, "%s: lmod mismatch: %d %d\n", __func__,
       po->operand[0].lmod, po->operand[1].lmod);
   lmod = po->operand[0].lmod;
@@ -1773,7 +2300,10 @@ static void out_cmp_for_cc(char *buf, size_t buf_size,
   }
 
   out_src_opr(buf1, sizeof(buf1), po, &po->operand[0], cast_use, 0);
-  out_src_opr(buf2, sizeof(buf2), po, &po->operand[1], cast_use, 0);
+  if (po->op == OP_DEC)
+    snprintf(buf2, sizeof(buf2), "1");
+  else
+    out_src_opr(buf2, sizeof(buf2), po, &po->operand[1], cast_use, 0);
 
   switch (pfo) {
   case PFO_C:
@@ -1813,7 +2343,7 @@ static void out_cmp_for_cc(char *buf, size_t buf_size,
       buf1, is_inv ? ">=" : "<", buf2);
     break;
 
-  case PFO_LE:
+  case PFO_LE: // !g
     snprintf(buf, buf_size, "(%s %s %s)",
       buf1, is_inv ? ">" : "<=", buf2);
     break;
@@ -1904,65 +2434,60 @@ static const char *op_to_c(struct parsed_op *po)
   }
 }
 
-static void op_set_clear_flag(struct parsed_op *po,
-  enum op_flags flag_set, enum op_flags flag_clear)
-{
-  po->flags |= flag_set;
-  po->flags &= ~flag_clear;
-}
-
 // last op in stream - unconditional branch or ret
 #define LAST_OP(_i) ((ops[_i].flags & OPF_TAIL) \
-  || (ops[_i].flags & (OPF_JMP|OPF_CC)) == OPF_JMP)
+  || ((ops[_i].flags & (OPF_JMP|OPF_CJMP|OPF_RMD)) == OPF_JMP \
+      && ops[_i].op != OP_CALL))
+
+#define check_i(po, i) \
+  if ((i) < 0) \
+    ferr(po, "bad " #i ": %d\n", i)
 
-static int scan_for_pop(int i, int opcnt, const char *reg,
-  int magic, int depth, int *maxdepth, int do_flags)
+// note: this skips over calls and rm'd stuff assuming they're handled
+// so it's intended to use at one of final passes
+static int scan_for_pop(int i, int opcnt, int magic, int reg,
+  int depth, int flags_set)
 {
-  const struct parsed_proto *pp;
   struct parsed_op *po;
+  int relevant;
   int ret = 0;
   int j;
 
   for (; i < opcnt; i++) {
     po = &ops[i];
     if (po->cc_scratch == magic)
-      break; // already checked
+      return ret; // already checked
     po->cc_scratch = magic;
 
     if (po->flags & OPF_TAIL) {
       if (po->op == OP_CALL) {
-        pp = proto_parse(g_fhdr, po->operand[0].name, 0);
-        if (pp != NULL && pp->is_noreturn)
-          // no stack cleanup for noreturn
-          return ret;
+        if (po->pp != NULL && po->pp->is_noreturn)
+          // assume no stack cleanup for noreturn
+          return 1;
       }
       return -1; // deadend
     }
 
-    if ((po->flags & OPF_RMD)
-        || (po->op == OP_PUSH && po->argnum != 0)) // arg push
+    if (po->flags & (OPF_RMD|OPF_DONE|OPF_FARG))
       continue;
 
     if ((po->flags & OPF_JMP) && po->op != OP_CALL) {
       if (po->btj != NULL) {
         // jumptable
         for (j = 0; j < po->btj->count; j++) {
-          ret |= scan_for_pop(po->btj->d[j].bt_i, opcnt, reg, magic,
-                   depth, maxdepth, do_flags);
+          check_i(po, po->btj->d[j].bt_i);
+          ret |= scan_for_pop(po->btj->d[j].bt_i, opcnt, magic, reg,
+                   depth, flags_set);
           if (ret < 0)
             return ret; // dead end
         }
         return ret;
       }
 
-      if (po->bt_i < 0) {
-        ferr(po, "dead branch\n");
-        return -1;
-      }
-
-      if (po->flags & OPF_CC) {
-        ret |= scan_for_pop(po->bt_i, opcnt, reg, magic,
-                 depth, maxdepth, do_flags);
+      check_i(po, po->bt_i);
+      if (po->flags & OPF_CJMP) {
+        ret |= scan_for_pop(po->bt_i, opcnt, magic, reg,
+                 depth, flags_set);
         if (ret < 0)
           return ret; // dead end
       }
@@ -1972,94 +2497,447 @@ static int scan_for_pop(int i, int opcnt, const char *reg,
       continue;
     }
 
+    relevant = 0;
     if ((po->op == OP_POP || po->op == OP_PUSH)
-        && po->operand[0].type == OPT_REG
-        && IS(po->operand[0].name, reg))
+      && po->operand[0].type == OPT_REG && po->operand[0].reg == reg)
     {
-      if (po->op == OP_PUSH && !(po->flags & OPF_FARG)) {
-        depth++;
-        if (depth > *maxdepth)
-          *maxdepth = depth;
-        if (do_flags)
-          op_set_clear_flag(po, OPF_RSAVE, OPF_RMD);
-      }
-      else if (po->op == OP_POP) {
-        if (depth == 0) {
-          if (do_flags)
-            op_set_clear_flag(po, OPF_RMD, OPF_RSAVE);
-          return 1;
-        }
-        else {
-          depth--;
-          if (depth < 0) // should not happen
-            ferr(po, "fail with depth\n");
-          if (do_flags)
-            op_set_clear_flag(po, OPF_RSAVE, OPF_RMD);
-        }
+      relevant = 1;
+    }
+
+    if (po->op == OP_PUSH) {
+      depth++;
+    }
+    else if (po->op == OP_POP) {
+      if (relevant && depth == 0) {
+        po->flags |= flags_set;
+        return 1;
       }
+      depth--;
     }
   }
 
-  return ret;
+  return -1;
 }
 
-// scan for pop starting from 'ret' op (all paths)
-static int scan_for_pop_ret(int i, int opcnt, const char *reg,
-  int flag_set)
+// scan for 'reg' pop backwards starting from i
+// intended to use for register restore search, so other reg
+// references are considered an error
+static int scan_for_rsave_pop_reg(int i, int magic, int reg, int set_flags)
 {
-  int found = 0;
-  int j;
-
-  for (; i < opcnt; i++) {
-    if (!(ops[i].flags & OPF_TAIL))
-      continue;
+  struct parsed_op *po;
+  struct label_ref *lr;
+  int ret = 0;
 
-    for (j = i - 1; j >= 0; j--) {
-      if (ops[j].flags & OPF_RMD)
-        continue;
-      if (ops[j].flags & OPF_JMP)
-        return -1;
+  ops[i].cc_scratch = magic;
 
-      if (ops[j].op == OP_POP && ops[j].operand[0].type == OPT_REG
-          && IS(ops[j].operand[0].name, reg))
-      {
-        found = 1;
-        ops[j].flags |= flag_set;
-        break;
+  while (1)
+  {
+    if (g_labels[i] != NULL) {
+      lr = &g_label_refs[i];
+      for (; lr != NULL; lr = lr->next) {
+        check_i(&ops[i], lr->i);
+        ret |= scan_for_rsave_pop_reg(lr->i, magic, reg, set_flags);
+        if (ret < 0)
+          return ret;
       }
+      if (i > 0 && LAST_OP(i - 1))
+        return ret;
+    }
+
+    i--;
+    if (i < 0)
+      break;
+
+    if (ops[i].cc_scratch == magic)
+      return ret;
+    ops[i].cc_scratch = magic;
 
-      if (g_labels[j][0] != 0)
+    po = &ops[i];
+    if (po->op == OP_POP && po->operand[0].reg == reg) {
+      if (po->flags & (OPF_RMD|OPF_DONE))
         return -1;
+
+      po->flags |= set_flags;
+      return 1;
     }
+
+    // this also covers the case where we reach corresponding push
+    if ((po->regmask_dst | po->regmask_src) & (1 << reg))
+      return -1;
   }
 
-  return found ? 0 : -1;
+  // nothing interesting on this path
+  return 0;
 }
 
-// is operand 'opr' modified by parsed_op 'po'?
-static int is_opr_modified(const struct parsed_opr *opr,
-  const struct parsed_op *po)
+static void find_reachable_exits(int i, int opcnt, int magic,
+  int *exits, int *exit_count)
 {
-  int mask;
-
-  if ((po->flags & OPF_RMD) || !(po->flags & OPF_DATA))
-    return 0;
+  struct parsed_op *po;
+  int j;
 
-  if (opr->type == OPT_REG) {
-    if (po->op == OP_CALL) {
-      mask = (1 << xAX) | (1 << xCX) | (1 << xDX);
-      if ((1 << opr->reg) & mask)
-        return 1;
+  for (; i < opcnt; i++)
+  {
+    po = &ops[i];
+    if (po->cc_scratch == magic)
+      return;
+    po->cc_scratch = magic;
+
+    if (po->flags & OPF_TAIL) {
+      ferr_assert(po, *exit_count < MAX_EXITS);
+      exits[*exit_count] = i;
+      (*exit_count)++;
+      return;
+    }
+
+    if ((po->flags & OPF_JMP) && po->op != OP_CALL) {
+      if (po->flags & OPF_RMD)
+        continue;
+
+      if (po->btj != NULL) {
+        for (j = 0; j < po->btj->count; j++) {
+          check_i(po, po->btj->d[j].bt_i);
+          find_reachable_exits(po->btj->d[j].bt_i, opcnt, magic,
+                  exits, exit_count);
+        }
+        return;
+      }
+
+      check_i(po, po->bt_i);
+      if (po->flags & OPF_CJMP)
+        find_reachable_exits(po->bt_i, opcnt, magic, exits, exit_count);
       else
+        i = po->bt_i - 1;
+      continue;
+    }
+  }
+}
+
+// scan for 'reg' pop backwards starting from exits (all paths)
+static int scan_for_pop_ret(int i, int opcnt, int reg, int set_flags)
+{
+  static int exits[MAX_EXITS];
+  static int exit_count;
+  int j, ret;
+
+  if (!set_flags) {
+    exit_count = 0;
+    find_reachable_exits(i, opcnt, i + opcnt * 15, exits,
+      &exit_count);
+    ferr_assert(&ops[i], exit_count > 0);
+  }
+
+  for (j = 0; j < exit_count; j++) {
+    ret = scan_for_rsave_pop_reg(exits[j], i + opcnt * 16 + set_flags,
+            reg, set_flags);
+    if (ret == -1)
+      return -1;
+  }
+
+  return 1;
+}
+
+// scan for one or more pop of push <const>
+static int scan_for_pop_const_r(int i, int opcnt, int magic,
+  int push_i, int is_probe)
+{
+  struct parsed_op *po;
+  struct label_ref *lr;
+  int ret = 0;
+  int j;
+
+  for (; i < opcnt; i++)
+  {
+    po = &ops[i];
+    if (po->cc_scratch == magic)
+      return ret; // already checked
+    po->cc_scratch = magic;
+
+    if (po->flags & OPF_JMP) {
+      if (po->flags & OPF_RMD)
+        continue;
+      if (po->op == OP_CALL)
+        return -1;
+
+      if (po->btj != NULL) {
+        for (j = 0; j < po->btj->count; j++) {
+          check_i(po, po->btj->d[j].bt_i);
+          ret |= scan_for_pop_const_r(po->btj->d[j].bt_i, opcnt, magic,
+                  push_i, is_probe);
+          if (ret < 0)
+            return ret;
+        }
+        return ret;
+      }
+
+      check_i(po, po->bt_i);
+      if (po->flags & OPF_CJMP) {
+        ret |= scan_for_pop_const_r(po->bt_i, opcnt, magic, push_i,
+                 is_probe);
+        if (ret < 0)
+          return ret;
+      }
+      else {
+        i = po->bt_i - 1;
+      }
+      continue;
+    }
+
+    if ((po->flags & (OPF_TAIL|OPF_RSAVE)) || po->op == OP_PUSH)
+      return -1;
+
+    if (g_labels[i] != NULL) {
+      // all refs must be visited
+      lr = &g_label_refs[i];
+      for (; lr != NULL; lr = lr->next) {
+        check_i(po, lr->i);
+        if (ops[lr->i].cc_scratch != magic)
+          return -1;
+      }
+      if (i > 0 && !LAST_OP(i - 1) && ops[i - 1].cc_scratch != magic)
+        return -1;
+    }
+
+    if (po->op == OP_POP)
+    {
+      if (po->flags & (OPF_RMD|OPF_DONE))
+        return -1;
+
+      if (!is_probe) {
+        po->flags |= OPF_DONE;
+        po->datap = &ops[push_i];
+      }
+      return 1;
+    }
+  }
+
+  return -1;
+}
+
+static void scan_for_pop_const(int i, int opcnt, int magic)
+{
+  int ret;
+
+  ret = scan_for_pop_const_r(i + 1, opcnt, magic, i, 1);
+  if (ret == 1) {
+    ops[i].flags |= OPF_RMD | OPF_DONE;
+    scan_for_pop_const_r(i + 1, opcnt, magic + 1, i, 0);
+  }
+}
+
+// check if all branch targets within a marked path are also marked
+// note: the path checked must not be empty or end with a branch
+static int check_path_branches(int opcnt, int magic)
+{
+  struct parsed_op *po;
+  int i, j;
+
+  for (i = 0; i < opcnt; i++) {
+    po = &ops[i];
+    if (po->cc_scratch != magic)
+      continue;
+
+    if (po->flags & OPF_JMP) {
+      if ((po->flags & OPF_RMD) || po->op == OP_CALL)
+        continue;
+
+      if (po->btj != NULL) {
+        for (j = 0; j < po->btj->count; j++) {
+          check_i(po, po->btj->d[j].bt_i);
+          if (ops[po->btj->d[j].bt_i].cc_scratch != magic)
+            return 0;
+        }
+      }
+
+      check_i(po, po->bt_i);
+      if (ops[po->bt_i].cc_scratch != magic)
+        return 0;
+      if ((po->flags & OPF_CJMP) && ops[i + 1].cc_scratch != magic)
         return 0;
     }
+  }
+
+  return 1;
+}
+
+// scan for multiple pushes for given pop
+static int scan_pushes_for_pop_r(int i, int magic, int pop_i,
+  int is_probe)
+{
+  int reg = ops[pop_i].operand[0].reg;
+  struct parsed_op *po;
+  struct label_ref *lr;
+  int ret = 0;
+
+  ops[i].cc_scratch = magic;
+
+  while (1)
+  {
+    if (g_labels[i] != NULL) {
+      lr = &g_label_refs[i];
+      for (; lr != NULL; lr = lr->next) {
+        check_i(&ops[i], lr->i);
+        ret |= scan_pushes_for_pop_r(lr->i, magic, pop_i, is_probe);
+        if (ret < 0)
+          return ret;
+      }
+      if (i > 0 && LAST_OP(i - 1))
+        return ret;
+    }
+
+    i--;
+    if (i < 0)
+      break;
+
+    if (ops[i].cc_scratch == magic)
+      return ret;
+    ops[i].cc_scratch = magic;
+
+    po = &ops[i];
+    if (po->op == OP_CALL)
+      return -1;
+    if ((po->flags & (OPF_TAIL|OPF_RSAVE)) || po->op == OP_POP)
+      return -1;
+
+    if (po->op == OP_PUSH)
+    {
+      if (po->datap != NULL)
+        return -1;
+      if (po->operand[0].type == OPT_REG && po->operand[0].reg == reg)
+        // leave this case for reg save/restore handlers
+        return -1;
+
+      if (!is_probe) {
+        po->flags |= OPF_PPUSH | OPF_DONE;
+        po->datap = &ops[pop_i];
+      }
+      return 1;
+    }
+  }
+
+  return -1;
+}
+
+static void scan_pushes_for_pop(int i, int opcnt, int *regmask_pp)
+{
+  int magic = i + opcnt * 14;
+  int ret;
+
+  ret = scan_pushes_for_pop_r(i, magic, i, 1);
+  if (ret == 1) {
+    ret = check_path_branches(opcnt, magic);
+    if (ret == 1) {
+      ops[i].flags |= OPF_PPUSH | OPF_DONE;
+      *regmask_pp |= 1 << ops[i].operand[0].reg;
+      scan_pushes_for_pop_r(i, magic + 1, i, 0);
+    }
+  }
+}
+
+static void scan_propagate_df(int i, int opcnt)
+{
+  struct parsed_op *po = &ops[i];
+  int j;
+
+  for (; i < opcnt; i++) {
+    po = &ops[i];
+    if (po->flags & OPF_DF)
+      return; // already resolved
+    po->flags |= OPF_DF;
+
+    if (po->op == OP_CALL)
+      ferr(po, "call with DF set?\n");
+
+    if (po->flags & OPF_JMP) {
+      if (po->btj != NULL) {
+        // jumptable
+        for (j = 0; j < po->btj->count; j++) {
+          check_i(po, po->btj->d[j].bt_i);
+          scan_propagate_df(po->btj->d[j].bt_i, opcnt);
+        }
+        return;
+      }
+
+      if (po->flags & OPF_RMD)
+        continue;
+      check_i(po, po->bt_i);
+      if (po->flags & OPF_CJMP)
+        scan_propagate_df(po->bt_i, opcnt);
+      else
+        i = po->bt_i - 1;
+      continue;
+    }
+
+    if (po->flags & OPF_TAIL)
+      break;
+
+    if (po->op == OP_CLD) {
+      po->flags |= OPF_RMD | OPF_DONE;
+      return;
+    }
+  }
+
+  ferr(po, "missing DF clear?\n");
+}
+
+// is operand 'opr' referenced by parsed_op 'po'?
+static int is_opr_referenced(const struct parsed_opr *opr,
+  const struct parsed_op *po)
+{
+  int i, mask;
+
+  if (opr->type == OPT_REG) {
+    mask = po->regmask_dst | po->regmask_src;
+    if (po->op == OP_CALL)
+      mask |= (1 << xAX) | (1 << xCX) | (1 << xDX);
+    if ((1 << opr->reg) & mask)
+      return 1;
+    else
+      return 0;
+  }
+
+  for (i = 0; i < po->operand_cnt; i++)
+    if (IS(po->operand[0].name, opr->name))
+      return 1;
+
+  return 0;
+}
+
+// is operand 'opr' read by parsed_op 'po'?
+static int is_opr_read(const struct parsed_opr *opr,
+  const struct parsed_op *po)
+{
+  if (opr->type == OPT_REG) {
+    if (po->regmask_src & (1 << opr->reg))
+      return 1;
+    else
+      return 0;
+  }
+
+  // yes I'm lazy
+  return 0;
+}
+
+// is operand 'opr' modified by parsed_op 'po'?
+static int is_opr_modified(const struct parsed_opr *opr,
+  const struct parsed_op *po)
+{
+  int mask;
 
-    if (po->operand[0].type == OPT_REG) {
-      if (po->regmask_dst & (1 << opr->reg))
+  if (opr->type == OPT_REG) {
+    if (po->op == OP_CALL) {
+      mask = po->regmask_dst;
+      mask |= (1 << xAX) | (1 << xCX) | (1 << xDX); // ?
+      if (mask & (1 << opr->reg))
         return 1;
       else
         return 0;
     }
+
+    if (po->regmask_dst & (1 << opr->reg))
+      return 1;
+    else
+      return 0;
   }
 
   return IS(po->operand[0].name, opr->name);
@@ -2133,22 +3011,25 @@ static int scan_for_flag_set(int i, int magic, int *branched,
 
   while (i >= 0) {
     if (ops[i].cc_scratch == magic) {
-      ferr(&ops[i], "%s looped\n", __func__);
-      return -1;
+      // is this a problem?
+      //ferr(&ops[i], "%s looped\n", __func__);
+      return 0;
     }
     ops[i].cc_scratch = magic;
 
-    if (g_labels[i][0] != 0) {
+    if (g_labels[i] != NULL) {
       *branched = 1;
 
       lr = &g_label_refs[i];
       for (; lr->next; lr = lr->next) {
+        check_i(&ops[i], lr->i);
         ret = scan_for_flag_set(lr->i, magic,
                 branched, setters, setter_cnt);
         if (ret < 0)
           return ret;
       }
 
+      check_i(&ops[i], lr->i);
       if (i > 0 && LAST_OP(i - 1)) {
         i = lr->i;
         continue;
@@ -2166,7 +3047,7 @@ static int scan_for_flag_set(int i, int magic, int *branched,
       return 0;
     }
 
-    if ((ops[i].flags & OPF_JMP) && !(ops[i].flags & OPF_CC))
+    if ((ops[i].flags & (OPF_JMP|OPF_CJMP)) == OPF_JMP)
       return -1;
   }
 
@@ -2177,7 +3058,7 @@ static int scan_for_flag_set(int i, int magic, int *branched,
 static int scan_for_cdq_edx(int i)
 {
   while (i >= 0) {
-    if (g_labels[i][0] != 0) {
+    if (g_labels[i] != NULL) {
       if (g_label_refs[i].next != NULL)
         return -1;
       if (i > 0 && LAST_OP(i - 1)) {
@@ -2201,7 +3082,7 @@ static int scan_for_cdq_edx(int i)
 static int scan_for_reg_clear(int i, int reg)
 {
   while (i >= 0) {
-    if (g_labels[i][0] != 0) {
+    if (g_labels[i] != NULL) {
       if (g_label_refs[i].next != NULL)
         return -1;
       if (i > 0 && LAST_OP(i - 1)) {
@@ -2225,19 +3106,44 @@ static int scan_for_reg_clear(int i, int reg)
   return -1;
 }
 
+static void patch_esp_adjust(struct parsed_op *po, int adj)
+{
+  ferr_assert(po, po->op == OP_ADD);
+  ferr_assert(po, IS(opr_name(po, 0), "esp"));
+  ferr_assert(po, po->operand[1].type == OPT_CONST);
+
+  // this is a bit of a hack, but deals with use of
+  // single adj for multiple calls
+  po->operand[1].val -= adj;
+  po->flags |= OPF_RMD;
+  if (po->operand[1].val == 0)
+    po->flags |= OPF_DONE;
+  ferr_assert(po, (int)po->operand[1].val >= 0);
+}
+
 // scan for positive, constant esp adjust
-static int scan_for_esp_adjust(int i, int opcnt, int *adj)
+// multipath case is preliminary
+static int scan_for_esp_adjust(int i, int opcnt,
+  int adj_expect, int *adj, int *is_multipath, int do_update)
 {
-  const struct parsed_proto *pp;
+  int adj_expect_unknown = 0;
   struct parsed_op *po;
   int first_pop = -1;
-  *adj = 0;
+  int adj_best = 0;
 
-  for (; i < opcnt; i++) {
-    po = &ops[i];
+  *adj = *is_multipath = 0;
+  if (adj_expect < 0) {
+    adj_expect_unknown = 1;
+    adj_expect = 32 * 4; // enough?
+  }
 
-    if (g_labels[i][0] != 0)
-      break;
+  for (; i < opcnt && *adj < adj_expect; i++) {
+    if (g_labels[i] != NULL)
+      *is_multipath = 1;
+
+    po = &ops[i];
+    if (po->flags & OPF_DONE)
+      continue;
 
     if (po->op == OP_ADD && po->operand[0].reg == xSP) {
       if (po->operand[1].type != OPT_CONST)
@@ -2245,34 +3151,59 @@ static int scan_for_esp_adjust(int i, int opcnt, int *adj)
       *adj += po->operand[1].val;
       if (*adj & 3)
         ferr(&ops[i], "unaligned esp adjust: %x\n", *adj);
+      if (do_update) {
+        if (!*is_multipath)
+          patch_esp_adjust(po, adj_expect);
+        else
+          po->flags |= OPF_RMD;
+      }
       return i;
     }
     else if (po->op == OP_PUSH) {
-      if (first_pop == -1)
-        first_pop = -2; // none
+      //if (first_pop == -1)
+      //  first_pop = -2; // none
       *adj -= lmod_bytes(po, po->operand[0].lmod);
     }
     else if (po->op == OP_POP) {
-      if (first_pop == -1)
-        first_pop = i;
+      if (!(po->flags & OPF_DONE)) {
+        // seems like msvc only uses 'pop ecx' for stack realignment..
+        if (po->operand[0].type != OPT_REG || po->operand[0].reg != xCX)
+          break;
+        if (first_pop == -1 && *adj >= 0)
+          first_pop = i;
+      }
+      if (do_update && *adj >= 0) {
+        po->flags |= OPF_RMD;
+        if (!*is_multipath)
+          po->flags |= OPF_DONE | OPF_NOREGS;
+      }
+
       *adj += lmod_bytes(po, po->operand[0].lmod);
+      if (*adj > adj_best)
+        adj_best = *adj;
     }
     else if (po->flags & (OPF_JMP|OPF_TAIL)) {
+      if (po->op == OP_JMP && po->btj == NULL) {
+        if (po->bt_i <= i)
+          break;
+        i = po->bt_i - 1;
+        continue;
+      }
       if (po->op != OP_CALL)
         break;
       if (po->operand[0].type != OPT_LABEL)
         break;
-      pp = po->datap;
-      if (pp != NULL && pp->is_stdcall)
+      if (po->pp != NULL && po->pp->is_stdcall)
+        break;
+      if (adj_expect_unknown && first_pop >= 0)
         break;
+      // assume it's another cdecl call
     }
   }
 
-  if (*adj == 4 && first_pop >= 0 && ops[first_pop].op == OP_POP
-    && ops[first_pop].operand[0].type == OPT_REG
-    && ops[first_pop].operand[0].reg == xCX)
-  {
-    // probably 'pop ecx' was used..
+  if (first_pop >= 0) {
+    // probably only 'pop ecx' was used
+    *adj = adj_best;
     return first_pop;
   }
 
@@ -2303,7 +3234,7 @@ static void scan_fwd_set_flags(int i, int opcnt, int magic, int flags)
       }
 
       scan_fwd_set_flags(po->bt_i, opcnt, magic, flags);
-      if (!(po->flags & OPF_CC))
+      if (!(po->flags & OPF_CJMP))
         return;
     }
     if (po->flags & OPF_TAIL)
@@ -2326,8 +3257,11 @@ static const struct parsed_proto *try_recover_pp(
     int stack_ra = 0;
     int offset = 0;
 
+    if (g_header_mode)
+      return NULL;
+
     parse_stack_access(po, opr->name, ofs_reg,
-      &offset, &stack_ra, NULL);
+      &offset, &stack_ra, NULL, 0);
     if (ofs_reg[0] != 0)
       ferr(po, "offset reg on arg access?\n");
     if (offset <= stack_ra) {
@@ -2358,38 +3292,49 @@ static const struct parsed_proto *try_recover_pp(
     p = strchr(opr->name + 1, '[');
     memcpy(buf, opr->name, p - opr->name);
     buf[p - opr->name] = 0;
-    pp = proto_parse(g_fhdr, buf, 0);
+    pp = proto_parse(g_fhdr, buf, g_quiet_pp);
   }
   else if (opr->type == OPT_OFFSET || opr->type == OPT_LABEL) {
-    pp = proto_parse(g_fhdr, opr->name, 0);
-    if (pp == NULL)
-      ferr(po, "proto_parse failed for icall from '%s'\n", opr->name);
-    check_func_pp(po, pp, "reg-fptr ref");
-  }
+    pp = proto_parse(g_fhdr, opr->name, g_quiet_pp);
+    if (pp == NULL) {
+      if (!g_header_mode)
+        ferr(po, "proto_parse failed for icall to '%s'\n", opr->name);
+    }
+    else
+      check_func_pp(po, pp, "reg-fptr ref");
+  }
 
   return pp;
 }
 
 static void scan_for_call_type(int i, const struct parsed_opr *opr,
-  int magic, const struct parsed_proto **pp_found, int *multi)
+  int magic, const struct parsed_proto **pp_found, int *pp_i,
+  int *multi)
 {
   const struct parsed_proto *pp = NULL;
   struct parsed_op *po;
   struct label_ref *lr;
 
-  while (i >= 0) {
-    if (ops[i].cc_scratch == magic)
-      return;
-    ops[i].cc_scratch = magic;
+  ops[i].cc_scratch = magic;
 
-    if (g_labels[i][0] != 0) {
+  while (1) {
+    if (g_labels[i] != NULL) {
       lr = &g_label_refs[i];
-      for (; lr != NULL; lr = lr->next)
-        scan_for_call_type(lr->i, opr, magic, pp_found, multi);
+      for (; lr != NULL; lr = lr->next) {
+        check_i(&ops[i], lr->i);
+        scan_for_call_type(lr->i, opr, magic, pp_found, pp_i, multi);
+      }
       if (i > 0 && LAST_OP(i - 1))
         return;
     }
+
     i--;
+    if (i < 0)
+      break;
+
+    if (ops[i].cc_scratch == magic)
+      return;
+    ops[i].cc_scratch = magic;
 
     if (!(ops[i].flags & OPF_DATA))
       continue;
@@ -2410,7 +3355,7 @@ static void scan_for_call_type(int i, const struct parsed_opr *opr,
 
   if (i < 0) {
     // reached the top - can only be an arg-reg
-    if (opr->type != OPT_REG)
+    if (opr->type != OPT_REG || g_func_pp == NULL)
       return;
 
     for (i = 0; i < g_func_pp->argc; i++) {
@@ -2442,363 +3387,277 @@ static void scan_for_call_type(int i, const struct parsed_opr *opr,
     }
     *multi = 1;
   }
-  if (pp != NULL)
+  if (pp != NULL) {
     *pp_found = pp;
+    *pp_i = po - ops;
+  }
 }
 
-static const struct parsed_proto *resolve_icall(int i, int opcnt,
-  int *multi_src)
+static void add_label_ref(struct label_ref *lr, int op_i)
 {
-  const struct parsed_proto *pp = NULL;
-  int search_advice = 0;
-
-  *multi_src = 0;
+  struct label_ref *lr_new;
 
-  switch (ops[i].operand[0].type) {
-  case OPT_REGMEM:
-  case OPT_LABEL:
-  case OPT_OFFSET:
-    pp = try_recover_pp(&ops[i], &ops[i].operand[0], &search_advice);
-    if (!search_advice)
-      break;
-    // fallthrough
-  default:
-    scan_for_call_type(i, &ops[i].operand[0], i + opcnt * 9, &pp,
-      multi_src);
-    break;
+  if (lr->i == -1) {
+    lr->i = op_i;
+    return;
   }
 
-  return pp;
+  lr_new = calloc(1, sizeof(*lr_new));
+  lr_new->i = op_i;
+  lr_new->next = lr->next;
+  lr->next = lr_new;
 }
 
-static int collect_call_args_r(struct parsed_op *po, int i,
-  struct parsed_proto *pp, int *regmask, int *save_arg_vars, int arg,
-  int magic, int need_op_saving, int may_reuse)
+static struct parsed_data *try_resolve_jumptab(int i, int opcnt)
 {
-  struct parsed_proto *pp_tmp;
-  struct label_ref *lr;
-  int need_to_save_current;
-  int ret = 0;
-  int j;
+  struct parsed_op *po = &ops[i];
+  struct parsed_data *pd;
+  char label[NAMELEN], *p;
+  int len, j, l;
 
-  if (i < 0) {
-    ferr(po, "dead label encountered\n");
-    return -1;
-  }
+  p = strchr(po->operand[0].name, '[');
+  if (p == NULL)
+    return NULL;
 
-  for (; arg < pp->argc; arg++)
-    if (pp->arg[arg].reg == NULL)
-      break;
-  magic = (magic & 0xffffff) | (arg << 24);
+  len = p - po->operand[0].name;
+  strncpy(label, po->operand[0].name, len);
+  label[len] = 0;
 
-  for (j = i; j >= 0 && (arg < pp->argc || pp->is_unresolved); )
-  {
-    if (((ops[j].cc_scratch ^ magic) & 0xffffff) == 0) {
-      if (ops[j].cc_scratch != magic) {
-        ferr(&ops[j], "arg collect hit same path with diff args for %s\n",
-           pp->name);
-        return -1;
-      }
-      // ok: have already been here
-      return 0;
+  for (j = 0, pd = NULL; j < g_func_pd_cnt; j++) {
+    if (IS(g_func_pd[j].label, label)) {
+      pd = &g_func_pd[j];
+      break;
     }
-    ops[j].cc_scratch = magic;
+  }
+  if (pd == NULL)
+    //ferr(po, "label '%s' not parsed?\n", label);
+    return NULL;
 
-    if (g_labels[j][0] != 0 && g_label_refs[j].i != -1) {
-      lr = &g_label_refs[j];
-      if (lr->next != NULL)
-        need_op_saving = 1;
-      for (; lr->next; lr = lr->next) {
-        if ((ops[lr->i].flags & (OPF_JMP|OPF_CC)) != OPF_JMP)
-          may_reuse = 1;
-        ret = collect_call_args_r(po, lr->i, pp, regmask, save_arg_vars,
-                arg, magic, need_op_saving, may_reuse);
-        if (ret < 0)
-          return ret;
-      }
+  if (pd->type != OPT_OFFSET)
+    ferr(po, "label '%s' with non-offset data?\n", label);
 
-      if ((ops[lr->i].flags & (OPF_JMP|OPF_CC)) != OPF_JMP)
-        may_reuse = 1;
-      if (j > 0 && LAST_OP(j - 1)) {
-        // follow last branch in reverse
-        j = lr->i;
-        continue;
+  // find all labels, link
+  for (j = 0; j < pd->count; j++) {
+    for (l = 0; l < opcnt; l++) {
+      if (g_labels[l] != NULL && IS(g_labels[l], pd->d[j].u.label)) {
+        add_label_ref(&g_label_refs[l], i);
+        pd->d[j].bt_i = l;
+        break;
       }
-      need_op_saving = 1;
-      ret = collect_call_args_r(po, lr->i, pp, regmask, save_arg_vars,
-               arg, magic, need_op_saving, may_reuse);
-      if (ret < 0)
-        return ret;
     }
-    j--;
+  }
 
-    if (ops[j].op == OP_CALL)
-    {
-      if (pp->is_unresolved)
-        break;
+  return pd;
+}
 
-      pp_tmp = ops[j].datap;
-      if (pp_tmp == NULL)
-        ferr(po, "arg collect hit unparsed call '%s'\n",
-          ops[j].operand[0].name);
-      if (may_reuse && pp_tmp->argc_stack > 0)
-        ferr(po, "arg collect %d/%d hit '%s' with %d stack args\n",
-          arg, pp->argc, opr_name(&ops[j], 0), pp_tmp->argc_stack);
-    }
-    else if (ops[j].op == OP_ADD && ops[j].operand[0].reg == xSP) {
-      if (pp->is_unresolved)
-        break;
+static void clear_labels(int count)
+{
+  int i;
 
-      ferr(po, "arg collect %d/%d hit esp adjust\n",
-        arg, pp->argc);
+  for (i = 0; i < count; i++) {
+    if (g_labels[i] != NULL) {
+      free(g_labels[i]);
+      g_labels[i] = NULL;
     }
-    else if (ops[j].op == OP_POP) {
-      if (pp->is_unresolved)
-        break;
+  }
+}
 
-      ferr(po, "arg collect %d/%d hit pop\n", arg, pp->argc);
-    }
-    else if ((ops[j].flags & (OPF_JMP|OPF_CC)) == (OPF_JMP|OPF_CC))
-    {
-      if (pp->is_unresolved)
-        break;
+static int get_pp_arg_regmask_src(const struct parsed_proto *pp)
+{
+  int regmask = 0;
+  int i, reg;
 
-      may_reuse = 1;
+  for (i = 0; i < pp->argc; i++) {
+    if (pp->arg[i].reg != NULL) {
+      reg = char_array_i(regs_r32,
+              ARRAY_SIZE(regs_r32), pp->arg[i].reg);
+      if (reg < 0)
+        ferr(ops, "arg '%s' of func '%s' is not a reg?\n",
+          pp->arg[i].reg, pp->name);
+      regmask |= 1 << reg;
     }
-    else if (ops[j].op == OP_PUSH && !(ops[j].flags & OPF_FARG))
-    {
-      if (pp->is_unresolved && (ops[j].flags & OPF_RMD))
-        break;
-
-      pp->arg[arg].datap = &ops[j];
-      need_to_save_current = 0;
-      if (!need_op_saving) {
-        ret = scan_for_mod(&ops[j], j + 1, i, 1);
-        need_to_save_current = (ret >= 0);
-      }
-      if (need_op_saving || need_to_save_current) {
-        // mark this push as one that needs operand saving
-        ops[j].flags &= ~OPF_RMD;
-        if (ops[j].argnum == 0) {
-          ops[j].argnum = arg + 1;
-          *save_arg_vars |= 1 << arg;
-        }
-        else if (ops[j].argnum < arg + 1)
-          ferr(&ops[j], "argnum conflict (%d<%d) for '%s'\n",
-            ops[j].argnum, arg + 1, pp->name);
-      }
-      else if (ops[j].argnum == 0)
-        ops[j].flags |= OPF_RMD;
-
-      // some PUSHes are reused by different calls on other branches,
-      // but that can't happen if we didn't branch, so they
-      // can be removed from future searches (handles nested calls)
-      if (!may_reuse)
-        ops[j].flags |= OPF_FARG;
+  }
 
-      ops[j].flags &= ~OPF_RSAVE;
+  return regmask;
+}
 
-      arg++;
-      if (!pp->is_unresolved) {
-        // next arg
-        for (; arg < pp->argc; arg++)
-          if (pp->arg[arg].reg == NULL)
-            break;
+static int get_pp_arg_regmask_dst(const struct parsed_proto *pp)
+{
+  int regmask = 0;
+  int i, reg;
+
+  if (pp->has_retreg) {
+    for (i = 0; i < pp->argc; i++) {
+      if (pp->arg[i].type.is_retreg) {
+        reg = char_array_i(regs_r32,
+                ARRAY_SIZE(regs_r32), pp->arg[i].reg);
+        ferr_assert(ops, reg >= 0);
+        regmask |= 1 << reg;
       }
-      magic = (magic & 0xffffff) | (arg << 24);
-
-      // tracking reg usage
-      if (ops[j].operand[0].type == OPT_REG)
-        *regmask |= 1 << ops[j].operand[0].reg;
     }
   }
 
-  if (arg < pp->argc) {
-    ferr(po, "arg collect failed for '%s': %d/%d\n",
-      pp->name, arg, pp->argc);
-    return -1;
+  if (strstr(pp->ret_type.name, "int64"))
+    return regmask | (1 << xAX) | (1 << xDX);
+  if (IS(pp->ret_type.name, "float")
+   || IS(pp->ret_type.name, "double"))
+  {
+    return regmask | mxST0;
   }
+  if (strcasecmp(pp->ret_type.name, "void") == 0)
+    return regmask;
 
-  return arg;
+  return regmask | mxAX;
 }
 
-static int collect_call_args(struct parsed_op *po, int i,
-  struct parsed_proto *pp, int *regmask, int *save_arg_vars,
-  int magic)
+static void resolve_branches_parse_calls(int opcnt)
 {
+  static const struct {
+    const char *name;
+    enum op_op op;
+    unsigned int flags;
+    unsigned int regmask_src;
+    unsigned int regmask_dst;
+  } pseudo_ops[] = {
+    { "__ftol", OPP_FTOL, OPF_FPOP, mxST0, mxAX | mxDX },
+  };
+  const struct parsed_proto *pp_c;
+  struct parsed_proto *pp;
+  struct parsed_data *pd;
+  struct parsed_op *po;
+  const char *tmpname;
+  int i, l;
   int ret;
-  int a;
-
-  ret = collect_call_args_r(po, i, pp, regmask, save_arg_vars,
-          0, magic, 0, 0);
-  if (ret < 0)
-    return ret;
 
-  if (pp->is_unresolved) {
-    pp->argc += ret;
-    pp->argc_stack += ret;
-    for (a = 0; a < pp->argc; a++)
-      if (pp->arg[a].type.name == NULL)
-        pp->arg[a].type.name = strdup("int");
-  }
+  for (i = 0; i < opcnt; i++)
+  {
+    po = &ops[i];
+    po->bt_i = -1;
+    po->btj = NULL;
 
-  return ret;
-}
+    if (po->datap != NULL) {
+      pp = calloc(1, sizeof(*pp));
+      my_assert_not(pp, NULL);
 
-static void pp_insert_reg_arg(struct parsed_proto *pp, const char *reg)
-{
-  int i;
+      ret = parse_protostr(po->datap, pp);
+      if (ret < 0)
+        ferr(po, "bad protostr supplied: %s\n", (char *)po->datap);
+      free(po->datap);
+      po->datap = NULL;
+      po->pp = pp;
+    }
 
-  for (i = 0; i < pp->argc; i++)
-    if (pp->arg[i].reg == NULL)
-      break;
+    if (po->op == OP_CALL) {
+      pp = NULL;
 
-  if (pp->argc_stack)
-    memmove(&pp->arg[i + 1], &pp->arg[i],
-      sizeof(pp->arg[0]) * pp->argc_stack);
-  memset(&pp->arg[i], 0, sizeof(pp->arg[i]));
-  pp->arg[i].reg = strdup(reg);
-  pp->arg[i].type.name = strdup("int");
-  pp->argc++;
-  pp->argc_reg++;
-}
+      if (po->pp != NULL)
+        pp = po->pp;
+      else if (po->operand[0].type == OPT_LABEL)
+      {
+        tmpname = opr_name(po, 0);
+        if (IS_START(tmpname, "loc_"))
+          ferr(po, "call to loc_*\n");
 
-static void add_label_ref(struct label_ref *lr, int op_i)
-{
-  struct label_ref *lr_new;
+        // convert some calls to pseudo-ops
+        for (l = 0; l < ARRAY_SIZE(pseudo_ops); l++) {
+          if (!IS(tmpname, pseudo_ops[l].name))
+            continue;
 
-  if (lr->i == -1) {
-    lr->i = op_i;
-    return;
-  }
+          po->op = pseudo_ops[l].op;
+          po->operand_cnt = 0;
+          po->regmask_src = pseudo_ops[l].regmask_src;
+          po->regmask_dst = pseudo_ops[l].regmask_dst;
+          po->flags = pseudo_ops[l].flags;
+          po->flags |= po->regmask_dst ? OPF_DATA : 0;
+          break;
+        }
+        if (l < ARRAY_SIZE(pseudo_ops))
+          continue;
 
-  lr_new = calloc(1, sizeof(*lr_new));
-  lr_new->i = op_i;
-  lr_new->next = lr->next;
-  lr->next = lr_new;
-}
+        pp_c = proto_parse(g_fhdr, tmpname, g_header_mode);
+        if (!g_header_mode && pp_c == NULL)
+          ferr(po, "proto_parse failed for call '%s'\n", tmpname);
 
-static void output_std_flags(FILE *fout, struct parsed_op *po,
-  int *pfomask, const char *dst_opr_text)
-{
-  if (*pfomask & (1 << PFO_Z)) {
-    fprintf(fout, "\n  cond_z = (%s%s == 0);",
-      lmod_cast_u(po, po->operand[0].lmod), dst_opr_text);
-    *pfomask &= ~(1 << PFO_Z);
-  }
-  if (*pfomask & (1 << PFO_S)) {
-    fprintf(fout, "\n  cond_s = (%s%s < 0);",
-      lmod_cast_s(po, po->operand[0].lmod), dst_opr_text);
-    *pfomask &= ~(1 << PFO_S);
-  }
-}
+        if (pp_c != NULL) {
+          pp = proto_clone(pp_c);
+          my_assert_not(pp, NULL);
+        }
+      }
 
-static void output_pp_attrs(FILE *fout, const struct parsed_proto *pp,
-  int is_noreturn)
-{
-  if (pp->is_fastcall)
-    fprintf(fout, "__fastcall ");
-  else if (pp->is_stdcall && pp->argc_reg == 0)
-    fprintf(fout, "__stdcall ");
-  if (pp->is_noreturn || is_noreturn)
-    fprintf(fout, "noreturn ");
-}
-
-static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt)
-{
-  struct parsed_op *po, *delayed_flag_op = NULL, *tmp_op;
-  struct parsed_opr *last_arith_dst = NULL;
-  char buf1[256], buf2[256], buf3[256], cast[64];
-  const struct parsed_proto *pp_c;
-  struct parsed_proto *pp, *pp_tmp;
-  struct parsed_data *pd;
-  const char *tmpname;
-  enum parsed_flag_op pfo;
-  int save_arg_vars = 0;
-  int cmp_result_vars = 0;
-  int need_mul_var = 0;
-  int had_decl = 0;
-  int label_pending = 0;
-  int regmask_save = 0;
-  int regmask_arg = 0;
-  int regmask_now = 0;
-  int regmask_init = 0;
-  int regmask = 0;
-  int pfomask = 0;
-  int found = 0;
-  int depth = 0;
-  int no_output;
-  int i, j, l;
-  int dummy;
-  int arg;
-  int reg;
-  int ret;
+      if (pp != NULL) {
+        if (pp->is_fptr)
+          check_func_pp(po, pp, "fptr var call");
+        if (pp->is_noreturn)
+          po->flags |= OPF_TAIL;
+      }
+      po->pp = pp;
+      continue;
+    }
 
-  g_bp_frame = g_sp_frame = g_stack_fsz = 0;
-  g_stack_frame_used = 0;
+    if (!(po->flags & OPF_JMP) || po->op == OP_RET)
+      continue;
 
-  g_func_pp = proto_parse(fhdr, funcn, 0);
-  if (g_func_pp == NULL)
-    ferr(ops, "proto_parse failed for '%s'\n", funcn);
+    if (po->operand[0].type == OPT_REGMEM) {
+      pd = try_resolve_jumptab(i, opcnt);
+      if (pd == NULL)
+        goto tailcall;
 
-  fprintf(fout, "%s ", g_func_pp->ret_type.name);
-  output_pp_attrs(fout, g_func_pp, g_ida_func_attr & IDAFA_NORETURN);
-  fprintf(fout, "%s(", funcn);
+      po->btj = pd;
+      continue;
+    }
 
-  for (i = 0; i < g_func_pp->argc; i++) {
-    if (i > 0)
-      fprintf(fout, ", ");
-    if (g_func_pp->arg[i].fptr != NULL) {
-      // func pointer..
-      pp = g_func_pp->arg[i].fptr;
-      fprintf(fout, "%s (", pp->ret_type.name);
-      output_pp_attrs(fout, pp, 0);
-      fprintf(fout, "*a%d)(", i + 1);
-      for (j = 0; j < pp->argc; j++) {
-        if (j > 0)
-          fprintf(fout, ", ");
-        if (pp->arg[j].fptr)
-          ferr(ops, "nested fptr\n");
-        fprintf(fout, "%s", pp->arg[j].type.name);
+    for (l = 0; l < opcnt; l++) {
+      if (g_labels[l] != NULL
+          && IS(po->operand[0].name, g_labels[l]))
+      {
+        if (l == i + 1 && po->op == OP_JMP) {
+          // yet another alignment type..
+          po->flags |= OPF_RMD|OPF_DONE;
+          break;
+        }
+        add_label_ref(&g_label_refs[l], i);
+        po->bt_i = l;
+        break;
       }
-      fprintf(fout, ")");
     }
-    else {
-      fprintf(fout, "%s a%d", g_func_pp->arg[i].type.name, i + 1);
-    }
-    if (g_func_pp->arg[i].reg != NULL) {
-      reg = char_array_i(regs_r32,
-              ARRAY_SIZE(regs_r32), g_func_pp->arg[i].reg);
-      if (reg < 0)
-        ferr(ops, "arg '%s' is not a reg?\n", g_func_pp->arg[i].reg);
-      regmask_arg |= 1 << reg;
-    }
-  }
-  if (g_func_pp->is_vararg) {
-    if (i > 0)
-      fprintf(fout, ", ");
-    fprintf(fout, "...");
+
+    if (po->bt_i != -1 || (po->flags & OPF_RMD))
+      continue;
+
+    if (po->operand[0].type == OPT_LABEL)
+      // assume tail call
+      goto tailcall;
+
+    ferr(po, "unhandled branch\n");
+
+tailcall:
+    po->op = OP_CALL;
+    po->flags |= OPF_TAIL;
+    if (i > 0 && ops[i - 1].op == OP_POP)
+      po->flags |= OPF_ATAIL;
+    i--; // reprocess
   }
+}
 
-  fprintf(fout, ")\n{\n");
+static void scan_prologue_epilogue(int opcnt)
+{
+  int ecx_push = 0, esp_sub = 0;
+  int found;
+  int i, j, l;
 
-  // pass1:
-  // - handle ebp/esp frame, remove ops related to it
   if (ops[0].op == OP_PUSH && IS(opr_name(&ops[0], 0), "ebp")
       && ops[1].op == OP_MOV
       && IS(opr_name(&ops[1], 0), "ebp")
       && IS(opr_name(&ops[1], 1), "esp"))
   {
-    int ecx_push = 0;
-
     g_bp_frame = 1;
-    ops[0].flags |= OPF_RMD;
-    ops[1].flags |= OPF_RMD;
+    ops[0].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+    ops[1].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
     i = 2;
 
     if (ops[2].op == OP_SUB && IS(opr_name(&ops[2], 0), "esp")) {
       g_stack_fsz = opr_const(&ops[2], 1);
-      ops[2].flags |= OPF_RMD;
+      ops[2].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
       i++;
     }
     else {
@@ -2806,7 +3665,7 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt)
       i = 2;
       while (ops[i].op == OP_PUSH && IS(opr_name(&ops[i], 0), "ecx")) {
         g_stack_fsz += 4;
-        ops[i].flags |= OPF_RMD;
+        ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
         ecx_push++;
         i++;
       }
@@ -2817,9 +3676,9 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt)
           && IS(opr_name(&ops[i + 1], 0), "__alloca_probe"))
       {
         g_stack_fsz += ops[i].operand[1].val;
-        ops[i].flags |= OPF_RMD;
+        ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
         i++;
-        ops[i].flags |= OPF_RMD;
+        ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
         i++;
       }
     }
@@ -2827,929 +3686,2454 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt)
     found = 0;
     do {
       for (; i < opcnt; i++)
-        if (ops[i].op == OP_RET)
+        if (ops[i].flags & OPF_TAIL)
           break;
-      if (i == opcnt && (ops[i - 1].flags & OPF_JMP) && found)
-        break;
+      j = i - 1;
+      if (i == opcnt && (ops[j].flags & OPF_JMP)) {
+        if (ops[j].bt_i != -1 || ops[j].btj != NULL)
+          break;
+        i--;
+        j--;
+      }
 
-      if (ops[i - 1].op == OP_POP && IS(opr_name(&ops[i - 1], 0), "ebp"))
-        ops[i - 1].flags |= OPF_RMD;
+      if ((ops[j].op == OP_POP && IS(opr_name(&ops[j], 0), "ebp"))
+          || ops[j].op == OP_LEAVE)
+      {
+        ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+      }
+      else if (ops[i].op == OP_CALL && ops[i].pp != NULL
+        && ops[i].pp->is_noreturn)
+      {
+        // on noreturn, msvc sometimes cleans stack, sometimes not
+        i++;
+        found = 1;
+        continue;
+      }
       else if (!(g_ida_func_attr & IDAFA_NORETURN))
-        ferr(&ops[i - 1], "'pop ebp' expected\n");
+        ferr(&ops[j], "'pop ebp' expected\n");
 
       if (g_stack_fsz != 0) {
-        if (ops[i - 2].op == OP_MOV
-            && IS(opr_name(&ops[i - 2], 0), "esp")
-            && IS(opr_name(&ops[i - 2], 1), "ebp"))
+        if (ops[j].op == OP_LEAVE)
+          j--;
+        else if (ops[j].op == OP_POP
+            && ops[j - 1].op == OP_MOV
+            && IS(opr_name(&ops[j - 1], 0), "esp")
+            && IS(opr_name(&ops[j - 1], 1), "ebp"))
         {
-          ops[i - 2].flags |= OPF_RMD;
+          ops[j - 1].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+          j -= 2;
         }
         else if (!(g_ida_func_attr & IDAFA_NORETURN))
-          ferr(&ops[i - 2], "esp restore expected\n");
+        {
+          ferr(&ops[j], "esp restore expected\n");
+        }
 
-        if (ecx_push && ops[i - 3].op == OP_POP
-          && IS(opr_name(&ops[i - 3], 0), "ecx"))
+        if (ecx_push && j >= 0 && ops[j].op == OP_POP
+          && IS(opr_name(&ops[j], 0), "ecx"))
         {
-          ferr(&ops[i - 3], "unexpected ecx pop\n");
+          ferr(&ops[j], "unexpected ecx pop\n");
         }
       }
 
       found = 1;
       i++;
     } while (i < opcnt);
+
+    if (!found)
+      ferr(ops, "missing ebp epilogue\n");
+    return;
+  }
+
+  // non-bp frame
+  i = 0;
+  while (ops[i].op == OP_PUSH && IS(opr_name(&ops[i], 0), "ecx")) {
+    ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+    g_stack_fsz += 4;
+    ecx_push++;
+    i++;
   }
-  else {
-    for (i = 0; i < opcnt; i++) {
-      if (ops[i].op == OP_PUSH || (ops[i].flags & (OPF_JMP|OPF_TAIL)))
-        break;
-      if (ops[i].op == OP_SUB && ops[i].operand[0].reg == xSP
-        && ops[i].operand[1].type == OPT_CONST)
-      {
-        g_sp_frame = 1;
-        break;
-      }
-    }
 
-    if (g_sp_frame)
+  for (; i < opcnt; i++) {
+    if (ops[i].op == OP_PUSH || (ops[i].flags & (OPF_JMP|OPF_TAIL)))
+      break;
+    if (ops[i].op == OP_SUB && ops[i].operand[0].reg == xSP
+      && ops[i].operand[1].type == OPT_CONST)
     {
       g_stack_fsz = ops[i].operand[1].val;
-      ops[i].flags |= OPF_RMD;
+      ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+      esp_sub = 1;
+      break;
+    }
+  }
 
-      i++;
-      do {
-        for (; i < opcnt; i++)
-          if (ops[i].op == OP_RET)
-            break;
-        if (ops[i - 1].op != OP_ADD
-            || !IS(opr_name(&ops[i - 1], 0), "esp")
-            || ops[i - 1].operand[1].type != OPT_CONST
-            || ops[i - 1].operand[1].val != g_stack_fsz)
-          ferr(&ops[i - 1], "'add esp' expected\n");
-        ops[i - 1].flags |= OPF_RMD;
+  if (ecx_push && !esp_sub) {
+    // could actually be args for a call..
+    for (; i < opcnt; i++)
+      if (ops[i].op != OP_PUSH)
+        break;
+
+    if (ops[i].op == OP_CALL && ops[i].operand[0].type == OPT_LABEL) {
+      const struct parsed_proto *pp;
+      pp = proto_parse(g_fhdr, opr_name(&ops[i], 0), 1);
+      j = pp ? pp->argc_stack : 0;
+      while (i > 0 && j > 0) {
+        i--;
+        if (ops[i].op == OP_PUSH) {
+          ops[i].flags &= ~(OPF_RMD | OPF_DONE | OPF_NOREGS);
+          j--;
+        }
+      }
+      if (j != 0)
+        ferr(&ops[i], "unhandled prologue\n");
 
+      // recheck
+      i = g_stack_fsz = ecx_push = 0;
+      while (ops[i].op == OP_PUSH && IS(opr_name(&ops[i], 0), "ecx")) {
+        if (!(ops[i].flags & OPF_RMD))
+          break;
+        g_stack_fsz += 4;
+        ecx_push++;
         i++;
-      } while (i < opcnt);
+      }
     }
   }
 
-  // pass2:
-  // - parse calls with labels
-  // - resolve all branches
-  for (i = 0; i < opcnt; i++)
+  found = 0;
+  if (ecx_push || esp_sub)
   {
-    po = &ops[i];
-    po->bt_i = -1;
-    po->btj = NULL;
+    g_sp_frame = 1;
 
-    if (po->flags & OPF_RMD)
-      continue;
+    i++;
+    do {
+      for (; i < opcnt; i++)
+        if (ops[i].flags & OPF_TAIL)
+          break;
+      j = i - 1;
+      if (i == opcnt && (ops[j].flags & OPF_JMP)) {
+        if (ops[j].bt_i != -1 || ops[j].btj != NULL)
+          break;
+        i--;
+        j--;
+      }
 
-    if (po->op == OP_CALL) {
-      pp = NULL;
+      if (ecx_push > 0) {
+        for (l = 0; l < ecx_push; l++) {
+          if (ops[j].op == OP_POP && IS(opr_name(&ops[j], 0), "ecx"))
+            /* pop ecx */;
+          else if (ops[j].op == OP_ADD
+                   && IS(opr_name(&ops[j], 0), "esp")
+                   && ops[j].operand[1].type == OPT_CONST)
+          {
+            /* add esp, N */
+            l += ops[j].operand[1].val / 4 - 1;
+          }
+          else
+            ferr(&ops[j], "'pop ecx' expected\n");
 
-      if (po->operand[0].type == OPT_LABEL) {
-        tmpname = opr_name(po, 0);
-        if (IS_START(tmpname, "loc_"))
-          ferr(po, "call to loc_*\n");
-        pp_c = proto_parse(fhdr, tmpname, 0);
-        if (pp_c == NULL)
-          ferr(po, "proto_parse failed for call '%s'\n", tmpname);
+          ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+          j--;
+        }
+        if (l != ecx_push)
+          ferr(&ops[j], "epilogue scan failed\n");
 
-        pp = proto_clone(pp_c);
-        my_assert_not(pp, NULL);
+        found = 1;
       }
-      else if (po->datap != NULL) {
-        pp = calloc(1, sizeof(*pp));
-        my_assert_not(pp, NULL);
 
-        ret = parse_protostr(po->datap, pp);
-        if (ret < 0)
-          ferr(po, "bad protostr supplied: %s\n", (char *)po->datap);
-        free(po->datap);
-        po->datap = NULL;
-      }
+      if (esp_sub) {
+        if (ops[j].op != OP_ADD
+            || !IS(opr_name(&ops[j], 0), "esp")
+            || ops[j].operand[1].type != OPT_CONST
+            || ops[j].operand[1].val != g_stack_fsz)
+          ferr(&ops[j], "'add esp' expected\n");
 
-      if (pp != NULL) {
-        if (pp->is_fptr)
-          check_func_pp(po, pp, "fptr var call");
-        if (pp->is_noreturn)
-          po->flags |= OPF_TAIL;
-        po->datap = pp;
+        ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+        ops[j].operand[1].val = 0; // hack for stack arg scanner
+        found = 1;
       }
-      continue;
-    }
 
-    if (!(po->flags & OPF_JMP) || po->op == OP_RET)
-      continue;
+      i++;
+    } while (i < opcnt);
 
-    if (po->operand[0].type == OPT_REGMEM) {
-      char *p = strchr(po->operand[0].name, '[');
-      if (p == NULL)
-        goto tailcall;
-      ret = p - po->operand[0].name;
-      strncpy(buf1, po->operand[0].name, ret);
-      buf1[ret] = 0;
+    if (!found)
+      ferr(ops, "missing esp epilogue\n");
+  }
+}
 
-      for (j = 0, pd = NULL; j < g_func_pd_cnt; j++) {
-        if (IS(g_func_pd[j].label, buf1)) {
-          pd = &g_func_pd[j];
-          break;
-        }
-      }
-      if (pd == NULL)
-        //ferr(po, "label '%s' not parsed?\n", buf1);
-        goto tailcall;
-      if (pd->type != OPT_OFFSET)
-        ferr(po, "label '%s' with non-offset data?\n", buf1);
+// find an instruction that changed opr before i op
+// *op_i must be set to -1 by the caller
+// *is_caller is set to 1 if one source is determined to be g_func arg
+// returns 1 if found, *op_i is then set to origin
+// returns -1 if multiple origins are found
+static int resolve_origin(int i, const struct parsed_opr *opr,
+  int magic, int *op_i, int *is_caller)
+{
+  struct label_ref *lr;
+  int ret = 0;
 
-      // find all labels, link
-      for (j = 0; j < pd->count; j++) {
-        for (l = 0; l < opcnt; l++) {
-          if (g_labels[l][0] && IS(g_labels[l], pd->d[j].u.label)) {
-            add_label_ref(&g_label_refs[l], i);
-            pd->d[j].bt_i = l;
-            break;
-          }
-        }
+  if (ops[i].cc_scratch == magic)
+    return 0;
+  ops[i].cc_scratch = magic;
+
+  while (1) {
+    if (g_labels[i] != NULL) {
+      lr = &g_label_refs[i];
+      for (; lr != NULL; lr = lr->next) {
+        check_i(&ops[i], lr->i);
+        ret |= resolve_origin(lr->i, opr, magic, op_i, is_caller);
       }
+      if (i > 0 && LAST_OP(i - 1))
+        return ret;
+    }
 
-      po->btj = pd;
+    i--;
+    if (i < 0) {
+      if (is_caller != NULL)
+        *is_caller = 1;
+      return -1;
+    }
+
+    if (ops[i].cc_scratch == magic)
+      return ret;
+    ops[i].cc_scratch = magic;
+
+    if (!(ops[i].flags & OPF_DATA))
+      continue;
+    if (!is_opr_modified(opr, &ops[i]))
       continue;
+
+    if (*op_i >= 0) {
+      if (*op_i == i)
+        return ret | 1;
+
+      // XXX: could check if the other op does the same
+      return -1;
     }
 
-    for (l = 0; l < opcnt; l++) {
-      if (g_labels[l][0] && IS(po->operand[0].name, g_labels[l])) {
-        add_label_ref(&g_label_refs[l], i);
-        po->bt_i = l;
-        break;
+    *op_i = i;
+    return ret | 1;
+  }
+}
+
+// find an instruction that previously referenced opr
+// if multiple results are found - fail
+// *op_i must be set to -1 by the caller
+// returns 1 if found, *op_i is then set to referencer insn
+static int resolve_last_ref(int i, const struct parsed_opr *opr,
+  int magic, int *op_i)
+{
+  struct label_ref *lr;
+  int ret = 0;
+
+  if (ops[i].cc_scratch == magic)
+    return 0;
+  ops[i].cc_scratch = magic;
+
+  while (1) {
+    if (g_labels[i] != NULL) {
+      lr = &g_label_refs[i];
+      for (; lr != NULL; lr = lr->next) {
+        check_i(&ops[i], lr->i);
+        ret |= resolve_last_ref(lr->i, opr, magic, op_i);
       }
+      if (i > 0 && LAST_OP(i - 1))
+        return ret;
     }
 
-    if (po->bt_i != -1)
-      continue;
+    i--;
+    if (i < 0)
+      return -1;
 
-    if (po->operand[0].type == OPT_LABEL)
-      // assume tail call
-      goto tailcall;
+    if (ops[i].cc_scratch == magic)
+      return 0;
+    ops[i].cc_scratch = magic;
 
-    ferr(po, "unhandled branch\n");
+    if (!is_opr_referenced(opr, &ops[i]))
+      continue;
 
-tailcall:
-    po->op = OP_CALL;
-    po->flags |= OPF_TAIL;
-    i--; // reprocess
+    if (*op_i >= 0)
+      return -1;
+
+    *op_i = i;
+    return 1;
   }
+}
 
-  // pass3:
-  // - remove dead labels
-  // - process calls
-  for (i = 0; i < opcnt; i++)
+// find next instruction that reads opr
+// *op_i must be set to -1 by the caller
+// on return, *op_i is set to first referencer insn
+// returns 1 if exactly 1 referencer is found
+static int find_next_read(int i, int opcnt,
+  const struct parsed_opr *opr, int magic, int *op_i)
+{
+  struct parsed_op *po;
+  int j, ret = 0;
+
+  for (; i < opcnt; i++)
   {
-    if (g_labels[i][0] != 0 && g_label_refs[i].i == -1)
-      g_labels[i][0] = 0;
+    if (ops[i].cc_scratch == magic)
+      return ret;
+    ops[i].cc_scratch = magic;
 
     po = &ops[i];
-    if (po->flags & OPF_RMD)
-      continue;
+    if ((po->flags & OPF_JMP) && po->op != OP_CALL) {
+      if (po->btj != NULL) {
+        // jumptable
+        for (j = 0; j < po->btj->count; j++) {
+          check_i(po, po->btj->d[j].bt_i);
+          ret |= find_next_read(po->btj->d[j].bt_i, opcnt, opr,
+                   magic, op_i);
+        }
+        return ret;
+      }
 
-    if (po->op == OP_CALL)
-    {
-      tmpname = opr_name(po, 0);
-      pp = po->datap;
-      if (pp == NULL)
-      {
-        // indirect call
-        pp_c = resolve_icall(i, opcnt, &l);
-        if (pp_c != NULL) {
-          pp = proto_clone(pp_c);
-          my_assert_not(pp, NULL);
-          if (l)
-            // not resolved just to single func
-            pp->is_fptr = 1;
-        }
-        if (pp == NULL) {
-          pp = calloc(1, sizeof(*pp));
-          my_assert_not(pp, NULL);
-          pp->is_fptr = 1;
-          ret = scan_for_esp_adjust(i + 1, opcnt, &j);
-          if (ret < 0) {
-            if (!g_allow_regfunc)
-              ferr(po, "non-__cdecl indirect call unhandled yet\n");
-            pp->is_unresolved = 1;
-            j = 0;
-          }
-          j /= 4;
-          if (j > ARRAY_SIZE(pp->arg))
-            ferr(po, "esp adjust too large: %d\n", j);
-          pp->ret_type.name = strdup("int");
-          pp->argc = pp->argc_stack = j;
-          for (arg = 0; arg < pp->argc; arg++)
-            pp->arg[arg].type.name = strdup("int");
-        }
-        po->datap = pp;
-      }
-
-      // look for and make use of esp adjust
-      ret = -1;
-      if (!pp->is_stdcall && pp->argc_stack > 0)
-        ret = scan_for_esp_adjust(i + 1, opcnt, &j);
-      if (ret >= 0) {
-        if (pp->is_vararg) {
-          if (j / 4 < pp->argc_stack)
-            ferr(po, "esp adjust is too small: %x < %x\n",
-              j, pp->argc_stack * 4);
-          // modify pp to make it have varargs as normal args
-          arg = pp->argc;
-          pp->argc += j / 4 - pp->argc_stack;
-          for (; arg < pp->argc; arg++) {
-            pp->arg[arg].type.name = strdup("int");
-            pp->argc_stack++;
-          }
-          if (pp->argc > ARRAY_SIZE(pp->arg))
-            ferr(po, "too many args for '%s'\n", tmpname);
-        }
-        if (pp->argc_stack != j / 4)
-          ferr(po, "stack tracking failed for '%s': %x %x\n",
-            tmpname, pp->argc_stack * 4, j);
-
-        ops[ret].flags |= OPF_RMD;
-        // a bit of a hack, but deals with use of
-        // single adj for multiple calls
-        ops[ret].operand[1].val -= j;
-      }
-      else if (pp->is_vararg)
-        ferr(po, "missing esp_adjust for vararg func '%s'\n",
-          pp->name);
-
-      if (!pp->is_unresolved) {
-        // since we know the args, collect them
-        collect_call_args(po, i, pp, &regmask, &save_arg_vars,
-          i + opcnt * 2);
+      if (po->flags & OPF_RMD)
+        continue;
+      check_i(po, po->bt_i);
+      if (po->flags & OPF_CJMP) {
+        ret |= find_next_read(po->bt_i, opcnt, opr, magic, op_i);
+        if (ret < 0)
+          return ret;
       }
-
-      if (strstr(pp->ret_type.name, "int64"))
-        need_mul_var = 1;
+      else
+        i = po->bt_i - 1;
+      continue;
     }
-  }
 
-  // pass4:
-  // - find POPs for PUSHes, rm both
-  // - scan for all used registers
-  // - find flag set ops for their users
-  // - do unreselved calls
-  // - declare indirect functions
-  for (i = 0; i < opcnt; i++) {
-    po = &ops[i];
-    if (po->flags & OPF_RMD)
+    if (!is_opr_read(opr, po)) {
+      if (is_opr_modified(opr, po)
+        && (po->op == OP_CALL
+         || ((po->flags & OPF_DATA)
+           && po->operand[0].lmod == OPLM_DWORD)))
+      {
+        // it's overwritten
+        return ret;
+      }
+      if (po->flags & OPF_TAIL)
+        return ret;
       continue;
-
-    if (po->op == OP_PUSH && (po->flags & OPF_RSAVE)) {
-      reg = po->operand[0].reg;
-      if (!(regmask & (1 << reg)))
-        // not a reg save after all, rerun scan_for_pop
-        po->flags &= ~OPF_RSAVE;
-      else
-        regmask_save |= 1 << reg;
     }
 
-    if (po->op == OP_PUSH
-        && po->argnum == 0 && !(po->flags & OPF_RSAVE)
-        && po->operand[0].type == OPT_REG)
-    {
-      reg = po->operand[0].reg;
-      if (reg < 0)
-        ferr(po, "reg not set for push?\n");
+    if (*op_i >= 0)
+      return -1;
 
-      depth = 0;
-      ret = scan_for_pop(i + 1, opcnt,
-              po->operand[0].name, i + opcnt * 3, 0, &depth, 0);
-      if (ret == 1) {
-        if (depth > 1)
-          ferr(po, "too much depth: %d\n", depth);
+    *op_i = i;
+    return 1;
+  }
 
-        po->flags |= OPF_RMD;
-        scan_for_pop(i + 1, opcnt, po->operand[0].name,
-          i + opcnt * 4, 0, &depth, 1);
-        continue;
-      }
-      ret = scan_for_pop_ret(i + 1, opcnt, po->operand[0].name, 0);
-      if (ret == 0) {
-        arg = OPF_RMD;
-        if (regmask & (1 << reg)) {
-          if (regmask_save & (1 << reg))
-            ferr(po, "%s already saved?\n", po->operand[0].name);
-          arg = OPF_RSAVE;
-        }
-        po->flags |= arg;
-        scan_for_pop_ret(i + 1, opcnt, po->operand[0].name, arg);
-        continue;
-      }
-    }
+  return 0;
+}
 
-    regmask_now = po->regmask_src | po->regmask_dst;
-    if (regmask_now & (1 << xBP)) {
-      if (g_bp_frame && !(po->flags & OPF_EBP_S)) {
-        if (po->regmask_dst & (1 << xBP))
-          // compiler decided to drop bp frame and use ebp as scratch
-          scan_fwd_set_flags(i, opcnt, i + opcnt * 5, OPF_EBP_S);
-        else
-          regmask_now &= ~(1 << xBP);
-      }
-    }
+static int try_resolve_const(int i, const struct parsed_opr *opr,
+  int magic, unsigned int *val)
+{
+  int s_i = -1;
+  int ret;
 
-    regmask |= regmask_now;
+  ret = resolve_origin(i, opr, magic, &s_i, NULL);
+  if (ret == 1) {
+    i = s_i;
+    if (ops[i].op != OP_MOV && ops[i].operand[1].type != OPT_CONST)
+      return -1;
 
-    if (po->flags & OPF_CC)
-    {
-      int setters[16], cnt = 0, branched = 0;
+    *val = ops[i].operand[1].val;
+    return 1;
+  }
 
-      ret = scan_for_flag_set(i, i + opcnt * 6,
-              &branched, setters, &cnt);
-      if (ret < 0 || cnt <= 0)
-        ferr(po, "unable to trace flag setter(s)\n");
-      if (cnt > ARRAY_SIZE(setters))
-        ferr(po, "too many flag setters\n");
+  return -1;
+}
 
-      pfo = split_cond(po, po->op, &dummy);
-      for (j = 0; j < cnt; j++)
-      {
-        tmp_op = &ops[setters[j]]; // flag setter
-        pfomask = 0;
+static const struct parsed_proto *resolve_icall(int i, int opcnt,
+  int *pp_i, int *multi_src)
+{
+  const struct parsed_proto *pp = NULL;
+  int search_advice = 0;
+  int offset = -1;
+  char name[256];
+  char s_reg[4];
+  int reg, len;
+  int ret;
 
-        // to get nicer code, we try to delay test and cmp;
-        // if we can't because of operand modification, or if we
-        // have math op, or branch, make it calculate flags explicitly
-        if (tmp_op->op == OP_TEST || tmp_op->op == OP_CMP) {
-          if (branched || scan_for_mod(tmp_op, setters[j] + 1, i, 0) >= 0)
-            pfomask = 1 << pfo;
+  *multi_src = 0;
+  *pp_i = -1;
+
+  switch (ops[i].operand[0].type) {
+  case OPT_REGMEM:
+    // try to resolve struct member calls
+    ret = sscanf(ops[i].operand[0].name, "%3s+%x%n",
+            s_reg, &offset, &len);
+    if (ret == 2 && len == strlen(ops[i].operand[0].name))
+    {
+      reg = char_array_i(regs_r32, ARRAY_SIZE(regs_r32), s_reg);
+      if (reg >= 0) {
+        struct parsed_opr opr = OPR_INIT(OPT_REG, OPLM_DWORD, reg);
+        int j = -1;
+        ret = resolve_origin(i, &opr, i + opcnt * 19, &j, NULL);
+        if (ret != 1)
+          break;
+        if (ops[j].op == OP_MOV && ops[j].operand[1].type == OPT_REGMEM
+          && ops[j].operand[0].lmod == OPLM_DWORD
+          && ops[j].pp == NULL) // no hint
+        {
+          // allow one simple dereference (directx)
+          reg = char_array_i(regs_r32, ARRAY_SIZE(regs_r32),
+                  ops[j].operand[1].name);
+          if (reg < 0)
+            break;
+          struct parsed_opr opr2 = OPR_INIT(OPT_REG, OPLM_DWORD, reg);
+          int k = -1;
+          ret = resolve_origin(j, &opr2, j + opcnt * 19, &k, NULL);
+          if (ret != 1)
+            break;
+          j = k;
         }
-        else if (tmp_op->op == OP_CMPS) {
-          pfomask = 1 << PFO_Z;
+        if (ops[j].op != OP_MOV)
+          break;
+        if (ops[j].operand[0].lmod != OPLM_DWORD)
+          break;
+        if (ops[j].pp != NULL) {
+          // type hint in asm
+          pp = ops[j].pp;
         }
-        else {
-          // see if we'll be able to handle based on op result
-          if ((tmp_op->op != OP_AND && tmp_op->op != OP_OR
-               && pfo != PFO_Z && pfo != PFO_S && pfo != PFO_P)
-              || branched
-              || scan_for_mod_opr0(tmp_op, setters[j] + 1, i) >= 0)
-            pfomask = 1 << pfo;
+        else if (ops[j].operand[1].type == OPT_REGMEM) {
+          // allow 'hello[ecx]' - assume array of same type items
+          ret = sscanf(ops[j].operand[1].name, "%[^[][e%2s]",
+                  name, s_reg);
+          if (ret != 2)
+            break;
+          pp = proto_parse(g_fhdr, name, g_quiet_pp);
         }
-        if (pfomask) {
-          tmp_op->pfomask |= pfomask;
-          cmp_result_vars |= pfomask;
+        else if (ops[j].operand[1].type == OPT_LABEL)
+          pp = proto_parse(g_fhdr, ops[j].operand[1].name, g_quiet_pp);
+        else
+          break;
+        if (pp == NULL)
+          break;
+        if (pp->is_func || pp->is_fptr || !pp->type.is_struct) {
+          pp = NULL;
+          break;
         }
-        // note: may overwrite, currently not a problem
-        po->datap = tmp_op;
+        pp = proto_lookup_struct(g_fhdr, pp->type.name, offset);
       }
-
-      if (po->op == OP_ADC || po->op == OP_SBB)
-        cmp_result_vars |= 1 << PFO_C;
-    }
-    else if (po->op == OP_MUL
-      || (po->op == OP_IMUL && po->operand_cnt == 1))
-    {
-      need_mul_var = 1;
+      break;
     }
-    else if (po->op == OP_CALL) {
-      pp = po->datap;
-      if (pp == NULL)
-        ferr(po, "NULL pp\n");
-
-      if (pp->is_unresolved) {
-        int regmask_stack = 0;
-        collect_call_args(po, i, pp, &regmask, &save_arg_vars,
-          i + opcnt * 2);
+    // fallthrough
+  case OPT_LABEL:
+  case OPT_OFFSET:
+    pp = try_recover_pp(&ops[i], &ops[i].operand[0], &search_advice);
+    if (!search_advice)
+      break;
+    // fallthrough
+  default:
+    scan_for_call_type(i, &ops[i].operand[0], i + opcnt * 9, &pp,
+      pp_i, multi_src);
+    break;
+  }
 
-        // this is pretty rough guess:
-        // see ecx and edx were pushed (and not their saved versions)
-        for (arg = 0; arg < pp->argc; arg++) {
-          if (pp->arg[arg].reg != NULL)
-            continue;
+  return pp;
+}
 
-          tmp_op = pp->arg[arg].datap;
-          if (tmp_op == NULL)
-            ferr(po, "parsed_op missing for arg%d\n", arg);
-          if (tmp_op->argnum == 0 && tmp_op->operand[0].type == OPT_REG)
-            regmask_stack |= 1 << tmp_op->operand[0].reg;
-        }
+static struct parsed_proto *process_call_early(int i, int opcnt,
+  int *adj_i)
+{
+  struct parsed_op *po = &ops[i];
+  struct parsed_proto *pp;
+  int multipath = 0;
+  int adj = 0;
+  int j, ret;
+
+  pp = po->pp;
+  if (pp == NULL || pp->is_vararg || pp->argc_reg != 0)
+    // leave for later
+    return NULL;
 
-        if (!((regmask_stack & (1 << xCX))
-          && (regmask_stack & (1 << xDX))))
+  // look for and make use of esp adjust
+  *adj_i = ret = -1;
+  if (!pp->is_stdcall && pp->argc_stack > 0)
+    ret = scan_for_esp_adjust(i + 1, opcnt,
+            pp->argc_stack * 4, &adj, &multipath, 0);
+  if (ret >= 0) {
+    if (pp->argc_stack > adj / 4)
+      return NULL;
+    if (multipath)
+      return NULL;
+    if (ops[ret].op == OP_POP) {
+      for (j = 1; j < adj / 4; j++) {
+        if (ops[ret + j].op != OP_POP
+          || ops[ret + j].operand[0].reg != xCX)
         {
-          if (pp->argc_stack != 0
-           || ((regmask | regmask_arg) & ((1 << xCX)|(1 << xDX))))
-          {
-            pp_insert_reg_arg(pp, "ecx");
-            pp->is_fastcall = 1;
-            regmask_init |= 1 << xCX;
-            regmask |= 1 << xCX;
-          }
-          if (pp->argc_stack != 0
-           || ((regmask | regmask_arg) & (1 << xDX)))
-          {
-            pp_insert_reg_arg(pp, "edx");
-            regmask_init |= 1 << xDX;
-            regmask |= 1 << xDX;
-          }
+          return NULL;
         }
-
-        // note: __cdecl doesn't fall into is_unresolved category
-        if (pp->argc_stack > 0)
-          pp->is_stdcall = 1;
       }
+    }
+  }
 
-      for (arg = 0; arg < pp->argc; arg++) {
-        if (pp->arg[arg].reg != NULL) {
-          reg = char_array_i(regs_r32,
-                  ARRAY_SIZE(regs_r32), pp->arg[arg].reg);
-          if (reg < 0)
-            ferr(ops, "arg '%s' is not a reg?\n", pp->arg[arg].reg);
-          if (!(regmask & (1 << reg))) {
-            regmask_init |= 1 << reg;
-            regmask |= 1 << reg;
-          }
-        }
-      }
+  *adj_i = ret;
+  return pp;
+}
 
-      if (pp->is_fptr) {
-        if (pp->name[0] != 0) {
-          memmove(pp->name + 2, pp->name, strlen(pp->name) + 1);
-          memcpy(pp->name, "i_", 2);
+static struct parsed_proto *process_call(int i, int opcnt)
+{
+  struct parsed_op *po = &ops[i];
+  const struct parsed_proto *pp_c;
+  struct parsed_proto *pp;
+  const char *tmpname;
+  int call_i = -1, ref_i = -1;
+  int adj = 0, multipath = 0;
+  int ret, arg;
 
-          // might be declared already
-          found = 0;
-          for (j = 0; j < i; j++) {
-            if (ops[j].op == OP_CALL && (pp_tmp = ops[j].datap)) {
-              if (pp_tmp->is_fptr && IS(pp->name, pp_tmp->name)) {
-                found = 1;
-                break;
-              }
-            }
+  tmpname = opr_name(po, 0);
+  pp = po->pp;
+  if (pp == NULL)
+  {
+    // indirect call
+    pp_c = resolve_icall(i, opcnt, &call_i, &multipath);
+    if (pp_c != NULL) {
+      if (!pp_c->is_func && !pp_c->is_fptr)
+        ferr(po, "call to non-func: %s\n", pp_c->name);
+      pp = proto_clone(pp_c);
+      my_assert_not(pp, NULL);
+      if (multipath)
+        // not resolved just to single func
+        pp->is_fptr = 1;
+
+      switch (po->operand[0].type) {
+      case OPT_REG:
+        // we resolved this call and no longer need the register
+        po->regmask_src &= ~(1 << po->operand[0].reg);
+
+        if (!multipath && i != call_i && ops[call_i].op == OP_MOV
+          && ops[call_i].operand[1].type == OPT_LABEL)
+        {
+          // no other source users?
+          ret = resolve_last_ref(i, &po->operand[0], i + opcnt * 10,
+                  &ref_i);
+          if (ret == 1 && call_i == ref_i) {
+            // and nothing uses it after us?
+            ref_i = -1;
+            find_next_read(i + 1, opcnt, &po->operand[0],
+              i + opcnt * 11, &ref_i);
+            if (ref_i == -1)
+              // then also don't need the source mov
+              ops[call_i].flags |= OPF_RMD | OPF_NOREGS;
           }
-          if (found)
-            continue;
         }
-        else
-          snprintf(pp->name, sizeof(pp->name), "icall%d", i);
+        break;
+      case OPT_REGMEM:
+        pp->is_fptr = 1;
+        break;
+      default:
+        break;
+      }
+    }
+    if (pp == NULL) {
+      pp = calloc(1, sizeof(*pp));
+      my_assert_not(pp, NULL);
+
+      pp->is_fptr = 1;
+      ret = scan_for_esp_adjust(i + 1, opcnt,
+              -1, &adj, &multipath, 0);
+      if (ret < 0 || adj < 0) {
+        if (!g_allow_regfunc)
+          ferr(po, "non-__cdecl indirect call unhandled yet\n");
+        pp->is_unresolved = 1;
+        adj = 0;
+      }
+      adj /= 4;
+      if (adj > ARRAY_SIZE(pp->arg))
+        ferr(po, "esp adjust too large: %d\n", adj);
+      pp->ret_type.name = strdup("int");
+      pp->argc = pp->argc_stack = adj;
+      for (arg = 0; arg < pp->argc; arg++)
+        pp->arg[arg].type.name = strdup("int");
+    }
+    po->pp = pp;
+  }
 
-        fprintf(fout, "  %s (", pp->ret_type.name);
-        output_pp_attrs(fout, pp, 0);
-        fprintf(fout, "*%s)(", pp->name);
-        for (j = 0; j < pp->argc; j++) {
-          if (j > 0)
-            fprintf(fout, ", ");
-          fprintf(fout, "%s a%d", pp->arg[j].type.name, j + 1);
-        }
-        fprintf(fout, ");\n");
+  // look for and make use of esp adjust
+  multipath = 0;
+  ret = -1;
+  if (!pp->is_stdcall && pp->argc_stack > 0) {
+    int adj_expect = pp->is_vararg ? -1 : pp->argc_stack * 4;
+    ret = scan_for_esp_adjust(i + 1, opcnt,
+            adj_expect, &adj, &multipath, 0);
+  }
+  if (ret >= 0) {
+    if (pp->is_vararg) {
+      if (adj / 4 < pp->argc_stack) {
+        fnote(po, "(this call)\n");
+        ferr(&ops[ret], "esp adjust is too small: %x < %x\n",
+          adj, pp->argc_stack * 4);
       }
+      // modify pp to make it have varargs as normal args
+      arg = pp->argc;
+      pp->argc += adj / 4 - pp->argc_stack;
+      for (; arg < pp->argc; arg++) {
+        pp->arg[arg].type.name = strdup("int");
+        pp->argc_stack++;
+      }
+      if (pp->argc > ARRAY_SIZE(pp->arg))
+        ferr(po, "too many args for '%s'\n", tmpname);
+    }
+    if (pp->argc_stack > adj / 4) {
+      fnote(po, "(this call)\n");
+      ferr(&ops[ret], "stack tracking failed for '%s': %x %x\n",
+        tmpname, pp->argc_stack * 4, adj);
     }
-    else if (po->op == OP_RET && !IS(g_func_pp->ret_type.name, "void"))
-      regmask |= 1 << xAX;
+
+    scan_for_esp_adjust(i + 1, opcnt,
+      pp->argc_stack * 4, &adj, &multipath, 1);
   }
+  else if (pp->is_vararg)
+    ferr(po, "missing esp_adjust for vararg func '%s'\n",
+      pp->name);
 
-  // pass4:
-  // - confirm regmask_save, it might have been reduced
-  if (regmask_save != 0)
+  return pp;
+}
+
+static int collect_call_args_early(struct parsed_op *po, int i,
+  struct parsed_proto *pp, int *regmask)
+{
+  int arg, ret;
+  int j;
+
+  for (arg = 0; arg < pp->argc; arg++)
+    if (pp->arg[arg].reg == NULL)
+      break;
+
+  // first see if it can be easily done
+  for (j = i; j > 0 && arg < pp->argc; )
   {
-    regmask_save = 0;
-    for (i = 0; i < opcnt; i++) {
-      po = &ops[i];
-      if (po->flags & OPF_RMD)
-        continue;
+    if (g_labels[j] != NULL)
+      return -1;
+    j--;
+
+    if (ops[j].op == OP_CALL)
+      return -1;
+    else if (ops[j].op == OP_ADD && ops[j].operand[0].reg == xSP)
+      return -1;
+    else if (ops[j].op == OP_POP)
+      return -1;
+    else if (ops[j].flags & OPF_CJMP)
+      return -1;
+    else if (ops[j].op == OP_PUSH) {
+      if (ops[j].flags & (OPF_FARG|OPF_FARGNR))
+        return -1;
+      ret = scan_for_mod(&ops[j], j + 1, i, 1);
+      if (ret >= 0)
+        return -1;
+
+      if (pp->arg[arg].type.is_va_list)
+        return -1;
 
-      if (po->op == OP_PUSH && (po->flags & OPF_RSAVE))
-        regmask_save |= 1 << po->operand[0].reg;
+      // next arg
+      for (arg++; arg < pp->argc; arg++)
+        if (pp->arg[arg].reg == NULL)
+          break;
     }
   }
 
+  if (arg < pp->argc)
+    return -1;
 
-  // output LUTs/jumptables
-  for (i = 0; i < g_func_pd_cnt; i++) {
-    pd = &g_func_pd[i];
-    fprintf(fout, "  static const ");
-    if (pd->type == OPT_OFFSET) {
-      fprintf(fout, "void *jt_%s[] =\n    { ", pd->label);
+  // now do it
+  for (arg = 0; arg < pp->argc; arg++)
+    if (pp->arg[arg].reg == NULL)
+      break;
 
-      for (j = 0; j < pd->count; j++) {
-        if (j > 0)
-          fprintf(fout, ", ");
-        fprintf(fout, "&&%s", pd->d[j].u.label);
-      }
-    }
-    else {
-      fprintf(fout, "%s %s[] =\n    { ",
-        lmod_type_u(ops, pd->lmod), pd->label);
+  for (j = i; j > 0 && arg < pp->argc; )
+  {
+    j--;
 
-      for (j = 0; j < pd->count; j++) {
-        if (j > 0)
-          fprintf(fout, ", ");
-        fprintf(fout, "%u", pd->d[j].u.val);
-      }
-    }
-    fprintf(fout, " };\n");
-  }
+    if (ops[j].op == OP_PUSH)
+    {
+      ops[j].p_argnext = -1;
+      ferr_assert(&ops[j], pp->arg[arg].datap == NULL);
+      pp->arg[arg].datap = &ops[j];
 
-  // declare stack frame, va_arg
-  if (g_stack_fsz)
-    fprintf(fout, "  union { u32 d[%d]; u16 w[%d]; u8 b[%d]; } sf;\n",
-      (g_stack_fsz + 3) / 4, (g_stack_fsz + 1) / 2, g_stack_fsz);
+      if (ops[j].operand[0].type == OPT_REG)
+        *regmask |= 1 << ops[j].operand[0].reg;
 
-  if (g_func_pp->is_vararg)
-    fprintf(fout, "  va_list ap;\n");
+      ops[j].flags |= OPF_RMD | OPF_DONE | OPF_FARGNR | OPF_FARG;
+      ops[j].flags &= ~OPF_RSAVE;
 
-  // declare arg-registers
-  for (i = 0; i < g_func_pp->argc; i++) {
-    if (g_func_pp->arg[i].reg != NULL) {
-      reg = char_array_i(regs_r32,
-              ARRAY_SIZE(regs_r32), g_func_pp->arg[i].reg);
-      if (regmask & (1 << reg)) {
-        fprintf(fout, "  u32 %s = (u32)a%d;\n",
-          g_func_pp->arg[i].reg, i + 1);
-      }
-      else
-        fprintf(fout, "  // %s = a%d; // unused\n",
-          g_func_pp->arg[i].reg, i + 1);
-      had_decl = 1;
+      // next arg
+      for (arg++; arg < pp->argc; arg++)
+        if (pp->arg[arg].reg == NULL)
+          break;
     }
   }
 
-  regmask_now = regmask & ~regmask_arg;
-  regmask_now &= ~(1 << xSP);
-  if (regmask_now) {
-    for (reg = 0; reg < 8; reg++) {
-      if (regmask_now & (1 << reg)) {
-        fprintf(fout, "  u32 %s", regs_r32[reg]);
-        if (regmask_init & (1 << reg))
-          fprintf(fout, " = 0");
-        fprintf(fout, ";\n");
-        had_decl = 1;
-      }
-    }
+  return 0;
+}
+
+static int collect_call_args_r(struct parsed_op *po, int i,
+  struct parsed_proto *pp, int *regmask, int *save_arg_vars,
+  int *arg_grp, int arg, int magic, int need_op_saving, int may_reuse)
+{
+  struct parsed_proto *pp_tmp;
+  struct parsed_op *po_tmp;
+  struct label_ref *lr;
+  int need_to_save_current;
+  int arg_grp_current = 0;
+  int save_args_seen = 0;
+  int save_args;
+  int ret = 0;
+  int reg;
+  char buf[32];
+  int j, k;
+
+  if (i < 0) {
+    ferr(po, "dead label encountered\n");
+    return -1;
   }
 
-  if (regmask_save) {
-    for (reg = 0; reg < 8; reg++) {
-      if (regmask_save & (1 << reg)) {
-        fprintf(fout, "  u32 s_%s;\n", regs_r32[reg]);
-        had_decl = 1;
+  for (; arg < pp->argc; arg++)
+    if (pp->arg[arg].reg == NULL)
+      break;
+  magic = (magic & 0xffffff) | (arg << 24);
+
+  for (j = i; j >= 0 && (arg < pp->argc || pp->is_unresolved); )
+  {
+    if (((ops[j].cc_scratch ^ magic) & 0xffffff) == 0) {
+      if (ops[j].cc_scratch != magic) {
+        ferr(&ops[j], "arg collect hit same path with diff args for %s\n",
+           pp->name);
+        return -1;
       }
+      // ok: have already been here
+      return 0;
     }
-  }
+    ops[j].cc_scratch = magic;
 
-  if (save_arg_vars) {
-    for (reg = 0; reg < 32; reg++) {
-      if (save_arg_vars & (1 << reg)) {
-        fprintf(fout, "  u32 s_a%d;\n", reg + 1);
-        had_decl = 1;
+    if (g_labels[j] != NULL && g_label_refs[j].i != -1) {
+      lr = &g_label_refs[j];
+      if (lr->next != NULL)
+        need_op_saving = 1;
+      for (; lr->next; lr = lr->next) {
+        check_i(&ops[j], lr->i);
+        if ((ops[lr->i].flags & (OPF_JMP|OPF_CJMP)) != OPF_JMP)
+          may_reuse = 1;
+        ret = collect_call_args_r(po, lr->i, pp, regmask, save_arg_vars,
+                arg_grp, arg, magic, need_op_saving, may_reuse);
+        if (ret < 0)
+          return ret;
       }
-    }
-  }
 
-  if (cmp_result_vars) {
-    for (i = 0; i < 8; i++) {
-      if (cmp_result_vars & (1 << i)) {
-        fprintf(fout, "  u32 cond_%s;\n", parsed_flag_op_names[i]);
-        had_decl = 1;
+      check_i(&ops[j], lr->i);
+      if ((ops[lr->i].flags & (OPF_JMP|OPF_CJMP)) != OPF_JMP)
+        may_reuse = 1;
+      if (j > 0 && LAST_OP(j - 1)) {
+        // follow last branch in reverse
+        j = lr->i;
+        continue;
       }
+      need_op_saving = 1;
+      ret = collect_call_args_r(po, lr->i, pp, regmask, save_arg_vars,
+               arg_grp, arg, magic, need_op_saving, may_reuse);
+      if (ret < 0)
+        return ret;
     }
-  }
+    j--;
 
-  if (need_mul_var) {
-    fprintf(fout, "  u64 mul_tmp;\n");
-    had_decl = 1;
-  }
+    if (ops[j].op == OP_CALL)
+    {
+      if (pp->is_unresolved)
+        break;
 
-  if (had_decl)
-    fprintf(fout, "\n");
+      pp_tmp = ops[j].pp;
+      if (pp_tmp == NULL)
+        ferr(po, "arg collect hit unparsed call '%s'\n",
+          ops[j].operand[0].name);
+      if (may_reuse && pp_tmp->argc_stack > 0)
+        ferr(po, "arg collect %d/%d hit '%s' with %d stack args\n",
+          arg, pp->argc, opr_name(&ops[j], 0), pp_tmp->argc_stack);
+    }
+    // esp adjust of 0 means we collected it before
+    else if (ops[j].op == OP_ADD && ops[j].operand[0].reg == xSP
+      && (ops[j].operand[1].type != OPT_CONST
+          || ops[j].operand[1].val != 0))
+    {
+      if (pp->is_unresolved)
+        break;
 
-  if (g_func_pp->is_vararg) {
-    if (g_func_pp->argc_stack == 0)
-      ferr(ops, "vararg func without stack args?\n");
-    fprintf(fout, "  va_start(ap, a%d);\n", g_func_pp->argc);
-  }
+      fnote(po, "(this call)\n");
+      ferr(&ops[j], "arg collect %d/%d hit esp adjust of %d\n",
+        arg, pp->argc, ops[j].operand[1].val);
+    }
+    else if (ops[j].op == OP_POP && !(ops[j].flags & OPF_DONE))
+    {
+      if (pp->is_unresolved)
+        break;
 
-  // output ops
-  for (i = 0; i < opcnt; i++)
-  {
-    if (g_labels[i][0] != 0) {
-      fprintf(fout, "\n%s:\n", g_labels[i]);
-      label_pending = 1;
+      fnote(po, "(this call)\n");
+      ferr(&ops[j], "arg collect %d/%d hit pop\n", arg, pp->argc);
+    }
+    else if (ops[j].flags & OPF_CJMP)
+    {
+      if (pp->is_unresolved)
+        break;
 
-      delayed_flag_op = NULL;
-      last_arith_dst = NULL;
+      may_reuse = 1;
     }
+    else if (ops[j].op == OP_PUSH
+      && !(ops[j].flags & (OPF_FARGNR|OPF_DONE)))
+    {
+      if (pp->is_unresolved && (ops[j].flags & OPF_RMD))
+        break;
 
-    po = &ops[i];
-    if (po->flags & OPF_RMD)
-      continue;
+      ops[j].p_argnext = -1;
+      po_tmp = pp->arg[arg].datap;
+      if (po_tmp != NULL)
+        ops[j].p_argnext = po_tmp - ops;
+      pp->arg[arg].datap = &ops[j];
 
-    no_output = 0;
+      need_to_save_current = 0;
+      save_args = 0;
+      reg = -1;
+      if (ops[j].operand[0].type == OPT_REG)
+        reg = ops[j].operand[0].reg;
 
-    #define assert_operand_cnt(n_) \
-      if (po->operand_cnt != n_) \
-        ferr(po, "operand_cnt is %d/%d\n", po->operand_cnt, n_)
+      if (!need_op_saving) {
+        ret = scan_for_mod(&ops[j], j + 1, i, 1);
+        need_to_save_current = (ret >= 0);
+      }
+      if (need_op_saving || need_to_save_current) {
+        // mark this push as one that needs operand saving
+        ops[j].flags &= ~OPF_RMD;
+        if (ops[j].p_argnum == 0) {
+          ops[j].p_argnum = arg + 1;
+          save_args |= 1 << arg;
+        }
+        else if (ops[j].p_argnum < arg + 1) {
+          // XXX: might kill valid var..
+          //*save_arg_vars &= ~(1 << (ops[j].p_argnum - 1));
+          ops[j].p_argnum = arg + 1;
+          save_args |= 1 << arg;
+        }
 
-    // conditional/flag using op?
-    if (po->flags & OPF_CC)
-    {
-      int is_delayed = 0;
-      int is_inv = 0;
+        if (save_args_seen & (1 << (ops[j].p_argnum - 1))) {
+          save_args_seen = 0;
+          arg_grp_current++;
+          if (arg_grp_current >= MAX_ARG_GRP)
+            ferr(&ops[j], "out of arg groups (arg%d), f %s\n",
+              ops[j].p_argnum, pp->name);
+        }
+      }
+      else if (ops[j].p_argnum == 0)
+        ops[j].flags |= OPF_RMD;
 
-      pfo = split_cond(po, po->op, &is_inv);
-      tmp_op = po->datap;
+      // some PUSHes are reused by different calls on other branches,
+      // but that can't happen if we didn't branch, so they
+      // can be removed from future searches (handles nested calls)
+      if (!may_reuse)
+        ops[j].flags |= OPF_FARGNR;
 
-      // we go through all this trouble to avoid using parsed_flag_op,
-      // which makes generated code much nicer
-      if (delayed_flag_op != NULL)
-      {
-        out_cmp_test(buf1, sizeof(buf1), delayed_flag_op, pfo, is_inv);
-        is_delayed = 1;
-      }
-      else if (last_arith_dst != NULL
-        && (pfo == PFO_Z || pfo == PFO_S || pfo == PFO_P
-           || (tmp_op && (tmp_op->op == OP_AND || tmp_op->op == OP_OR))
-           ))
+      ops[j].flags |= OPF_FARG;
+      ops[j].flags &= ~OPF_RSAVE;
+
+      // check for __VALIST
+      if (!pp->is_unresolved && g_func_pp != NULL
+        && pp->arg[arg].type.is_va_list)
       {
-        out_src_opr_u32(buf3, sizeof(buf3), po, last_arith_dst);
-        out_test_for_cc(buf1, sizeof(buf1), po, pfo, is_inv,
-          last_arith_dst->lmod, buf3);
-        is_delayed = 1;
+        k = -1;
+        ret = resolve_origin(j, &ops[j].operand[0],
+                magic + 1, &k, NULL);
+        if (ret == 1 && k >= 0)
+        {
+          if (ops[k].op == OP_LEA) {
+            if (!g_func_pp->is_vararg)
+              ferr(&ops[k], "lea <arg> used, but %s is not vararg?\n",
+                   g_func_pp->name);
+
+            snprintf(buf, sizeof(buf), "arg_%X",
+              g_func_pp->argc_stack * 4);
+            if (strstr(ops[k].operand[1].name, buf)
+             || strstr(ops[k].operand[1].name, "arglist"))
+            {
+              ops[k].flags |= OPF_RMD | OPF_NOREGS | OPF_DONE;
+              ops[j].flags |= OPF_RMD | OPF_NOREGS | OPF_VAPUSH;
+              save_args &= ~(1 << arg);
+              reg = -1;
+            }
+            else
+              ferr(&ops[k], "va_list arg detection failed\n");
+          }
+          // check for va_list from g_func_pp arg too
+          else if (ops[k].op == OP_MOV
+            && is_stack_access(&ops[k], &ops[k].operand[1]))
+          {
+            ret = stack_frame_access(&ops[k], &ops[k].operand[1],
+              buf, sizeof(buf), ops[k].operand[1].name, "", 1, 0);
+            if (ret >= 0) {
+              ops[k].flags |= OPF_RMD | OPF_DONE;
+              ops[j].flags |= OPF_RMD;
+              ops[j].p_argpass = ret + 1;
+              save_args &= ~(1 << arg);
+              reg = -1;
+            }
+          }
+        }
       }
-      else if (tmp_op != NULL) {
-        // use preprocessed flag calc results
-        if (!(tmp_op->pfomask & (1 << pfo)))
-          ferr(po, "not prepared for pfo %d\n", pfo);
 
-        // note: is_inv was not yet applied
-        snprintf(buf1, sizeof(buf1), "(%scond_%s)",
-          is_inv ? "!" : "", parsed_flag_op_names[pfo]);
-      }
-      else {
-        ferr(po, "all methods of finding comparison failed\n");
-      }
- 
-      if (po->flags & OPF_JMP) {
-        fprintf(fout, "  if %s\n", buf1);
-      }
-      else if (po->op == OP_ADC || po->op == OP_SBB) {
-        if (is_delayed)
-          fprintf(fout, "  cond_%s = %s;\n",
-            parsed_flag_op_names[pfo], buf1);
-      }
-      else if (po->flags & OPF_DATA) { // SETcc
-        out_dst_opr(buf2, sizeof(buf2), po, &po->operand[0]);
-        fprintf(fout, "  %s = %s;", buf2, buf1);
-      }
-      else {
-        ferr(po, "unhandled conditional op\n");
+      *save_arg_vars |= save_args;
+
+      // tracking reg usage
+      if (reg >= 0)
+        *regmask |= 1 << reg;
+
+      arg++;
+      if (!pp->is_unresolved) {
+        // next arg
+        for (; arg < pp->argc; arg++)
+          if (pp->arg[arg].reg == NULL)
+            break;
       }
+      magic = (magic & 0xffffff) | (arg << 24);
     }
 
-    pfomask = po->pfomask;
+    if (ops[j].p_arggrp > arg_grp_current) {
+      save_args_seen = 0;
+      arg_grp_current = ops[j].p_arggrp;
+    }
+    if (ops[j].p_argnum > 0)
+      save_args_seen |= 1 << (ops[j].p_argnum - 1);
+  }
 
-    switch (po->op)
-    {
-      case OP_MOV:
-        assert_operand_cnt(2);
-        propagate_lmod(po, &po->operand[0], &po->operand[1]);
-        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
-        fprintf(fout, "  %s = %s;", buf1,
-            out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
-              po->operand[0].is_ptr ? "(void *)" : "", 0));
-        break;
+  if (arg < pp->argc) {
+    ferr(po, "arg collect failed for '%s': %d/%d\n",
+      pp->name, arg, pp->argc);
+    return -1;
+  }
 
-      case OP_LEA:
-        assert_operand_cnt(2);
-        po->operand[1].lmod = OPLM_DWORD; // always
-        fprintf(fout, "  %s = %s;",
-            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
-            out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
-              NULL, 1));
-        break;
+  if (arg_grp_current > *arg_grp)
+    *arg_grp = arg_grp_current;
 
-      case OP_MOVZX:
-        assert_operand_cnt(2);
-        fprintf(fout, "  %s = %s;",
-            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
-            out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
-        break;
+  return arg;
+}
 
-      case OP_MOVSX:
-        assert_operand_cnt(2);
-        switch (po->operand[1].lmod) {
-        case OPLM_BYTE:
-          strcpy(buf3, "(s8)");
-          break;
-        case OPLM_WORD:
-          strcpy(buf3, "(s16)");
-          break;
-        default:
-          ferr(po, "invalid src lmod: %d\n", po->operand[1].lmod);
-        }
-        fprintf(fout, "  %s = %s;",
-            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
-            out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
-              buf3, 0));
-        break;
+static int collect_call_args(struct parsed_op *po, int i,
+  struct parsed_proto *pp, int *regmask, int *save_arg_vars,
+  int magic)
+{
+  // arg group is for cases when pushes for
+  // multiple funcs are going on
+  struct parsed_op *po_tmp;
+  int save_arg_vars_current = 0;
+  int arg_grp = 0;
+  int ret;
+  int a;
 
-      case OP_NOT:
-        assert_operand_cnt(1);
-        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
-        fprintf(fout, "  %s = ~%s;", buf1, buf1);
-        break;
+  ret = collect_call_args_r(po, i, pp, regmask,
+          &save_arg_vars_current, &arg_grp, 0, magic, 0, 0);
+  if (ret < 0)
+    return ret;
 
-      case OP_CDQ:
-        assert_operand_cnt(2);
-        fprintf(fout, "  %s = (s32)%s >> 31;",
-            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
-            out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
-        strcpy(g_comment, "cdq");
-        break;
+  if (arg_grp != 0) {
+    // propagate arg_grp
+    for (a = 0; a < pp->argc; a++) {
+      if (pp->arg[a].reg != NULL)
+        continue;
 
-      case OP_STOS:
-        // assumes DF=0
-        assert_operand_cnt(3);
-        if (po->flags & OPF_REP) {
-          fprintf(fout, "  for (; ecx != 0; ecx--, edi += %d)\n",
-            lmod_bytes(po, po->operand[0].lmod));
-          fprintf(fout, "    %sedi = eax;",
-            lmod_cast_u_ptr(po, po->operand[0].lmod));
-          strcpy(g_comment, "rep stos");
-        }
-        else {
-          fprintf(fout, "    %sedi = eax; edi += %d;",
-            lmod_cast_u_ptr(po, po->operand[0].lmod),
-            lmod_bytes(po, po->operand[0].lmod));
-          strcpy(g_comment, "stos");
-        }
-        break;
+      po_tmp = pp->arg[a].datap;
+      while (po_tmp != NULL) {
+        po_tmp->p_arggrp = arg_grp;
+        if (po_tmp->p_argnext > 0)
+          po_tmp = &ops[po_tmp->p_argnext];
+        else
+          po_tmp = NULL;
+      }
+    }
+  }
+  save_arg_vars[arg_grp] |= save_arg_vars_current;
 
-      case OP_MOVS:
-        // assumes DF=0
-        assert_operand_cnt(3);
-        j = lmod_bytes(po, po->operand[0].lmod);
-        strcpy(buf1, lmod_cast_u_ptr(po, po->operand[0].lmod));
-        if (po->flags & OPF_REP) {
-          fprintf(fout,
-            "  for (; ecx != 0; ecx--, edi += %d, esi += %d)\n",
-            j, j);
-          fprintf(fout,
-            "    %sedi = %sesi;", buf1, buf1);
-          strcpy(g_comment, "rep movs");
-        }
-        else {
-          fprintf(fout, "    %sedi = %sesi; edi += %d; esi += %d;",
-            buf1, buf1, j, j);
-          strcpy(g_comment, "movs");
-        }
-        break;
+  if (pp->is_unresolved) {
+    pp->argc += ret;
+    pp->argc_stack += ret;
+    for (a = 0; a < pp->argc; a++)
+      if (pp->arg[a].type.name == NULL)
+        pp->arg[a].type.name = strdup("int");
+  }
 
-      case OP_CMPS:
-        // assumes DF=0
-        // repe ~ repeat while ZF=1
-        assert_operand_cnt(3);
-        j = lmod_bytes(po, po->operand[0].lmod);
-        strcpy(buf1, lmod_cast_u_ptr(po, po->operand[0].lmod));
-        if (po->flags & OPF_REP) {
-          fprintf(fout,
-            "  for (; ecx != 0; ecx--, edi += %d, esi += %d)\n",
-            j, j);
-          fprintf(fout,
-            "    if ((cond_z = (%sedi == %sesi)) %s 0)\n",
-              buf1, buf1, (po->flags & OPF_REPZ) ? "==" : "!=");
-          fprintf(fout,
-            "      break;");
-          snprintf(g_comment, sizeof(g_comment), "rep%s cmps",
-            (po->flags & OPF_REPZ) ? "e" : "ne");
-        }
-        else {
-          fprintf(fout,
-            "    cond_z = (%sedi = %sesi); edi += %d; esi += %d;",
-            buf1, buf1, j, j);
-          strcpy(g_comment, "cmps");
-        }
-        pfomask &= ~(1 << PFO_Z);
-        last_arith_dst = NULL;
-        delayed_flag_op = NULL;
-        break;
+  return ret;
+}
 
-      // arithmetic w/flags
-      case OP_ADD:
-      case OP_SUB:
-      case OP_AND:
-      case OP_OR:
-        propagate_lmod(po, &po->operand[0], &po->operand[1]);
-        // fallthrough
-      dualop_arith:
-        assert_operand_cnt(2);
-        fprintf(fout, "  %s %s= %s;",
-            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
-            op_to_c(po),
-            out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
-        output_std_flags(fout, po, &pfomask, buf1);
-        last_arith_dst = &po->operand[0];
-        delayed_flag_op = NULL;
-        break;
+static void reg_use_pass(int i, int opcnt, unsigned char *cbits,
+  int regmask_now, int *regmask,
+  int regmask_save_now, int *regmask_save,
+  int *regmask_init, int regmask_arg)
+{
+  struct parsed_op *po;
+  unsigned int mask;
+  int already_saved;
+  int regmask_new;
+  int regmask_op;
+  int flags_set;
+  int ret, reg;
+  int j;
 
-      case OP_SHL:
-      case OP_SHR:
-        assert_operand_cnt(2);
-        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
-        if (pfomask & (1 << PFO_C)) {
-          if (po->operand[1].type == OPT_CONST) {
-            l = lmod_bytes(po, po->operand[0].lmod) * 8;
-            j = po->operand[1].val;
-            j %= l;
-            if (j != 0) {
-              if (po->op == OP_SHL)
-                j = l - j;
-              else
-                j -= 1;
-              fprintf(fout, "  cond_c = (%s & 0x%02x) ? 1 : 0;\n",
-                buf1, 1 << j);
-            }
-            else
-              ferr(po, "zero shift?\n");
-          }
-          else
-            ferr(po, "TODO\n");
-          pfomask &= ~(1 << PFO_C);
+  for (; i < opcnt; i++)
+  {
+    po = &ops[i];
+    if (cbits[i >> 3] & (1 << (i & 7)))
+      return;
+    cbits[i >> 3] |= (1 << (i & 7));
+
+    if ((po->flags & OPF_JMP) && po->op != OP_CALL) {
+      if (po->flags & (OPF_RMD|OPF_DONE))
+        continue;
+      if (po->btj != NULL) {
+        for (j = 0; j < po->btj->count; j++) {
+          check_i(po, po->btj->d[j].bt_i);
+          reg_use_pass(po->btj->d[j].bt_i, opcnt, cbits,
+            regmask_now, regmask, regmask_save_now, regmask_save,
+            regmask_init, regmask_arg);
         }
-        fprintf(fout, "  %s %s= %s;", buf1, op_to_c(po),
-            out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
-        output_std_flags(fout, po, &pfomask, buf1);
-        last_arith_dst = &po->operand[0];
-        delayed_flag_op = NULL;
-        break;
+        return;
+      }
 
-      case OP_SAR:
-        assert_operand_cnt(2);
-        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
-        fprintf(fout, "  %s = %s%s >> %s;", buf1,
-          lmod_cast_s(po, po->operand[0].lmod), buf1,
-          out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
-        output_std_flags(fout, po, &pfomask, buf1);
-        last_arith_dst = &po->operand[0];
-        delayed_flag_op = NULL;
-        break;
+      check_i(po, po->bt_i);
+      if (po->flags & OPF_CJMP)
+        reg_use_pass(po->bt_i, opcnt, cbits,
+          regmask_now, regmask, regmask_save_now, regmask_save,
+          regmask_init, regmask_arg);
+      else
+        i = po->bt_i - 1;
+      continue;
+    }
 
-      case OP_ROL:
-      case OP_ROR:
-        assert_operand_cnt(2);
-        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
-        if (po->operand[1].type == OPT_CONST) {
-          j = po->operand[1].val;
-          j %= lmod_bytes(po, po->operand[0].lmod) * 8;
-          fprintf(fout, po->op == OP_ROL ?
-            "  %s = (%s << %d) | (%s >> %d);" :
-            "  %s = (%s >> %d) | (%s << %d);",
-            buf1, buf1, j, buf1,
-            lmod_bytes(po, po->operand[0].lmod) * 8 - j);
-        }
-        else
-          ferr(po, "TODO\n");
-        output_std_flags(fout, po, &pfomask, buf1);
-        last_arith_dst = &po->operand[0];
-        delayed_flag_op = NULL;
-        break;
+    if (po->op == OP_PUSH && !(po->flags & (OPF_FARG|OPF_DONE))
+      && !g_func_pp->is_userstack
+      && po->operand[0].type == OPT_REG)
+    {
+      reg = po->operand[0].reg;
+      ferr_assert(po, reg >= 0);
 
-      case OP_XOR:
-        assert_operand_cnt(2);
-        propagate_lmod(po, &po->operand[0], &po->operand[1]);
-        if (IS(opr_name(po, 0), opr_name(po, 1))) {
-          // special case for XOR
-          fprintf(fout, "  %s = 0;",
-            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]));
-          last_arith_dst = &po->operand[0];
-          delayed_flag_op = NULL;
-          break;
+      already_saved = 0;
+      flags_set = OPF_RSAVE | OPF_RMD | OPF_DONE;
+      if (regmask_now & (1 << reg)) {
+        already_saved = regmask_save_now & (1 << reg);
+        flags_set = OPF_RSAVE | OPF_DONE;
+      }
+
+      ret = scan_for_pop(i + 1, opcnt, i + opcnt * 3, reg, 0, 0);
+      if (ret == 1) {
+        scan_for_pop(i + 1, opcnt, i + opcnt * 4, reg, 0, flags_set);
+      }
+      else {
+        ret = scan_for_pop_ret(i + 1, opcnt, po->operand[0].reg, 0);
+        if (ret == 1) {
+          scan_for_pop_ret(i + 1, opcnt, po->operand[0].reg,
+            flags_set);
         }
-        goto dualop_arith;
+      }
+      if (ret == 1) {
+        ferr_assert(po, !already_saved);
+        po->flags |= flags_set;
 
-      case OP_ADC:
-      case OP_SBB:
-        assert_operand_cnt(2);
-        propagate_lmod(po, &po->operand[0], &po->operand[1]);
-        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
-        if (po->op == OP_SBB
-          && IS(po->operand[0].name, po->operand[1].name))
-        {
-          // avoid use of unitialized var
-          fprintf(fout, "  %s = -cond_c;", buf1);
-          // carry remains what it was
-          pfomask &= ~(1 << PFO_C);
+        if (regmask_now & (1 << reg)) {
+          regmask_save_now |= (1 << reg);
+          *regmask_save |= regmask_save_now;
         }
+        continue;
+      }
+    }
+    else if (po->op == OP_POP && (po->flags & OPF_RSAVE)) {
+      reg = po->operand[0].reg;
+      ferr_assert(po, reg >= 0);
+
+      if (regmask_save_now & (1 << reg))
+        regmask_save_now &= ~(1 << reg);
+      else
+        regmask_now &= ~(1 << reg);
+      continue;
+    }
+    else if (po->op == OP_CALL) {
+      if ((po->regmask_dst & (1 << xAX))
+        && !(po->regmask_dst & (1 << xDX)))
+      {
+        if (po->flags & OPF_TAIL)
+          // don't need eax, will do "return f();" or "f(); return;"
+          po->regmask_dst &= ~(1 << xAX);
         else {
-          fprintf(fout, "  %s %s= %s + cond_c;", buf1, op_to_c(po),
-            out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
+          struct parsed_opr opr = OPR_INIT(OPT_REG, OPLM_DWORD, xAX);
+          j = -1;
+          find_next_read(i + 1, opcnt, &opr, i + opcnt * 17, &j);
+          if (j == -1)
+            // not used
+            po->regmask_dst &= ~(1 << xAX);
         }
-        output_std_flags(fout, po, &pfomask, buf1);
-        last_arith_dst = &po->operand[0];
-        delayed_flag_op = NULL;
-        break;
+      }
+    }
+
+    if (po->flags & OPF_NOREGS)
+      continue;
+
+    if (po->flags & OPF_FPUSH) {
+      if (regmask_now & mxST1)
+        ferr(po, "TODO: FPUSH on active ST1\n");
+      if (regmask_now & mxST0)
+        po->flags |= OPF_FSHIFT;
+      mask = mxST0 | mxST1;
+      regmask_now = (regmask_now & ~mask) | ((regmask_now & mxST0) << 1);
+    }
+
+    // if incomplete register is used, clear it on init to avoid
+    // later use of uninitialized upper part in some situations
+    if ((po->flags & OPF_DATA) && po->operand[0].type == OPT_REG
+        && po->operand[0].lmod != OPLM_DWORD)
+    {
+      reg = po->operand[0].reg;
+      ferr_assert(po, reg >= 0);
+
+      if (!(regmask_now & (1 << reg)))
+        *regmask_init |= 1 << reg;
+    }
+
+    regmask_op = po->regmask_src | po->regmask_dst;
+
+    regmask_new = po->regmask_src & ~regmask_now & ~regmask_arg;
+    regmask_new &= ~(1 << xSP);
+    if (g_bp_frame && !(po->flags & OPF_EBP_S))
+      regmask_new &= ~(1 << xBP);
+
+    if (regmask_new != 0)
+      fnote(po, "uninitialized reg mask: %x\n", regmask_new);
+
+    if (regmask_op & (1 << xBP)) {
+      if (g_bp_frame && !(po->flags & OPF_EBP_S)) {
+        if (po->regmask_dst & (1 << xBP))
+          // compiler decided to drop bp frame and use ebp as scratch
+          scan_fwd_set_flags(i + 1, opcnt, i + opcnt * 5, OPF_EBP_S);
+        else
+          regmask_op &= ~(1 << xBP);
+      }
+    }
+
+    regmask_now |= regmask_op;
+    *regmask |= regmask_now;
+
+    // released regs
+    if (po->flags & OPF_FPOP) {
+      mask = mxST0 | mxST1;
+      if (!(regmask_now & mask))
+        ferr(po, "float pop on empty stack?\n");
+      if (regmask_now & mxST1)
+        po->flags |= OPF_FSHIFT;
+      regmask_now = (regmask_now & ~mask) | ((regmask_now & mxST1) >> 1);
+    }
+
+    if (po->flags & OPF_TAIL) {
+      if (regmask_now & (mxST0 | mxST1))
+        ferr(po, "float regs on tail: %x\n", regmask_now);
+
+      // there is support for "conditional tailcall", sort of
+      if (!(po->flags & OPF_CC))
+        return;
+    }
+  }
+}
+
+static void pp_insert_reg_arg(struct parsed_proto *pp, const char *reg)
+{
+  int i;
+
+  for (i = 0; i < pp->argc; i++)
+    if (pp->arg[i].reg == NULL)
+      break;
+
+  if (pp->argc_stack)
+    memmove(&pp->arg[i + 1], &pp->arg[i],
+      sizeof(pp->arg[0]) * pp->argc_stack);
+  memset(&pp->arg[i], 0, sizeof(pp->arg[i]));
+  pp->arg[i].reg = strdup(reg);
+  pp->arg[i].type.name = strdup("int");
+  pp->argc++;
+  pp->argc_reg++;
+}
+
+static void output_std_flags(FILE *fout, struct parsed_op *po,
+  int *pfomask, const char *dst_opr_text)
+{
+  if (*pfomask & (1 << PFO_Z)) {
+    fprintf(fout, "\n  cond_z = (%s%s == 0);",
+      lmod_cast_u(po, po->operand[0].lmod), dst_opr_text);
+    *pfomask &= ~(1 << PFO_Z);
+  }
+  if (*pfomask & (1 << PFO_S)) {
+    fprintf(fout, "\n  cond_s = (%s%s < 0);",
+      lmod_cast_s(po, po->operand[0].lmod), dst_opr_text);
+    *pfomask &= ~(1 << PFO_S);
+  }
+}
+
+enum {
+  OPP_FORCE_NORETURN = (1 << 0),
+  OPP_SIMPLE_ARGS    = (1 << 1),
+  OPP_ALIGN          = (1 << 2),
+};
+
+static void output_pp_attrs(FILE *fout, const struct parsed_proto *pp,
+  int flags)
+{
+  const char *cconv = "";
+
+  if (pp->is_fastcall)
+    cconv = "__fastcall ";
+  else if (pp->is_stdcall && pp->argc_reg == 0)
+    cconv = "__stdcall ";
+
+  fprintf(fout, (flags & OPP_ALIGN) ? "%-16s" : "%s", cconv);
+
+  if (pp->is_noreturn || (flags & OPP_FORCE_NORETURN))
+    fprintf(fout, "noreturn ");
+}
+
+static void output_pp(FILE *fout, const struct parsed_proto *pp,
+  int flags)
+{
+  int i;
+
+  fprintf(fout, (flags & OPP_ALIGN) ? "%-5s" : "%s ",
+    pp->ret_type.name);
+  if (pp->is_fptr)
+    fprintf(fout, "(");
+  output_pp_attrs(fout, pp, flags);
+  if (pp->is_fptr)
+    fprintf(fout, "*");
+  fprintf(fout, "%s", pp->name);
+  if (pp->is_fptr)
+    fprintf(fout, ")");
+
+  fprintf(fout, "(");
+  for (i = 0; i < pp->argc; i++) {
+    if (i > 0)
+      fprintf(fout, ", ");
+    if (pp->arg[i].fptr != NULL && !(flags & OPP_SIMPLE_ARGS)) {
+      // func pointer
+      output_pp(fout, pp->arg[i].fptr, 0);
+    }
+    else if (pp->arg[i].type.is_retreg) {
+      fprintf(fout, "u32 *r_%s", pp->arg[i].reg);
+    }
+    else {
+      fprintf(fout, "%s", pp->arg[i].type.name);
+      if (!pp->is_fptr)
+        fprintf(fout, " a%d", i + 1);
+    }
+  }
+  if (pp->is_vararg) {
+    if (i > 0)
+      fprintf(fout, ", ");
+    fprintf(fout, "...");
+  }
+  fprintf(fout, ")");
+}
+
+static char *saved_arg_name(char *buf, size_t buf_size, int grp, int num)
+{
+  char buf1[16];
+
+  buf1[0] = 0;
+  if (grp > 0)
+    snprintf(buf1, sizeof(buf1), "%d", grp);
+  snprintf(buf, buf_size, "s%s_a%d", buf1, num);
+
+  return buf;
+}
+
+static void gen_x_cleanup(int opcnt);
+
+static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt)
+{
+  struct parsed_op *po, *delayed_flag_op = NULL, *tmp_op;
+  struct parsed_opr *last_arith_dst = NULL;
+  char buf1[256], buf2[256], buf3[256], cast[64];
+  struct parsed_proto *pp, *pp_tmp;
+  struct parsed_data *pd;
+  unsigned int uval;
+  int save_arg_vars[MAX_ARG_GRP] = { 0, };
+  unsigned char cbits[MAX_OPS / 8];
+  const char *float_type;
+  int cond_vars = 0;
+  int need_tmp_var = 0;
+  int need_tmp64 = 0;
+  int had_decl = 0;
+  int label_pending = 0;
+  int need_double = 0;
+  int regmask_save = 0; // regs saved/restored in this func
+  int regmask_arg;      // regs from this function args (fastcall, etc)
+  int regmask_ret;      // regs needed on ret
+  int regmask_now;      // temp
+  int regmask_init = 0; // regs that need zero initialization
+  int regmask_pp = 0;   // regs used in complex push-pop graph
+  int regmask = 0;      // used regs
+  int pfomask = 0;
+  int found = 0;
+  int no_output;
+  int i, j, l;
+  int arg;
+  int reg;
+  int ret;
+
+  g_bp_frame = g_sp_frame = g_stack_fsz = 0;
+  g_stack_frame_used = 0;
+  if (g_sct_func_attr & SCTFA_CLEAR_REGS)
+    regmask_init = g_regmask_init;
+
+  g_func_pp = proto_parse(fhdr, funcn, 0);
+  if (g_func_pp == NULL)
+    ferr(ops, "proto_parse failed for '%s'\n", funcn);
+
+  regmask_arg = get_pp_arg_regmask_src(g_func_pp);
+  regmask_ret = get_pp_arg_regmask_dst(g_func_pp);
+
+  // pass1:
+  // - resolve all branches
+  // - parse calls with labels
+  resolve_branches_parse_calls(opcnt);
+
+  // pass2:
+  // - handle ebp/esp frame, remove ops related to it
+  scan_prologue_epilogue(opcnt);
+
+  // pass3:
+  // - remove dead labels
+  // - set regs needed at ret
+  for (i = 0; i < opcnt; i++)
+  {
+    if (g_labels[i] != NULL && g_label_refs[i].i == -1) {
+      free(g_labels[i]);
+      g_labels[i] = NULL;
+    }
+
+    if (ops[i].op == OP_RET)
+      ops[i].regmask_src |= regmask_ret;
+  }
+
+  // pass4:
+  // - process trivial calls
+  for (i = 0; i < opcnt; i++)
+  {
+    po = &ops[i];
+    if (po->flags & (OPF_RMD|OPF_DONE))
+      continue;
+
+    if (po->op == OP_CALL)
+    {
+      pp = process_call_early(i, opcnt, &j);
+      if (pp != NULL) {
+        if (!(po->flags & OPF_ATAIL))
+          // since we know the args, try to collect them
+          if (collect_call_args_early(po, i, pp, &regmask) != 0)
+            pp = NULL;
+      }
+
+      if (pp != NULL) {
+        if (j >= 0) {
+          // commit esp adjust
+          if (ops[j].op != OP_POP)
+            patch_esp_adjust(&ops[j], pp->argc_stack * 4);
+          else {
+            for (l = 0; l < pp->argc_stack; l++)
+              ops[j + l].flags |= OPF_DONE | OPF_RMD | OPF_NOREGS;
+          }
+        }
+
+        if (strstr(pp->ret_type.name, "int64"))
+          need_tmp64 = 1;
+
+        po->flags |= OPF_DONE;
+      }
+    }
+  }
+
+  // pass5:
+  // - process calls, stage 2
+  // - handle some push/pop pairs
+  // - scan for STD/CLD, propagate DF
+  for (i = 0; i < opcnt; i++)
+  {
+    po = &ops[i];
+    if (po->flags & OPF_RMD)
+      continue;
+
+    if (po->op == OP_CALL)
+    {
+      if (!(po->flags & OPF_DONE)) {
+        pp = process_call(i, opcnt);
+
+        if (!pp->is_unresolved && !(po->flags & OPF_ATAIL)) {
+          // since we know the args, collect them
+          collect_call_args(po, i, pp, &regmask, save_arg_vars,
+            i + opcnt * 2);
+        }
+        // for unresolved, collect after other passes
+      }
+
+      pp = po->pp;
+      ferr_assert(po, pp != NULL);
+
+      po->regmask_src |= get_pp_arg_regmask_src(pp);
+      po->regmask_dst |= get_pp_arg_regmask_dst(pp);
+
+      if (po->regmask_dst & mxST0)
+        po->flags |= OPF_FPUSH;
+
+      if (strstr(pp->ret_type.name, "int64"))
+        need_tmp64 = 1;
+
+      continue;
+    }
+
+    if (po->flags & OPF_DONE)
+      continue;
+
+    if (po->op == OP_PUSH && !(po->flags & OPF_FARG)
+      && !(po->flags & OPF_RSAVE) && po->operand[0].type == OPT_CONST)
+    {
+      scan_for_pop_const(i, opcnt, i + opcnt * 12);
+    }
+    else if (po->op == OP_POP)
+      scan_pushes_for_pop(i, opcnt, &regmask_pp);
+    else if (po->op == OP_STD) {
+      po->flags |= OPF_DF | OPF_RMD | OPF_DONE;
+      scan_propagate_df(i + 1, opcnt);
+    }
+  }
+
+  // pass6:
+  // - find POPs for PUSHes, rm both
+  // - scan for all used registers
+  memset(cbits, 0, sizeof(cbits));
+  reg_use_pass(0, opcnt, cbits, regmask_init, &regmask,
+    0, &regmask_save, &regmask_init, regmask_arg);
+
+  // pass7:
+  // - find flag set ops for their users
+  // - do unresolved calls
+  // - declare indirect functions
+  for (i = 0; i < opcnt; i++)
+  {
+    po = &ops[i];
+    if (po->flags & (OPF_RMD|OPF_DONE))
+      continue;
+
+    if (po->flags & OPF_CC)
+    {
+      int setters[16], cnt = 0, branched = 0;
+
+      ret = scan_for_flag_set(i, i + opcnt * 6,
+              &branched, setters, &cnt);
+      if (ret < 0 || cnt <= 0)
+        ferr(po, "unable to trace flag setter(s)\n");
+      if (cnt > ARRAY_SIZE(setters))
+        ferr(po, "too many flag setters\n");
+
+      for (j = 0; j < cnt; j++)
+      {
+        tmp_op = &ops[setters[j]]; // flag setter
+        pfomask = 0;
+
+        // to get nicer code, we try to delay test and cmp;
+        // if we can't because of operand modification, or if we
+        // have arith op, or branch, make it calculate flags explicitly
+        if (tmp_op->op == OP_TEST || tmp_op->op == OP_CMP)
+        {
+          if (branched || scan_for_mod(tmp_op, setters[j] + 1, i, 0) >= 0)
+            pfomask = 1 << po->pfo;
+        }
+        else if (tmp_op->op == OP_CMPS || tmp_op->op == OP_SCAS) {
+          pfomask = 1 << po->pfo;
+        }
+        else {
+          // see if we'll be able to handle based on op result
+          if ((tmp_op->op != OP_AND && tmp_op->op != OP_OR
+               && po->pfo != PFO_Z && po->pfo != PFO_S
+               && po->pfo != PFO_P)
+              || branched
+              || scan_for_mod_opr0(tmp_op, setters[j] + 1, i) >= 0)
+          {
+            pfomask = 1 << po->pfo;
+          }
+
+          if (tmp_op->op == OP_ADD && po->pfo == PFO_C) {
+            propagate_lmod(tmp_op, &tmp_op->operand[0],
+              &tmp_op->operand[1]);
+            if (tmp_op->operand[0].lmod == OPLM_DWORD)
+              need_tmp64 = 1;
+          }
+        }
+        if (pfomask) {
+          tmp_op->pfomask |= pfomask;
+          cond_vars |= pfomask;
+        }
+        // note: may overwrite, currently not a problem
+        po->datap = tmp_op;
+      }
+
+      if (po->op == OP_RCL || po->op == OP_RCR
+       || po->op == OP_ADC || po->op == OP_SBB)
+        cond_vars |= 1 << PFO_C;
+    }
+
+    if (po->op == OP_CMPS || po->op == OP_SCAS) {
+      cond_vars |= 1 << PFO_Z;
+    }
+    else if (po->op == OP_MUL
+      || (po->op == OP_IMUL && po->operand_cnt == 1))
+    {
+      if (po->operand[0].lmod == OPLM_DWORD)
+        need_tmp64 = 1;
+    }
+    else if (po->op == OP_CALL) {
+      // note: resolved non-reg calls are OPF_DONE already
+      pp = po->pp;
+      ferr_assert(po, pp != NULL);
+
+      if (pp->is_unresolved) {
+        int regmask_stack = 0;
+        collect_call_args(po, i, pp, &regmask, save_arg_vars,
+          i + opcnt * 2);
+
+        // this is pretty rough guess:
+        // see ecx and edx were pushed (and not their saved versions)
+        for (arg = 0; arg < pp->argc; arg++) {
+          if (pp->arg[arg].reg != NULL)
+            continue;
+
+          tmp_op = pp->arg[arg].datap;
+          if (tmp_op == NULL)
+            ferr(po, "parsed_op missing for arg%d\n", arg);
+          if (tmp_op->p_argnum == 0 && tmp_op->operand[0].type == OPT_REG)
+            regmask_stack |= 1 << tmp_op->operand[0].reg;
+        }
+
+        if (!((regmask_stack & (1 << xCX))
+          && (regmask_stack & (1 << xDX))))
+        {
+          if (pp->argc_stack != 0
+           || ((regmask | regmask_arg) & ((1 << xCX)|(1 << xDX))))
+          {
+            pp_insert_reg_arg(pp, "ecx");
+            pp->is_fastcall = 1;
+            regmask_init |= 1 << xCX;
+            regmask |= 1 << xCX;
+          }
+          if (pp->argc_stack != 0
+           || ((regmask | regmask_arg) & (1 << xDX)))
+          {
+            pp_insert_reg_arg(pp, "edx");
+            regmask_init |= 1 << xDX;
+            regmask |= 1 << xDX;
+          }
+        }
+
+        // note: __cdecl doesn't fall into is_unresolved category
+        if (pp->argc_stack > 0)
+          pp->is_stdcall = 1;
+      }
+    }
+    else if (po->op == OP_MOV && po->operand[0].pp != NULL
+      && po->operand[1].pp != NULL)
+    {
+      // <var> = offset <something>
+      if ((po->operand[1].pp->is_func || po->operand[1].pp->is_fptr)
+        && !IS_START(po->operand[1].name, "off_"))
+      {
+        if (!po->operand[0].pp->is_fptr)
+          ferr(po, "%s not declared as fptr when it should be\n",
+            po->operand[0].name);
+        if (pp_cmp_func(po->operand[0].pp, po->operand[1].pp)) {
+          pp_print(buf1, sizeof(buf1), po->operand[0].pp);
+          pp_print(buf2, sizeof(buf2), po->operand[1].pp);
+          fnote(po, "var:  %s\n", buf1);
+          fnote(po, "func: %s\n", buf2);
+          ferr(po, "^ mismatch\n");
+        }
+      }
+    }
+    else if (po->op == OP_DIV || po->op == OP_IDIV) {
+      if (po->operand[0].lmod == OPLM_DWORD) {
+        // 32bit division is common, look for it
+        if (po->op == OP_DIV)
+          ret = scan_for_reg_clear(i, xDX);
+        else
+          ret = scan_for_cdq_edx(i);
+        if (ret >= 0)
+          po->flags |= OPF_32BIT;
+        else
+          need_tmp64 = 1;
+      }
+      else
+        need_tmp_var = 1;
+    }
+    else if (po->op == OP_CLD)
+      po->flags |= OPF_RMD | OPF_DONE;
+    else if (po->op == OPP_FTOL) {
+      struct parsed_opr opr = OPR_INIT(OPT_REG, OPLM_DWORD, xDX);
+      j = -1;
+      find_next_read(i + 1, opcnt, &opr, i + opcnt * 18, &j);
+      if (j == -1)
+        po->flags |= OPF_32BIT;
+    }
+    else if (po->op == OP_FLD && po->operand[0].lmod == OPLM_QWORD)
+      need_double = 1;
+
+    if (po->op == OP_RCL || po->op == OP_RCR || po->op == OP_XCHG)
+      need_tmp_var = 1;
+  }
+
+  float_type = need_double ? "double" : "float";
+
+  // output starts here
+
+  // define userstack size
+  if (g_func_pp->is_userstack) {
+    fprintf(fout, "#ifndef US_SZ_%s\n", g_func_pp->name);
+    fprintf(fout, "#define US_SZ_%s USERSTACK_SIZE\n", g_func_pp->name);
+    fprintf(fout, "#endif\n");
+  }
+
+  // the function itself
+  ferr_assert(ops, !g_func_pp->is_fptr);
+  output_pp(fout, g_func_pp,
+    (g_ida_func_attr & IDAFA_NORETURN) ? OPP_FORCE_NORETURN : 0);
+  fprintf(fout, "\n{\n");
+
+  // declare indirect functions
+  for (i = 0; i < opcnt; i++) {
+    po = &ops[i];
+    if (po->flags & OPF_RMD)
+      continue;
+
+    if (po->op == OP_CALL) {
+      pp = po->pp;
+      if (pp == NULL)
+        ferr(po, "NULL pp\n");
+
+      if (pp->is_fptr && !(pp->name[0] != 0 && pp->is_arg)) {
+        if (pp->name[0] != 0) {
+          memmove(pp->name + 2, pp->name, strlen(pp->name) + 1);
+          memcpy(pp->name, "i_", 2);
+
+          // might be declared already
+          found = 0;
+          for (j = 0; j < i; j++) {
+            if (ops[j].op == OP_CALL && (pp_tmp = ops[j].pp)) {
+              if (pp_tmp->is_fptr && IS(pp->name, pp_tmp->name)) {
+                found = 1;
+                break;
+              }
+            }
+          }
+          if (found)
+            continue;
+        }
+        else
+          snprintf(pp->name, sizeof(pp->name), "icall%d", i);
+
+        fprintf(fout, "  ");
+        output_pp(fout, pp, OPP_SIMPLE_ARGS);
+        fprintf(fout, ";\n");
+      }
+    }
+  }
+
+  // output LUTs/jumptables
+  for (i = 0; i < g_func_pd_cnt; i++) {
+    pd = &g_func_pd[i];
+    fprintf(fout, "  static const ");
+    if (pd->type == OPT_OFFSET) {
+      fprintf(fout, "void *jt_%s[] =\n    { ", pd->label);
+
+      for (j = 0; j < pd->count; j++) {
+        if (j > 0)
+          fprintf(fout, ", ");
+        fprintf(fout, "&&%s", pd->d[j].u.label);
+      }
+    }
+    else {
+      fprintf(fout, "%s %s[] =\n    { ",
+        lmod_type_u(ops, pd->lmod), pd->label);
+
+      for (j = 0; j < pd->count; j++) {
+        if (j > 0)
+          fprintf(fout, ", ");
+        fprintf(fout, "%u", pd->d[j].u.val);
+      }
+    }
+    fprintf(fout, " };\n");
+    had_decl = 1;
+  }
+
+  // declare stack frame, va_arg
+  if (g_stack_fsz) {
+    fprintf(fout, "  union { u32 d[%d];", (g_stack_fsz + 3) / 4);
+    if (g_func_lmods & (1 << OPLM_WORD))
+      fprintf(fout, " u16 w[%d];", (g_stack_fsz + 1) / 2);
+    if (g_func_lmods & (1 << OPLM_BYTE))
+      fprintf(fout, " u8 b[%d];", g_stack_fsz);
+    if (g_func_lmods & (1 << OPLM_QWORD))
+      fprintf(fout, " double q[%d];", (g_stack_fsz + 7) / 8);
+    fprintf(fout, " } sf;\n");
+    had_decl = 1;
+  }
+
+  if (g_func_pp->is_userstack) {
+    fprintf(fout, "  u32 fake_sf[US_SZ_%s / 4];\n", g_func_pp->name);
+    fprintf(fout, "  u32 *esp = &fake_sf[sizeof(fake_sf) / 4];\n");
+    had_decl = 1;
+  }
+
+  if (g_func_pp->is_vararg) {
+    fprintf(fout, "  va_list ap;\n");
+    had_decl = 1;
+  }
+
+  // declare arg-registers
+  for (i = 0; i < g_func_pp->argc; i++) {
+    if (g_func_pp->arg[i].reg != NULL) {
+      reg = char_array_i(regs_r32,
+              ARRAY_SIZE(regs_r32), g_func_pp->arg[i].reg);
+      if (regmask & (1 << reg)) {
+        if (g_func_pp->arg[i].type.is_retreg)
+          fprintf(fout, "  u32 %s = *r_%s;\n",
+            g_func_pp->arg[i].reg, g_func_pp->arg[i].reg);
+        else
+          fprintf(fout, "  u32 %s = (u32)a%d;\n",
+            g_func_pp->arg[i].reg, i + 1);
+      }
+      else {
+        if (g_func_pp->arg[i].type.is_retreg)
+          ferr(ops, "retreg '%s' is unused?\n",
+            g_func_pp->arg[i].reg);
+        fprintf(fout, "  // %s = a%d; // unused\n",
+          g_func_pp->arg[i].reg, i + 1);
+      }
+      had_decl = 1;
+    }
+  }
+
+  // declare normal registers
+  regmask_now = regmask & ~regmask_arg;
+  regmask_now &= ~(1 << xSP);
+  if (regmask_now & 0x00ff) {
+    for (reg = 0; reg < 8; reg++) {
+      if (regmask_now & (1 << reg)) {
+        fprintf(fout, "  u32 %s", regs_r32[reg]);
+        if (regmask_init & (1 << reg))
+          fprintf(fout, " = 0");
+        fprintf(fout, ";\n");
+        had_decl = 1;
+      }
+    }
+  }
+  // ... mmx
+  if (regmask_now & 0xff00) {
+    for (reg = 8; reg < 16; reg++) {
+      if (regmask_now & (1 << reg)) {
+        fprintf(fout, "  mmxr %s", regs_r32[reg]);
+        if (regmask_init & (1 << reg))
+          fprintf(fout, " = { 0, }");
+        fprintf(fout, ";\n");
+        had_decl = 1;
+      }
+    }
+  }
+  // ... x87
+  if (regmask_now & 0xff0000) {
+    for (reg = 16; reg < 24; reg++) {
+      if (regmask_now & (1 << reg)) {
+        fprintf(fout, "  %s f_st%d", float_type, reg - 16);
+        if (regmask_init & (1 << reg))
+          fprintf(fout, " = 0");
+        fprintf(fout, ";\n");
+        had_decl = 1;
+      }
+    }
+  }
+
+  if (regmask_save) {
+    for (reg = 0; reg < 8; reg++) {
+      if (regmask_save & (1 << reg)) {
+        fprintf(fout, "  u32 s_%s;\n", regs_r32[reg]);
+        had_decl = 1;
+      }
+    }
+  }
+
+  for (i = 0; i < ARRAY_SIZE(save_arg_vars); i++) {
+    if (save_arg_vars[i] == 0)
+      continue;
+    for (reg = 0; reg < 32; reg++) {
+      if (save_arg_vars[i] & (1 << reg)) {
+        fprintf(fout, "  u32 %s;\n",
+          saved_arg_name(buf1, sizeof(buf1), i, reg + 1));
+        had_decl = 1;
+      }
+    }
+  }
+
+  // declare push-pop temporaries
+  if (regmask_pp) {
+    for (reg = 0; reg < 8; reg++) {
+      if (regmask_pp & (1 << reg)) {
+        fprintf(fout, "  u32 pp_%s;\n", regs_r32[reg]);
+        had_decl = 1;
+      }
+    }
+  }
+
+  if (cond_vars) {
+    for (i = 0; i < 8; i++) {
+      if (cond_vars & (1 << i)) {
+        fprintf(fout, "  u32 cond_%s;\n", parsed_flag_op_names[i]);
+        had_decl = 1;
+      }
+    }
+  }
+
+  if (need_tmp_var) {
+    fprintf(fout, "  u32 tmp;\n");
+    had_decl = 1;
+  }
+
+  if (need_tmp64) {
+    fprintf(fout, "  u64 tmp64;\n");
+    had_decl = 1;
+  }
+
+  if (had_decl)
+    fprintf(fout, "\n");
+
+  // do stack clear, if needed
+  if (g_sct_func_attr & SCTFA_CLEAR_SF) {
+    fprintf(fout, "  ");
+    if (g_stack_clear_len != 0) {
+      if (g_stack_clear_len <= 4) {
+        for (i = 0; i < g_stack_clear_len; i++)
+          fprintf(fout, "sf.d[%d] = ", g_stack_clear_start + i);
+        fprintf(fout, "0;\n");
+      }
+      else {
+        fprintf(fout, "memset(&sf[%d], 0, %d);\n",
+          g_stack_clear_start, g_stack_clear_len * 4);
+      }
+    }
+    else
+      fprintf(fout, "memset(&sf, 0, sizeof(sf));\n");
+  }
+
+  if (g_func_pp->is_vararg) {
+    if (g_func_pp->argc_stack == 0)
+      ferr(ops, "vararg func without stack args?\n");
+    fprintf(fout, "  va_start(ap, a%d);\n", g_func_pp->argc);
+  }
+
+  // output ops
+  for (i = 0; i < opcnt; i++)
+  {
+    if (g_labels[i] != NULL) {
+      fprintf(fout, "\n%s:\n", g_labels[i]);
+      label_pending = 1;
+
+      delayed_flag_op = NULL;
+      last_arith_dst = NULL;
+    }
+
+    po = &ops[i];
+    if (po->flags & OPF_RMD)
+      continue;
+
+    no_output = 0;
+
+    #define assert_operand_cnt(n_) \
+      if (po->operand_cnt != n_) \
+        ferr(po, "operand_cnt is %d/%d\n", po->operand_cnt, n_)
+
+    // conditional/flag using op?
+    if (po->flags & OPF_CC)
+    {
+      int is_delayed = 0;
+
+      tmp_op = po->datap;
+
+      // we go through all this trouble to avoid using parsed_flag_op,
+      // which makes generated code much nicer
+      if (delayed_flag_op != NULL)
+      {
+        out_cmp_test(buf1, sizeof(buf1), delayed_flag_op,
+          po->pfo, po->pfo_inv);
+        is_delayed = 1;
+      }
+      else if (last_arith_dst != NULL
+        && (po->pfo == PFO_Z || po->pfo == PFO_S || po->pfo == PFO_P
+           || (tmp_op && (tmp_op->op == OP_AND || tmp_op->op == OP_OR))
+           ))
+      {
+        out_src_opr_u32(buf3, sizeof(buf3), po, last_arith_dst);
+        out_test_for_cc(buf1, sizeof(buf1), po, po->pfo, po->pfo_inv,
+          last_arith_dst->lmod, buf3);
+        is_delayed = 1;
+      }
+      else if (tmp_op != NULL) {
+        // use preprocessed flag calc results
+        if (!(tmp_op->pfomask & (1 << po->pfo)))
+          ferr(po, "not prepared for pfo %d\n", po->pfo);
+
+        // note: pfo_inv was not yet applied
+        snprintf(buf1, sizeof(buf1), "(%scond_%s)",
+          po->pfo_inv ? "!" : "", parsed_flag_op_names[po->pfo]);
+      }
+      else {
+        ferr(po, "all methods of finding comparison failed\n");
+      }
+ 
+      if (po->flags & OPF_JMP) {
+        fprintf(fout, "  if %s", buf1);
+      }
+      else if (po->op == OP_RCL || po->op == OP_RCR
+               || po->op == OP_ADC || po->op == OP_SBB)
+      {
+        if (is_delayed)
+          fprintf(fout, "  cond_%s = %s;\n",
+            parsed_flag_op_names[po->pfo], buf1);
+      }
+      else if (po->flags & OPF_DATA) { // SETcc
+        out_dst_opr(buf2, sizeof(buf2), po, &po->operand[0]);
+        fprintf(fout, "  %s = %s;", buf2, buf1);
+      }
+      else {
+        ferr(po, "unhandled conditional op\n");
+      }
+    }
+
+    pfomask = po->pfomask;
+
+    if (po->flags & (OPF_REPZ|OPF_REPNZ)) {
+      struct parsed_opr opr = OPR_INIT(OPT_REG, OPLM_DWORD, xCX);
+      ret = try_resolve_const(i, &opr, opcnt * 7 + i, &uval);
+
+      if (ret != 1 || uval == 0) {
+        // we need initial flags for ecx=0 case..
+        if (i > 0 && ops[i - 1].op == OP_XOR
+          && IS(ops[i - 1].operand[0].name,
+                ops[i - 1].operand[1].name))
+        {
+          fprintf(fout, "  cond_z = ");
+          if (pfomask & (1 << PFO_C))
+            fprintf(fout, "cond_c = ");
+          fprintf(fout, "0;\n");
+        }
+        else if (last_arith_dst != NULL) {
+          out_src_opr_u32(buf3, sizeof(buf3), po, last_arith_dst);
+          out_test_for_cc(buf1, sizeof(buf1), po, PFO_Z, 0,
+            last_arith_dst->lmod, buf3);
+          fprintf(fout, "  cond_z = %s;\n", buf1);
+        }
+        else
+          ferr(po, "missing initial ZF\n");
+      }
+    }
+
+    switch (po->op)
+    {
+      case OP_MOV:
+        assert_operand_cnt(2);
+        propagate_lmod(po, &po->operand[0], &po->operand[1]);
+        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
+        default_cast_to(buf3, sizeof(buf3), &po->operand[0]);
+        fprintf(fout, "  %s = %s;", buf1,
+            out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
+              buf3, 0));
+        break;
+
+      case OP_LEA:
+        assert_operand_cnt(2);
+        po->operand[1].lmod = OPLM_DWORD; // always
+        fprintf(fout, "  %s = %s;",
+            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
+            out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
+              NULL, 1));
+        break;
+
+      case OP_MOVZX:
+        assert_operand_cnt(2);
+        fprintf(fout, "  %s = %s;",
+            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
+            out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
+        break;
+
+      case OP_MOVSX:
+        assert_operand_cnt(2);
+        switch (po->operand[1].lmod) {
+        case OPLM_BYTE:
+          strcpy(buf3, "(s8)");
+          break;
+        case OPLM_WORD:
+          strcpy(buf3, "(s16)");
+          break;
+        default:
+          ferr(po, "invalid src lmod: %d\n", po->operand[1].lmod);
+        }
+        fprintf(fout, "  %s = %s;",
+            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
+            out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
+              buf3, 0));
+        break;
+
+      case OP_XCHG:
+        assert_operand_cnt(2);
+        propagate_lmod(po, &po->operand[0], &po->operand[1]);
+        fprintf(fout, "  tmp = %s;",
+          out_src_opr(buf1, sizeof(buf1), po, &po->operand[0], "", 0));
+        fprintf(fout, " %s = %s;",
+          out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
+          out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
+            default_cast_to(buf3, sizeof(buf3), &po->operand[0]), 0));
+        fprintf(fout, " %s = %stmp;",
+          out_dst_opr(buf1, sizeof(buf1), po, &po->operand[1]),
+          default_cast_to(buf3, sizeof(buf3), &po->operand[1]));
+        snprintf(g_comment, sizeof(g_comment), "xchg");
+        break;
+
+      case OP_NOT:
+        assert_operand_cnt(1);
+        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
+        fprintf(fout, "  %s = ~%s;", buf1, buf1);
+        break;
+
+      case OP_XLAT:
+        assert_operand_cnt(2);
+        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
+        out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]);
+        fprintf(fout, "  %s = *(u8 *)(%s + %s);", buf1, buf2, buf1);
+        strcpy(g_comment, "xlat");
+        break;
+
+      case OP_CDQ:
+        assert_operand_cnt(2);
+        fprintf(fout, "  %s = (s32)%s >> 31;",
+            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
+            out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
+        strcpy(g_comment, "cdq");
+        break;
+
+      case OP_LODS:
+        if (po->flags & OPF_REP) {
+          assert_operand_cnt(3);
+          // hmh..
+          ferr(po, "TODO\n");
+        }
+        else {
+          assert_operand_cnt(2);
+          fprintf(fout, "  %s = %sesi; esi %c= %d;",
+            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[1]),
+            lmod_cast_u_ptr(po, po->operand[1].lmod),
+            (po->flags & OPF_DF) ? '-' : '+',
+            lmod_bytes(po, po->operand[1].lmod));
+          strcpy(g_comment, "lods");
+        }
+        break;
+
+      case OP_STOS:
+        if (po->flags & OPF_REP) {
+          assert_operand_cnt(3);
+          fprintf(fout, "  for (; ecx != 0; ecx--, edi %c= %d)\n",
+            (po->flags & OPF_DF) ? '-' : '+',
+            lmod_bytes(po, po->operand[1].lmod));
+          fprintf(fout, "    %sedi = eax;",
+            lmod_cast_u_ptr(po, po->operand[1].lmod));
+          strcpy(g_comment, "rep stos");
+        }
+        else {
+          assert_operand_cnt(2);
+          fprintf(fout, "  %sedi = eax; edi %c= %d;",
+            lmod_cast_u_ptr(po, po->operand[1].lmod),
+            (po->flags & OPF_DF) ? '-' : '+',
+            lmod_bytes(po, po->operand[1].lmod));
+          strcpy(g_comment, "stos");
+        }
+        break;
+
+      case OP_MOVS:
+        j = lmod_bytes(po, po->operand[0].lmod);
+        strcpy(buf1, lmod_cast_u_ptr(po, po->operand[0].lmod));
+        l = (po->flags & OPF_DF) ? '-' : '+';
+        if (po->flags & OPF_REP) {
+          assert_operand_cnt(3);
+          fprintf(fout,
+            "  for (; ecx != 0; ecx--, edi %c= %d, esi %c= %d)\n",
+            l, j, l, j);
+          fprintf(fout,
+            "    %sedi = %sesi;", buf1, buf1);
+          strcpy(g_comment, "rep movs");
+        }
+        else {
+          assert_operand_cnt(2);
+          fprintf(fout, "  %sedi = %sesi; edi %c= %d; esi %c= %d;",
+            buf1, buf1, l, j, l, j);
+          strcpy(g_comment, "movs");
+        }
+        break;
+
+      case OP_CMPS:
+        // repe ~ repeat while ZF=1
+        j = lmod_bytes(po, po->operand[0].lmod);
+        strcpy(buf1, lmod_cast_u_ptr(po, po->operand[0].lmod));
+        l = (po->flags & OPF_DF) ? '-' : '+';
+        if (po->flags & OPF_REP) {
+          assert_operand_cnt(3);
+          fprintf(fout,
+            "  for (; ecx != 0; ecx--) {\n");
+          if (pfomask & (1 << PFO_C)) {
+            // ugh..
+            fprintf(fout,
+            "    cond_c = %sesi < %sedi;\n", buf1, buf1);
+            pfomask &= ~(1 << PFO_C);
+          }
+          fprintf(fout,
+            "    cond_z = (%sesi == %sedi); esi %c= %d, edi %c= %d;\n",
+              buf1, buf1, l, j, l, j);
+          fprintf(fout,
+            "    if (cond_z %s 0) break;\n",
+              (po->flags & OPF_REPZ) ? "==" : "!=");
+          fprintf(fout,
+            "  }");
+          snprintf(g_comment, sizeof(g_comment), "rep%s cmps",
+            (po->flags & OPF_REPZ) ? "e" : "ne");
+        }
+        else {
+          assert_operand_cnt(2);
+          fprintf(fout,
+            "  cond_z = (%sesi == %sedi); esi %c= %d; edi %c= %d;",
+            buf1, buf1, l, j, l, j);
+          strcpy(g_comment, "cmps");
+        }
+        pfomask &= ~(1 << PFO_Z);
+        last_arith_dst = NULL;
+        delayed_flag_op = NULL;
+        break;
+
+      case OP_SCAS:
+        // only does ZF (for now)
+        // repe ~ repeat while ZF=1
+        j = lmod_bytes(po, po->operand[1].lmod);
+        l = (po->flags & OPF_DF) ? '-' : '+';
+        if (po->flags & OPF_REP) {
+          assert_operand_cnt(3);
+          fprintf(fout,
+            "  for (; ecx != 0; ecx--) {\n");
+          fprintf(fout,
+            "    cond_z = (%seax == %sedi); edi %c= %d;\n",
+              lmod_cast_u(po, po->operand[1].lmod),
+              lmod_cast_u_ptr(po, po->operand[1].lmod), l, j);
+          fprintf(fout,
+            "    if (cond_z %s 0) break;\n",
+              (po->flags & OPF_REPZ) ? "==" : "!=");
+          fprintf(fout,
+            "  }");
+          snprintf(g_comment, sizeof(g_comment), "rep%s scas",
+            (po->flags & OPF_REPZ) ? "e" : "ne");
+        }
+        else {
+          assert_operand_cnt(2);
+          fprintf(fout, "  cond_z = (%seax == %sedi); edi %c= %d;",
+              lmod_cast_u(po, po->operand[1].lmod),
+              lmod_cast_u_ptr(po, po->operand[1].lmod), l, j);
+          strcpy(g_comment, "scas");
+        }
+        pfomask &= ~(1 << PFO_Z);
+        last_arith_dst = NULL;
+        delayed_flag_op = NULL;
+        break;
+
+      // arithmetic w/flags
+      case OP_AND:
+        if (po->operand[1].type == OPT_CONST && !po->operand[1].val)
+          goto dualop_arith_const;
+        propagate_lmod(po, &po->operand[0], &po->operand[1]);
+        goto dualop_arith;
+
+      case OP_OR:
+        propagate_lmod(po, &po->operand[0], &po->operand[1]);
+        if (po->operand[1].type == OPT_CONST) {
+          j = lmod_bytes(po, po->operand[0].lmod);
+          if (((1ull << j * 8) - 1) == po->operand[1].val)
+            goto dualop_arith_const;
+        }
+        goto dualop_arith;
+
+      dualop_arith:
+        assert_operand_cnt(2);
+        fprintf(fout, "  %s %s= %s;",
+            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
+            op_to_c(po),
+            out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
+        output_std_flags(fout, po, &pfomask, buf1);
+        last_arith_dst = &po->operand[0];
+        delayed_flag_op = NULL;
+        break;
+
+      dualop_arith_const:
+        // and 0, or ~0 used instead mov
+        assert_operand_cnt(2);
+        fprintf(fout, "  %s = %s;",
+          out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
+          out_src_opr(buf2, sizeof(buf2), po, &po->operand[1],
+           default_cast_to(buf3, sizeof(buf3), &po->operand[0]), 0));
+        output_std_flags(fout, po, &pfomask, buf1);
+        last_arith_dst = &po->operand[0];
+        delayed_flag_op = NULL;
+        break;
+
+      case OP_SHL:
+      case OP_SHR:
+        assert_operand_cnt(2);
+        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
+        if (pfomask & (1 << PFO_C)) {
+          if (po->operand[1].type == OPT_CONST) {
+            l = lmod_bytes(po, po->operand[0].lmod) * 8;
+            j = po->operand[1].val;
+            j %= l;
+            if (j != 0) {
+              if (po->op == OP_SHL)
+                j = l - j;
+              else
+                j -= 1;
+              fprintf(fout, "  cond_c = (%s >> %d) & 1;\n",
+                buf1, j);
+            }
+            else
+              ferr(po, "zero shift?\n");
+          }
+          else
+            ferr(po, "TODO\n");
+          pfomask &= ~(1 << PFO_C);
+        }
+        fprintf(fout, "  %s %s= %s", buf1, op_to_c(po),
+            out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
+        if (po->operand[1].type != OPT_CONST)
+          fprintf(fout, " & 0x1f");
+        fprintf(fout, ";");
+        output_std_flags(fout, po, &pfomask, buf1);
+        last_arith_dst = &po->operand[0];
+        delayed_flag_op = NULL;
+        break;
+
+      case OP_SAR:
+        assert_operand_cnt(2);
+        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
+        fprintf(fout, "  %s = %s%s >> %s;", buf1,
+          lmod_cast_s(po, po->operand[0].lmod), buf1,
+          out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
+        output_std_flags(fout, po, &pfomask, buf1);
+        last_arith_dst = &po->operand[0];
+        delayed_flag_op = NULL;
+        break;
+
+      case OP_SHLD:
+      case OP_SHRD:
+        assert_operand_cnt(3);
+        propagate_lmod(po, &po->operand[0], &po->operand[1]);
+        l = lmod_bytes(po, po->operand[0].lmod) * 8;
+        out_src_opr_u32(buf3, sizeof(buf3), po, &po->operand[2]);
+        if (po->operand[2].type != OPT_CONST) {
+          // no handling for "undefined" case, hopefully not needed
+          snprintf(buf2, sizeof(buf2), "(%s & 0x1f)", buf3);
+          strcpy(buf3, buf2);
+        }
+        out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]);
+        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
+        if (po->op == OP_SHLD) {
+          fprintf(fout, "  %s <<= %s; %s |= %s >> (%d - %s);",
+            buf1, buf3, buf1, buf2, l, buf3);
+          strcpy(g_comment, "shld");
+        }
+        else {
+          fprintf(fout, "  %s >>= %s; %s |= %s << (%d - %s);",
+            buf1, buf3, buf1, buf2, l, buf3);
+          strcpy(g_comment, "shrd");
+        }
+        output_std_flags(fout, po, &pfomask, buf1);
+        last_arith_dst = &po->operand[0];
+        delayed_flag_op = NULL;
+        break;
+
+      case OP_ROL:
+      case OP_ROR:
+        assert_operand_cnt(2);
+        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
+        if (po->operand[1].type == OPT_CONST) {
+          j = po->operand[1].val;
+          j %= lmod_bytes(po, po->operand[0].lmod) * 8;
+          fprintf(fout, po->op == OP_ROL ?
+            "  %s = (%s << %d) | (%s >> %d);" :
+            "  %s = (%s >> %d) | (%s << %d);",
+            buf1, buf1, j, buf1,
+            lmod_bytes(po, po->operand[0].lmod) * 8 - j);
+        }
+        else
+          ferr(po, "TODO\n");
+        output_std_flags(fout, po, &pfomask, buf1);
+        last_arith_dst = &po->operand[0];
+        delayed_flag_op = NULL;
+        break;
+
+      case OP_RCL:
+      case OP_RCR:
+        assert_operand_cnt(2);
+        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
+        l = lmod_bytes(po, po->operand[0].lmod) * 8;
+        if (po->operand[1].type == OPT_CONST) {
+          j = po->operand[1].val % l;
+          if (j == 0)
+            ferr(po, "zero rotate\n");
+          fprintf(fout, "  tmp = (%s >> %d) & 1;\n",
+            buf1, (po->op == OP_RCL) ? (l - j) : (j - 1));
+          if (po->op == OP_RCL) {
+            fprintf(fout,
+              "  %s = (%s << %d) | (cond_c << %d)",
+              buf1, buf1, j, j - 1);
+            if (j != 1)
+              fprintf(fout, " | (%s >> %d)", buf1, l + 1 - j);
+          }
+          else {
+            fprintf(fout,
+              "  %s = (%s >> %d) | (cond_c << %d)",
+              buf1, buf1, j, l - j);
+            if (j != 1)
+              fprintf(fout, " | (%s << %d)", buf1, l + 1 - j);
+          }
+          fprintf(fout, ";\n");
+          fprintf(fout, "  cond_c = tmp;");
+        }
+        else
+          ferr(po, "TODO\n");
+        strcpy(g_comment, (po->op == OP_RCL) ? "rcl" : "rcr");
+        output_std_flags(fout, po, &pfomask, buf1);
+        last_arith_dst = &po->operand[0];
+        delayed_flag_op = NULL;
+        break;
+
+      case OP_XOR:
+        assert_operand_cnt(2);
+        propagate_lmod(po, &po->operand[0], &po->operand[1]);
+        if (IS(opr_name(po, 0), opr_name(po, 1))) {
+          // special case for XOR
+          if (pfomask & (1 << PFO_BE)) { // weird, but it happens..
+            fprintf(fout, "  cond_be = 1;\n");
+            pfomask &= ~(1 << PFO_BE);
+          }
+          fprintf(fout, "  %s = 0;",
+            out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]));
+          last_arith_dst = &po->operand[0];
+          delayed_flag_op = NULL;
+          break;
+        }
+        goto dualop_arith;
+
+      case OP_ADD:
+        assert_operand_cnt(2);
+        propagate_lmod(po, &po->operand[0], &po->operand[1]);
+        if (pfomask & (1 << PFO_C)) {
+          out_src_opr_u32(buf1, sizeof(buf1), po, &po->operand[0]);
+          out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]);
+          if (po->operand[0].lmod == OPLM_DWORD) {
+            fprintf(fout, "  tmp64 = (u64)%s + %s;\n", buf1, buf2);
+            fprintf(fout, "  cond_c = tmp64 >> 32;\n");
+            fprintf(fout, "  %s = (u32)tmp64;",
+              out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]));
+            strcat(g_comment, " add64");
+          }
+          else {
+            fprintf(fout, "  cond_c = ((u32)%s + %s) >> %d;\n",
+              buf1, buf2, lmod_bytes(po, po->operand[0].lmod) * 8);
+            fprintf(fout, "  %s += %s;",
+              out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
+              buf2);
+          }
+          pfomask &= ~(1 << PFO_C);
+          output_std_flags(fout, po, &pfomask, buf1);
+          last_arith_dst = &po->operand[0];
+          delayed_flag_op = NULL;
+          break;
+        }
+        goto dualop_arith;
+
+      case OP_SUB:
+        assert_operand_cnt(2);
+        propagate_lmod(po, &po->operand[0], &po->operand[1]);
+        if (pfomask & ~((1 << PFO_Z) | (1 << PFO_S))) {
+          for (j = 0; j <= PFO_LE; j++) {
+            if (!(pfomask & (1 << j)))
+              continue;
+            if (j == PFO_Z || j == PFO_S)
+              continue;
+
+            out_cmp_for_cc(buf1, sizeof(buf1), po, j, 0);
+            fprintf(fout, "  cond_%s = %s;\n",
+              parsed_flag_op_names[j], buf1);
+            pfomask &= ~(1 << j);
+          }
+        }
+        goto dualop_arith;
+
+      case OP_ADC:
+      case OP_SBB:
+        assert_operand_cnt(2);
+        propagate_lmod(po, &po->operand[0], &po->operand[1]);
+        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
+        if (po->op == OP_SBB
+          && IS(po->operand[0].name, po->operand[1].name))
+        {
+          // avoid use of unitialized var
+          fprintf(fout, "  %s = -cond_c;", buf1);
+          // carry remains what it was
+          pfomask &= ~(1 << PFO_C);
+        }
+        else {
+          fprintf(fout, "  %s %s= %s + cond_c;", buf1, op_to_c(po),
+            out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]));
+        }
+        output_std_flags(fout, po, &pfomask, buf1);
+        last_arith_dst = &po->operand[0];
+        delayed_flag_op = NULL;
+        break;
+
+      case OP_BSF:
+        assert_operand_cnt(2);
+        out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[1]);
+        fprintf(fout, "  %s = %s ? __builtin_ffs(%s) - 1 : 0;",
+          out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]),
+          buf2, buf2);
+        output_std_flags(fout, po, &pfomask, buf1);
+        last_arith_dst = &po->operand[0];
+        delayed_flag_op = NULL;
+        strcat(g_comment, " bsf");
+        break;
 
-      case OP_INC:
       case OP_DEC:
+        if (pfomask & ~(PFOB_S | PFOB_S | PFOB_C)) {
+          for (j = 0; j <= PFO_LE; j++) {
+            if (!(pfomask & (1 << j)))
+              continue;
+            if (j == PFO_Z || j == PFO_S || j == PFO_C)
+              continue;
+
+            out_cmp_for_cc(buf1, sizeof(buf1), po, j, 0);
+            fprintf(fout, "  cond_%s = %s;\n",
+              parsed_flag_op_names[j], buf1);
+            pfomask &= ~(1 << j);
+          }
+        }
+        // fallthrough
+
+      case OP_INC:
+        if (pfomask & (1 << PFO_C))
+          // carry is unaffected by inc/dec.. wtf?
+          ferr(po, "carry propagation needed\n");
+
         out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
         if (po->operand[0].type == OPT_REG) {
           strcpy(buf2, po->op == OP_INC ? "++" : "--");
@@ -3759,333 +6143,1556 @@ tailcall:
           strcpy(buf2, po->op == OP_INC ? "+" : "-");
           fprintf(fout, "  %s %s= 1;", buf1, buf2);
         }
-        output_std_flags(fout, po, &pfomask, buf1);
-        last_arith_dst = &po->operand[0];
+        output_std_flags(fout, po, &pfomask, buf1);
+        last_arith_dst = &po->operand[0];
+        delayed_flag_op = NULL;
+        break;
+
+      case OP_NEG:
+        out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
+        out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[0]);
+        fprintf(fout, "  %s = -%s%s;", buf1,
+          lmod_cast_s(po, po->operand[0].lmod), buf2);
+        last_arith_dst = &po->operand[0];
+        delayed_flag_op = NULL;
+        if (pfomask & (1 << PFO_C)) {
+          fprintf(fout, "\n  cond_c = (%s != 0);", buf1);
+          pfomask &= ~(1 << PFO_C);
+        }
+        break;
+
+      case OP_IMUL:
+        if (po->operand_cnt == 2) {
+          propagate_lmod(po, &po->operand[0], &po->operand[1]);
+          goto dualop_arith;
+        }
+        if (po->operand_cnt == 3)
+          ferr(po, "TODO imul3\n");
+        // fallthrough
+      case OP_MUL:
+        assert_operand_cnt(1);
+        switch (po->operand[0].lmod) {
+        case OPLM_DWORD:
+          strcpy(buf1, po->op == OP_IMUL ? "(s64)(s32)" : "(u64)");
+          fprintf(fout, "  tmp64 = %seax * %s%s;\n", buf1, buf1,
+            out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[0]));
+          fprintf(fout, "  edx = tmp64 >> 32;\n");
+          fprintf(fout, "  eax = tmp64;");
+          break;
+        case OPLM_BYTE:
+          strcpy(buf1, po->op == OP_IMUL ? "(s16)(s8)" : "(u16)(u8)");
+          fprintf(fout, "  LOWORD(eax) = %seax * %s;", buf1,
+            out_src_opr(buf2, sizeof(buf2), po, &po->operand[0],
+              buf1, 0));
+          break;
+        default:
+          ferr(po, "TODO: unhandled mul type\n");
+          break;
+        }
+        last_arith_dst = NULL;
+        delayed_flag_op = NULL;
+        break;
+
+      case OP_DIV:
+      case OP_IDIV:
+        assert_operand_cnt(1);
+        out_src_opr_u32(buf1, sizeof(buf1), po, &po->operand[0]);
+        strcpy(cast, lmod_cast(po, po->operand[0].lmod,
+          po->op == OP_IDIV));
+        switch (po->operand[0].lmod) {
+        case OPLM_DWORD:
+          if (po->flags & OPF_32BIT)
+            snprintf(buf2, sizeof(buf2), "%seax", cast);
+          else {
+            fprintf(fout, "  tmp64 = ((u64)edx << 32) | eax;\n");
+            snprintf(buf2, sizeof(buf2), "%stmp64",
+              (po->op == OP_IDIV) ? "(s64)" : "");
+          }
+          if (po->operand[0].type == OPT_REG
+            && po->operand[0].reg == xDX)
+          {
+            fprintf(fout, "  eax = %s / %s%s;\n", buf2, cast, buf1);
+            fprintf(fout, "  edx = %s %% %s%s;", buf2, cast, buf1);
+          }
+          else {
+            fprintf(fout, "  edx = %s %% %s%s;\n", buf2, cast, buf1);
+            fprintf(fout, "  eax = %s / %s%s;", buf2, cast, buf1);
+          }
+          break;
+        case OPLM_WORD:
+          fprintf(fout, "  tmp = (edx << 16) | (eax & 0xffff);\n");
+          snprintf(buf2, sizeof(buf2), "%stmp",
+            (po->op == OP_IDIV) ? "(s32)" : "");
+          if (po->operand[0].type == OPT_REG
+            && po->operand[0].reg == xDX)
+          {
+            fprintf(fout, "  LOWORD(eax) = %s / %s%s;\n",
+              buf2, cast, buf1);
+            fprintf(fout, "  LOWORD(edx) = %s %% %s%s;",
+              buf2, cast, buf1);
+          }
+          else {
+            fprintf(fout, "  LOWORD(edx) = %s %% %s%s;\n",
+              buf2, cast, buf1);
+            fprintf(fout, "  LOWORD(eax) = %s / %s%s;",
+              buf2, cast, buf1);
+          }
+          strcat(g_comment, " div16");
+          break;
+        default:
+          ferr(po, "unhandled div lmod %d\n", po->operand[0].lmod);
+        }
+        last_arith_dst = NULL;
+        delayed_flag_op = NULL;
+        break;
+
+      case OP_TEST:
+      case OP_CMP:
+        propagate_lmod(po, &po->operand[0], &po->operand[1]);
+        if (pfomask != 0) {
+          for (j = 0; j < 8; j++) {
+            if (pfomask & (1 << j)) {
+              out_cmp_test(buf1, sizeof(buf1), po, j, 0);
+              fprintf(fout, "  cond_%s = %s;",
+                parsed_flag_op_names[j], buf1);
+            }
+          }
+          pfomask = 0;
+        }
+        else
+          no_output = 1;
+        last_arith_dst = NULL;
+        delayed_flag_op = po;
+        break;
+
+      case OP_SCC:
+        // SETcc - should already be handled
+        break;
+
+      // note: we reuse OP_Jcc for SETcc, only flags differ
+      case OP_JCC:
+        fprintf(fout, "\n    goto %s;", po->operand[0].name);
+        break;
+
+      case OP_JECXZ:
+        fprintf(fout, "  if (ecx == 0)\n");
+        fprintf(fout, "    goto %s;", po->operand[0].name);
+        strcat(g_comment, " jecxz");
+        break;
+
+      case OP_LOOP:
+        fprintf(fout, "  if (--ecx != 0)\n");
+        fprintf(fout, "    goto %s;", po->operand[0].name);
+        strcat(g_comment, " loop");
+        break;
+
+      case OP_JMP:
+        assert_operand_cnt(1);
+        last_arith_dst = NULL;
+        delayed_flag_op = NULL;
+
+        if (po->operand[0].type == OPT_REGMEM) {
+          ret = sscanf(po->operand[0].name, "%[^[][%[^*]*4]",
+                  buf1, buf2);
+          if (ret != 2)
+            ferr(po, "parse failure for jmp '%s'\n",
+              po->operand[0].name);
+          fprintf(fout, "  goto *jt_%s[%s];", buf1, buf2);
+          break;
+        }
+        else if (po->operand[0].type != OPT_LABEL)
+          ferr(po, "unhandled jmp type\n");
+
+        fprintf(fout, "  goto %s;", po->operand[0].name);
+        break;
+
+      case OP_CALL:
+        assert_operand_cnt(1);
+        pp = po->pp;
+        my_assert_not(pp, NULL);
+
+        strcpy(buf3, "  ");
+        if (po->flags & OPF_CC) {
+          // we treat conditional branch to another func
+          // (yes such code exists..) as conditional tailcall
+          strcat(buf3, "  ");
+          fprintf(fout, " {\n");
+        }
+
+        if (pp->is_fptr && !pp->is_arg) {
+          fprintf(fout, "%s%s = %s;\n", buf3, pp->name,
+            out_src_opr(buf1, sizeof(buf1), po, &po->operand[0],
+              "(void *)", 0));
+          if (pp->is_unresolved)
+            fprintf(fout, "%sunresolved_call(\"%s:%d\", %s);\n",
+              buf3, asmfn, po->asmln, pp->name);
+        }
+
+        fprintf(fout, "%s", buf3);
+        if (strstr(pp->ret_type.name, "int64")) {
+          if (po->flags & OPF_TAIL)
+            ferr(po, "int64 and tail?\n");
+          fprintf(fout, "tmp64 = ");
+        }
+        else if (!IS(pp->ret_type.name, "void")) {
+          if (po->flags & OPF_TAIL) {
+            if (regmask_ret & mxAX) {
+              fprintf(fout, "return ");
+              if (g_func_pp->ret_type.is_ptr != pp->ret_type.is_ptr)
+                fprintf(fout, "(%s)", g_func_pp->ret_type.name);
+            }
+            else if (regmask_ret & mxST0)
+              ferr(po, "float tailcall\n");
+          }
+          else if (po->regmask_dst & mxAX) {
+            fprintf(fout, "eax = ");
+            if (pp->ret_type.is_ptr)
+              fprintf(fout, "(u32)");
+          }
+          else if (po->regmask_dst & mxST0) {
+            fprintf(fout, "f_st0 = ");
+          }
+        }
+
+        if (pp->name[0] == 0)
+          ferr(po, "missing pp->name\n");
+        fprintf(fout, "%s%s(", pp->name,
+          pp->has_structarg ? "_sa" : "");
+
+        if (po->flags & OPF_ATAIL) {
+          if (pp->argc_stack != g_func_pp->argc_stack
+            || (pp->argc_stack > 0
+                && pp->is_stdcall != g_func_pp->is_stdcall))
+            ferr(po, "incompatible tailcall\n");
+          if (g_func_pp->has_retreg)
+            ferr(po, "TODO: retreg+tailcall\n");
+
+          for (arg = j = 0; arg < pp->argc; arg++) {
+            if (arg > 0)
+              fprintf(fout, ", ");
+
+            cast[0] = 0;
+            if (pp->arg[arg].type.is_ptr)
+              snprintf(cast, sizeof(cast), "(%s)",
+                pp->arg[arg].type.name);
+
+            if (pp->arg[arg].reg != NULL) {
+              fprintf(fout, "%s%s", cast, pp->arg[arg].reg);
+              continue;
+            }
+            // stack arg
+            for (; j < g_func_pp->argc; j++)
+              if (g_func_pp->arg[j].reg == NULL)
+                break;
+            fprintf(fout, "%sa%d", cast, j + 1);
+            j++;
+          }
+        }
+        else {
+          for (arg = 0; arg < pp->argc; arg++) {
+            if (arg > 0)
+              fprintf(fout, ", ");
+
+            cast[0] = 0;
+            if (pp->arg[arg].type.is_ptr)
+              snprintf(cast, sizeof(cast), "(%s)",
+                pp->arg[arg].type.name);
+
+            if (pp->arg[arg].reg != NULL) {
+              if (pp->arg[arg].type.is_retreg)
+                fprintf(fout, "&%s", pp->arg[arg].reg);
+              else
+                fprintf(fout, "%s%s", cast, pp->arg[arg].reg);
+              continue;
+            }
+
+            // stack arg
+            tmp_op = pp->arg[arg].datap;
+            if (tmp_op == NULL)
+              ferr(po, "parsed_op missing for arg%d\n", arg);
+
+            if (tmp_op->flags & OPF_VAPUSH) {
+              fprintf(fout, "ap");
+            }
+            else if (tmp_op->p_argpass != 0) {
+              fprintf(fout, "a%d", tmp_op->p_argpass);
+            }
+            else if (tmp_op->p_argnum != 0) {
+              fprintf(fout, "%s%s", cast,
+                saved_arg_name(buf1, sizeof(buf1),
+                  tmp_op->p_arggrp, tmp_op->p_argnum));
+            }
+            else {
+              fprintf(fout, "%s",
+                out_src_opr(buf1, sizeof(buf1),
+                  tmp_op, &tmp_op->operand[0], cast, 0));
+            }
+          }
+        }
+        fprintf(fout, ");");
+
+        if (strstr(pp->ret_type.name, "int64")) {
+          fprintf(fout, "\n");
+          fprintf(fout, "%sedx = tmp64 >> 32;\n", buf3);
+          fprintf(fout, "%seax = tmp64;", buf3);
+        }
+
+        if (pp->is_unresolved) {
+          snprintf(buf2, sizeof(buf2), " unresolved %dreg",
+            pp->argc_reg);
+          strcat(g_comment, buf2);
+        }
+
+        if (po->flags & OPF_TAIL) {
+          ret = 0;
+          if (i == opcnt - 1 || pp->is_noreturn)
+            ret = 0;
+          else if (IS(pp->ret_type.name, "void"))
+            ret = 1;
+          else if (!(regmask_ret & (1 << xAX)))
+            ret = 1;
+          // else already handled as 'return f()'
+
+          if (ret) {
+            fprintf(fout, "\n%sreturn;", buf3);
+            strcat(g_comment, " ^ tailcall");
+          }
+          else
+            strcat(g_comment, " tailcall");
+
+          if ((regmask_ret & (1 << xAX))
+            && IS(pp->ret_type.name, "void") && !pp->is_noreturn)
+          {
+            ferr(po, "int func -> void func tailcall?\n");
+          }
+        }
+        if (pp->is_noreturn)
+          strcat(g_comment, " noreturn");
+        if ((po->flags & OPF_ATAIL) && pp->argc_stack > 0)
+          strcat(g_comment, " argframe");
+        if (po->flags & OPF_CC)
+          strcat(g_comment, " cond");
+
+        if (po->flags & OPF_CC)
+          fprintf(fout, "\n  }");
+
+        delayed_flag_op = NULL;
+        last_arith_dst = NULL;
+        break;
+
+      case OP_RET:
+        if (g_func_pp->is_vararg)
+          fprintf(fout, "  va_end(ap);\n");
+        if (g_func_pp->has_retreg) {
+          for (arg = 0; arg < g_func_pp->argc; arg++)
+            if (g_func_pp->arg[arg].type.is_retreg)
+              fprintf(fout, "  *r_%s = %s;\n",
+                g_func_pp->arg[arg].reg, g_func_pp->arg[arg].reg);
+        }
+ 
+        if (!(regmask_ret & (1 << xAX))) {
+          if (i != opcnt - 1 || label_pending)
+            fprintf(fout, "  return;");
+        }
+        else if (g_func_pp->ret_type.is_ptr) {
+          fprintf(fout, "  return (%s)eax;",
+            g_func_pp->ret_type.name);
+        }
+        else if (IS(g_func_pp->ret_type.name, "__int64"))
+          fprintf(fout, "  return ((u64)edx << 32) | eax;");
+        else
+          fprintf(fout, "  return eax;");
+
+        last_arith_dst = NULL;
         delayed_flag_op = NULL;
         break;
 
-      case OP_NEG:
+      case OP_PUSH:
+        out_src_opr_u32(buf1, sizeof(buf1), po, &po->operand[0]);
+        if (po->p_argnum != 0) {
+          // special case - saved func arg
+          fprintf(fout, "  %s = %s;",
+            saved_arg_name(buf2, sizeof(buf2),
+              po->p_arggrp, po->p_argnum), buf1);
+          break;
+        }
+        else if (po->flags & OPF_RSAVE) {
+          fprintf(fout, "  s_%s = %s;", buf1, buf1);
+          break;
+        }
+        else if (po->flags & OPF_PPUSH) {
+          tmp_op = po->datap;
+          ferr_assert(po, tmp_op != NULL);
+          out_dst_opr(buf2, sizeof(buf2), po, &tmp_op->operand[0]);
+          fprintf(fout, "  pp_%s = %s;", buf2, buf1);
+          break;
+        }
+        else if (g_func_pp->is_userstack) {
+          fprintf(fout, "  *(--esp) = %s;", buf1);
+          break;
+        }
+        if (!(g_ida_func_attr & IDAFA_NORETURN))
+          ferr(po, "stray push encountered\n");
+        no_output = 1;
+        break;
+
+      case OP_POP:
         out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
-        out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[0]);
-        fprintf(fout, "  %s = -%s%s;", buf1,
-          lmod_cast_s(po, po->operand[0].lmod), buf2);
-        last_arith_dst = &po->operand[0];
-        delayed_flag_op = NULL;
-        if (pfomask & (1 << PFO_C)) {
-          fprintf(fout, "\n  cond_c = (%s != 0);", buf1);
-          pfomask &= ~(1 << PFO_C);
+        if (po->flags & OPF_RSAVE) {
+          fprintf(fout, "  %s = s_%s;", buf1, buf1);
+          break;
+        }
+        else if (po->flags & OPF_PPUSH) {
+          // push/pop graph / non-const
+          ferr_assert(po, po->datap == NULL);
+          fprintf(fout, "  %s = pp_%s;", buf1, buf1);
+          break;
+        }
+        else if (po->datap != NULL) {
+          // push/pop pair
+          tmp_op = po->datap;
+          fprintf(fout, "  %s = %s;", buf1,
+            out_src_opr(buf2, sizeof(buf2),
+              tmp_op, &tmp_op->operand[0],
+              default_cast_to(buf3, sizeof(buf3), &po->operand[0]), 0));
+          break;
+        }
+        else if (g_func_pp->is_userstack) {
+          fprintf(fout, "  %s = *esp++;", buf1);
+          break;
         }
+        else
+          ferr(po, "stray pop encountered\n");
         break;
 
-      case OP_IMUL:
-        if (po->operand_cnt == 2) {
-          propagate_lmod(po, &po->operand[0], &po->operand[1]);
-          goto dualop_arith;
+      case OP_NOP:
+        no_output = 1;
+        break;
+
+      // x87
+      case OP_FLD:
+        if (po->flags & OPF_FSHIFT)
+          fprintf(fout, "  f_st1 = f_st0;\n");
+        if (po->operand[0].type == OPT_REG
+          && po->operand[0].reg == xST0)
+        {
+          strcat(g_comment, " fld st");
+          break;
         }
-        if (po->operand_cnt == 3)
-          ferr(po, "TODO imul3\n");
-        // fallthrough
-      case OP_MUL:
-        assert_operand_cnt(1);
-        strcpy(buf1, po->op == OP_IMUL ? "(s64)(s32)" : "(u64)");
-        fprintf(fout, "  mul_tmp = %seax * %s%s;\n", buf1, buf1,
-          out_src_opr_u32(buf2, sizeof(buf2), po, &po->operand[0]));
-        fprintf(fout, "  edx = mul_tmp >> 32;\n");
-        fprintf(fout, "  eax = mul_tmp;");
-        last_arith_dst = NULL;
-        delayed_flag_op = NULL;
+        fprintf(fout, "  f_st0 = %s;",
+          out_src_opr_float(buf1, sizeof(buf1), po, &po->operand[0]));
+        strcat(g_comment, " fld");
+        break;
+
+      case OP_FILD:
+        if (po->flags & OPF_FSHIFT)
+          fprintf(fout, "  f_st1 = f_st0;\n");
+        fprintf(fout, "  f_st0 = (%s)%s;", float_type,
+          out_src_opr(buf1, sizeof(buf1), po, &po->operand[0],
+            lmod_cast(po, po->operand[0].lmod, 1), 0));
+        strcat(g_comment, " fild");
+        break;
+
+      case OP_FLDc:
+        if (po->flags & OPF_FSHIFT)
+          fprintf(fout, "  f_st1 = f_st0;\n");
+        fprintf(fout, "  f_st0 = ");
+        switch (po->operand[0].val) {
+        case X87_CONST_1: fprintf(fout, "1.0;"); break;
+        case X87_CONST_Z: fprintf(fout, "0.0;"); break;
+        default: ferr(po, "TODO\n"); break;
+        }
+        break;
+
+      case OP_FST:
+        if ((po->flags & OPF_FPOP) && po->operand[0].type == OPT_REG
+          && po->operand[0].reg == xST0)
+        {
+          no_output = 1;
+          break;
+        }
+        fprintf(fout, "  %s = f_st0;",
+          out_dst_opr_float(buf1, sizeof(buf1), po, &po->operand[0]));
+        if (po->flags & OPF_FSHIFT)
+          fprintf(fout, "\n  f_st0 = f_st1;");
+        strcat(g_comment, " fst");
+        break;
+
+      case OP_FADD:
+      case OP_FDIV:
+      case OP_FMUL:
+      case OP_FSUB:
+        switch (po->op) {
+        case OP_FADD: j = '+'; break;
+        case OP_FDIV: j = '/'; break;
+        case OP_FMUL: j = '*'; break;
+        case OP_FSUB: j = '-'; break;
+        default: j = 'x'; break;
+        }
+        if (po->flags & OPF_FSHIFT) {
+          fprintf(fout, "  f_st0 = f_st1 %c f_st0;", j);
+        }
+        else {
+          fprintf(fout, "  %s %c= %s;",
+           out_dst_opr_float(buf1, sizeof(buf1), po, &po->operand[0]),
+           j,
+           out_src_opr_float(buf2, sizeof(buf2), po, &po->operand[1]));
+        }
+        break;
+
+      case OP_FDIVR:
+      case OP_FSUBR:
+        if (po->flags & OPF_FSHIFT)
+          snprintf(buf1, sizeof(buf1), "f_st0");
+        else
+          out_dst_opr_float(buf1, sizeof(buf1), po, &po->operand[0]);
+        fprintf(fout, "  %s = %s %c %s;", buf1,
+          out_src_opr_float(buf2, sizeof(buf2), po, &po->operand[1]),
+          po->op == OP_FDIVR ? '/' : '-',
+          out_src_opr_float(buf3, sizeof(buf3), po, &po->operand[0]));
+        break;
+
+      case OP_FIADD:
+      case OP_FIDIV:
+      case OP_FIMUL:
+      case OP_FISUB:
+        switch (po->op) {
+        case OP_FIADD: j = '+'; break;
+        case OP_FIDIV: j = '/'; break;
+        case OP_FIMUL: j = '*'; break;
+        case OP_FISUB: j = '-'; break;
+        default: j = 'x'; break;
+        }
+        fprintf(fout, "  f_st0 %c= (%s)%s;", j, float_type,
+          out_src_opr(buf1, sizeof(buf1), po, &po->operand[0],
+            lmod_cast(po, po->operand[0].lmod, 1), 0));
+        break;
+
+      case OP_FIDIVR:
+      case OP_FISUBR:
+        fprintf(fout, "  f_st0 = %s %c f_st0;",
+          out_src_opr_float(buf2, sizeof(buf2), po, &po->operand[1]),
+          po->op == OP_FIDIVR ? '/' : '-');
+        break;
+
+      case OP_FCOS:
+        fprintf(fout, "  f_st0 = cos%s(f_st0);",
+          need_double ? "" : "f");
+        break;
+
+      case OP_FPATAN:
+        fprintf(fout, "  f_st0 = atan%s(f_st1 / f_st0);",
+          need_double ? "" : "f");
+        break;
+
+      case OP_FSIN:
+        fprintf(fout, "  f_st0 = sin%s(f_st0);",
+          need_double ? "" : "f");
+        break;
+
+      case OP_FSQRT:
+        fprintf(fout, "  f_st0 = sqrt%s(f_st0);",
+          need_double ? "" : "f");
         break;
 
-      case OP_DIV:
-      case OP_IDIV:
-        assert_operand_cnt(1);
-        if (po->operand[0].lmod != OPLM_DWORD)
-          ferr(po, "unhandled lmod %d\n", po->operand[0].lmod);
+      case OPP_FTOL:
+        ferr_assert(po, po->flags & OPF_32BIT);
+        fprintf(fout, "  eax = (s32)f_st0;");
+        if (po->flags & OPF_FSHIFT)
+          fprintf(fout, "\n  f_st0 = f_st1;");
+        strcat(g_comment, " ftol");
+        break;
+
+      // mmx
+      case OP_EMMS:
+        strcpy(g_comment, " (emms)");
+        break;
+
+      default:
+        no_output = 1;
+        ferr(po, "unhandled op type %d, flags %x\n",
+          po->op, po->flags);
+        break;
+    }
+
+    if (g_comment[0] != 0) {
+      char *p = g_comment;
+      while (my_isblank(*p))
+        p++;
+      fprintf(fout, "  // %s", p);
+      g_comment[0] = 0;
+      no_output = 0;
+    }
+    if (!no_output)
+      fprintf(fout, "\n");
+
+    // some sanity checking
+    if (po->flags & OPF_REP) {
+      if (po->op != OP_STOS && po->op != OP_MOVS
+          && po->op != OP_CMPS && po->op != OP_SCAS)
+        ferr(po, "unexpected rep\n");
+      if (!(po->flags & (OPF_REPZ|OPF_REPNZ))
+          && (po->op == OP_CMPS || po->op == OP_SCAS))
+        ferr(po, "cmps/scas with plain rep\n");
+    }
+    if ((po->flags & (OPF_REPZ|OPF_REPNZ))
+        && po->op != OP_CMPS && po->op != OP_SCAS)
+      ferr(po, "unexpected repz/repnz\n");
+
+    if (pfomask != 0)
+      ferr(po, "missed flag calc, pfomask=%x\n", pfomask);
+
+    // see is delayed flag stuff is still valid
+    if (delayed_flag_op != NULL && delayed_flag_op != po) {
+      if (is_any_opr_modified(delayed_flag_op, po, 0))
+        delayed_flag_op = NULL;
+    }
+
+    if (last_arith_dst != NULL && last_arith_dst != &po->operand[0]) {
+      if (is_opr_modified(last_arith_dst, po))
+        last_arith_dst = NULL;
+    }
+
+    label_pending = 0;
+  }
+
+  if (g_stack_fsz && !g_stack_frame_used)
+    fprintf(fout, "  (void)sf;\n");
+
+  fprintf(fout, "}\n\n");
+
+  gen_x_cleanup(opcnt);
+}
+
+static void gen_x_cleanup(int opcnt)
+{
+  int i;
+
+  for (i = 0; i < opcnt; i++) {
+    struct label_ref *lr, *lr_del;
+
+    lr = g_label_refs[i].next;
+    while (lr != NULL) {
+      lr_del = lr;
+      lr = lr->next;
+      free(lr_del);
+    }
+    g_label_refs[i].i = -1;
+    g_label_refs[i].next = NULL;
+
+    if (ops[i].op == OP_CALL) {
+      if (ops[i].pp)
+        proto_release(ops[i].pp);
+    }
+  }
+  g_func_pp = NULL;
+}
+
+struct func_proto_dep;
+
+struct func_prototype {
+  char name[NAMELEN];
+  int id;
+  int argc_stack;
+  int regmask_dep;
+  int has_ret:3;                 // -1, 0, 1: unresolved, no, yes
+  unsigned int dep_resolved:1;
+  unsigned int is_stdcall:1;
+  struct func_proto_dep *dep_func;
+  int dep_func_cnt;
+  const struct parsed_proto *pp; // seed pp, if any
+};
+
+struct func_proto_dep {
+  char *name;
+  struct func_prototype *proto;
+  int regmask_live;             // .. at the time of call
+  unsigned int ret_dep:1;       // return from this is caller's return
+};
+
+static struct func_prototype *hg_fp;
+static int hg_fp_cnt;
+
+static struct scanned_var {
+  char name[NAMELEN];
+  enum opr_lenmod lmod;
+  unsigned int is_seeded:1;
+  unsigned int is_c_str:1;
+  const struct parsed_proto *pp; // seed pp, if any
+} *hg_vars;
+static int hg_var_cnt;
+
+static char **hg_refs;
+static int hg_ref_cnt;
+
+static void output_hdr_fp(FILE *fout, const struct func_prototype *fp,
+  int count);
+
+static struct func_prototype *hg_fp_add(const char *funcn)
+{
+  struct func_prototype *fp;
+
+  if ((hg_fp_cnt & 0xff) == 0) {
+    hg_fp = realloc(hg_fp, sizeof(hg_fp[0]) * (hg_fp_cnt + 0x100));
+    my_assert_not(hg_fp, NULL);
+    memset(hg_fp + hg_fp_cnt, 0, sizeof(hg_fp[0]) * 0x100);
+  }
+
+  fp = &hg_fp[hg_fp_cnt];
+  snprintf(fp->name, sizeof(fp->name), "%s", funcn);
+  fp->id = hg_fp_cnt;
+  fp->argc_stack = -1;
+  hg_fp_cnt++;
+
+  return fp;
+}
+
+static struct func_proto_dep *hg_fp_find_dep(struct func_prototype *fp,
+  const char *name)
+{
+  int i;
+
+  for (i = 0; i < fp->dep_func_cnt; i++)
+    if (IS(fp->dep_func[i].name, name))
+      return &fp->dep_func[i];
+
+  return NULL;
+}
+
+static void hg_fp_add_dep(struct func_prototype *fp, const char *name)
+{
+  // is it a dupe?
+  if (hg_fp_find_dep(fp, name))
+    return;
+
+  if ((fp->dep_func_cnt & 0xff) == 0) {
+    fp->dep_func = realloc(fp->dep_func,
+      sizeof(fp->dep_func[0]) * (fp->dep_func_cnt + 0x100));
+    my_assert_not(fp->dep_func, NULL);
+    memset(&fp->dep_func[fp->dep_func_cnt], 0,
+      sizeof(fp->dep_func[0]) * 0x100);
+  }
+  fp->dep_func[fp->dep_func_cnt].name = strdup(name);
+  fp->dep_func_cnt++;
+}
+
+static int hg_fp_cmp_name(const void *p1_, const void *p2_)
+{
+  const struct func_prototype *p1 = p1_, *p2 = p2_;
+  return strcmp(p1->name, p2->name);
+}
+
+#if 0
+static int hg_fp_cmp_id(const void *p1_, const void *p2_)
+{
+  const struct func_prototype *p1 = p1_, *p2 = p2_;
+  return p1->id - p2->id;
+}
+#endif
+
+static void hg_ref_add(const char *name)
+{
+  if ((hg_ref_cnt & 0xff) == 0) {
+    hg_refs = realloc(hg_refs, sizeof(hg_refs[0]) * (hg_ref_cnt + 0x100));
+    my_assert_not(hg_refs, NULL);
+    memset(hg_refs + hg_ref_cnt, 0, sizeof(hg_refs[0]) * 0x100);
+  }
+
+  hg_refs[hg_ref_cnt] = strdup(name);
+  my_assert_not(hg_refs[hg_ref_cnt], NULL);
+  hg_ref_cnt++;
+}
+
+// recursive register dep pass
+// - track saved regs (part 2)
+// - try to figure out arg-regs
+// - calculate reg deps
+static void gen_hdr_dep_pass(int i, int opcnt, unsigned char *cbits,
+  struct func_prototype *fp, int regmask_save, int regmask_dst,
+  int *regmask_dep, int *has_ret)
+{
+  struct func_proto_dep *dep;
+  struct parsed_op *po;
+  int from_caller = 0;
+  int j, l;
+  int reg;
+  int ret;
+
+  for (; i < opcnt; i++)
+  {
+    if (cbits[i >> 3] & (1 << (i & 7)))
+      return;
+    cbits[i >> 3] |= (1 << (i & 7));
+
+    po = &ops[i];
+
+    if ((po->flags & OPF_JMP) && po->op != OP_CALL) {
+      if (po->flags & OPF_RMD)
+        continue;
+
+      if (po->btj != NULL) {
+        // jumptable
+        for (j = 0; j < po->btj->count; j++) {
+          check_i(po, po->btj->d[j].bt_i);
+          gen_hdr_dep_pass(po->btj->d[j].bt_i, opcnt, cbits, fp,
+            regmask_save, regmask_dst, regmask_dep, has_ret);
+        }
+        return;
+      }
+
+      check_i(po, po->bt_i);
+      if (po->flags & OPF_CJMP) {
+        gen_hdr_dep_pass(po->bt_i, opcnt, cbits, fp,
+          regmask_save, regmask_dst, regmask_dep, has_ret);
+      }
+      else {
+        i = po->bt_i - 1;
+      }
+      continue;
+    }
+
+    if (po->flags & OPF_FARG)
+      /* (just calculate register deps) */;
+    else if (po->op == OP_PUSH && po->operand[0].type == OPT_REG)
+    {
+      reg = po->operand[0].reg;
+      ferr_assert(po, reg >= 0);
+
+      if (po->flags & OPF_RSAVE) {
+        regmask_save |= 1 << reg;
+        continue;
+      }
+      if (po->flags & OPF_DONE)
+        continue;
+
+      ret = scan_for_pop(i + 1, opcnt, i + opcnt * 2, reg, 0, 0);
+      if (ret == 1) {
+        regmask_save |= 1 << reg;
+        po->flags |= OPF_RMD;
+        scan_for_pop(i + 1, opcnt, i + opcnt * 3, reg, 0, OPF_RMD);
+        continue;
+      }
+    }
+    else if (po->flags & OPF_RMD)
+      continue;
+    else if (po->op == OP_CALL) {
+      po->regmask_dst |= 1 << xAX;
+
+      dep = hg_fp_find_dep(fp, po->operand[0].name);
+      if (dep != NULL)
+        dep->regmask_live = regmask_save | regmask_dst;
+    }
+    else if (po->op == OP_RET) {
+      if (po->operand_cnt > 0) {
+        fp->is_stdcall = 1;
+        if (fp->argc_stack >= 0
+            && fp->argc_stack != po->operand[0].val / 4)
+          ferr(po, "ret mismatch? (%d)\n", fp->argc_stack * 4);
+        fp->argc_stack = po->operand[0].val / 4;
+      }
+    }
+
+    // if has_ret is 0, there is uninitialized eax path,
+    // which means it's most likely void func
+    if (*has_ret != 0 && (po->flags & OPF_TAIL)) {
+      if (po->op == OP_CALL) {
+        j = i;
+        ret = 1;
+      }
+      else {
+        struct parsed_opr opr = OPR_INIT(OPT_REG, OPLM_DWORD, xAX);
+        j = -1;
+        from_caller = 0;
+        ret = resolve_origin(i, &opr, i + opcnt * 4, &j, &from_caller);
+      }
 
-        // 32bit division is common, look for it
-        if (po->op == OP_DIV)
-          ret = scan_for_reg_clear(i, xDX);
-        else
-          ret = scan_for_cdq_edx(i);
-        if (ret >= 0) {
-          out_src_opr_u32(buf1, sizeof(buf1), po, &po->operand[0]);
-          strcpy(buf2, lmod_cast(po, po->operand[0].lmod,
-            po->op == OP_IDIV));
-          fprintf(fout, "  edx = %seax %% %s%s;\n", buf2, buf2, buf1);
-          fprintf(fout, "  eax = %seax / %s%s;", buf2, buf2, buf1);
+      if (ret != 1 && from_caller) {
+        // unresolved eax - probably void func
+        *has_ret = 0;
+      }
+      else {
+        if (j >= 0 && ops[j].op == OP_CALL) {
+          dep = hg_fp_find_dep(fp, ops[j].operand[0].name);
+          if (dep != NULL)
+            dep->ret_dep = 1;
+          else
+            *has_ret = 1;
         }
         else
-          ferr(po, "TODO 64bit divident\n");
-        last_arith_dst = NULL;
-        delayed_flag_op = NULL;
-        break;
+          *has_ret = 1;
+      }
+    }
 
-      case OP_TEST:
-      case OP_CMP:
-        propagate_lmod(po, &po->operand[0], &po->operand[1]);
-        if (pfomask != 0) {
-          for (j = 0; j < 8; j++) {
-            if (pfomask & (1 << j)) {
-              out_cmp_test(buf1, sizeof(buf1), po, j, 0);
-              fprintf(fout, "  cond_%s = %s;",
-                parsed_flag_op_names[j], buf1);
-            }
+    l = regmask_save | regmask_dst;
+    if (g_bp_frame && !(po->flags & OPF_EBP_S))
+      l |= 1 << xBP;
+
+    l = po->regmask_src & ~l;
+#if 0
+    if (l)
+      fnote(po, "dep |= %04x, dst %04x, save %04x (f %x)\n",
+        l, regmask_dst, regmask_save, po->flags);
+#endif
+    *regmask_dep |= l;
+    regmask_dst |= po->regmask_dst;
+
+    if (po->flags & OPF_TAIL)
+      return;
+  }
+}
+
+static void gen_hdr(const char *funcn, int opcnt)
+{
+  int save_arg_vars[MAX_ARG_GRP] = { 0, };
+  unsigned char cbits[MAX_OPS / 8];
+  const struct parsed_proto *pp_c;
+  struct parsed_proto *pp;
+  struct func_prototype *fp;
+  struct parsed_op *po;
+  int regmask_dummy = 0;
+  int regmask_dep;
+  int max_bp_offset = 0;
+  int has_ret;
+  int i, j, l;
+  int ret;
+
+  pp_c = proto_parse(g_fhdr, funcn, 1);
+  if (pp_c != NULL)
+    // already in seed, will add to hg_fp later
+    return;
+
+  fp = hg_fp_add(funcn);
+
+  g_bp_frame = g_sp_frame = g_stack_fsz = 0;
+  g_stack_frame_used = 0;
+
+  // pass1:
+  // - resolve all branches
+  // - parse calls with labels
+  resolve_branches_parse_calls(opcnt);
+
+  // pass2:
+  // - handle ebp/esp frame, remove ops related to it
+  scan_prologue_epilogue(opcnt);
+
+  // pass3:
+  // - remove dead labels
+  // - collect calls
+  for (i = 0; i < opcnt; i++)
+  {
+    if (g_labels[i] != NULL && g_label_refs[i].i == -1) {
+      free(g_labels[i]);
+      g_labels[i] = NULL;
+    }
+
+    po = &ops[i];
+    if (po->flags & (OPF_RMD|OPF_DONE))
+      continue;
+
+    if (po->op == OP_CALL) {
+      if (po->operand[0].type == OPT_LABEL)
+        hg_fp_add_dep(fp, opr_name(po, 0));
+      else if (po->pp != NULL)
+        hg_fp_add_dep(fp, po->pp->name);
+    }
+  }
+
+  // pass4:
+  // - remove dead labels
+  // - handle push <const>/pop pairs
+  for (i = 0; i < opcnt; i++)
+  {
+    if (g_labels[i] != NULL && g_label_refs[i].i == -1) {
+      free(g_labels[i]);
+      g_labels[i] = NULL;
+    }
+
+    po = &ops[i];
+    if (po->flags & (OPF_RMD|OPF_DONE))
+      continue;
+
+    if (po->op == OP_PUSH && po->operand[0].type == OPT_CONST)
+      scan_for_pop_const(i, opcnt, i + opcnt * 13);
+  }
+
+  // pass5:
+  // - process trivial calls
+  for (i = 0; i < opcnt; i++)
+  {
+    po = &ops[i];
+    if (po->flags & (OPF_RMD|OPF_DONE))
+      continue;
+
+    if (po->op == OP_CALL)
+    {
+      pp = process_call_early(i, opcnt, &j);
+      if (pp != NULL) {
+        if (!(po->flags & OPF_ATAIL))
+          // since we know the args, try to collect them
+          if (collect_call_args_early(po, i, pp, &regmask_dummy) != 0)
+            pp = NULL;
+      }
+
+      if (pp != NULL) {
+        if (j >= 0) {
+          // commit esp adjust
+          if (ops[j].op != OP_POP)
+            patch_esp_adjust(&ops[j], pp->argc_stack * 4);
+          else {
+            for (l = 0; l < pp->argc_stack; l++)
+              ops[j + l].flags |= OPF_DONE | OPF_RMD | OPF_NOREGS;
           }
-          pfomask = 0;
         }
+
+        po->flags |= OPF_DONE;
+      }
+    }
+  }
+
+  // pass6:
+  // - track saved regs (simple)
+  // - process calls
+  for (i = 0; i < opcnt; i++)
+  {
+    po = &ops[i];
+    if (po->flags & (OPF_RMD|OPF_DONE))
+      continue;
+
+    if (po->op == OP_PUSH && po->operand[0].type == OPT_REG
+      && po->operand[0].reg != xCX)
+    {
+      ret = scan_for_pop_ret(i + 1, opcnt, po->operand[0].reg, 0);
+      if (ret == 1) {
+        // regmask_save |= 1 << po->operand[0].reg; // do it later
+        po->flags |= OPF_RSAVE | OPF_RMD | OPF_DONE;
+        scan_for_pop_ret(i + 1, opcnt, po->operand[0].reg, OPF_RMD);
+      }
+    }
+    else if (po->op == OP_CALL)
+    {
+      pp = process_call(i, opcnt);
+
+      if (!pp->is_unresolved && !(po->flags & OPF_ATAIL)) {
+        // since we know the args, collect them
+        ret = collect_call_args(po, i, pp, &regmask_dummy, save_arg_vars,
+          i + opcnt * 1);
+      }
+    }
+  }
+
+  // pass7
+  memset(cbits, 0, sizeof(cbits));
+  regmask_dep = 0;
+  has_ret = -1;
+
+  gen_hdr_dep_pass(0, opcnt, cbits, fp, 0, 0, &regmask_dep, &has_ret);
+
+  // find unreachable code - must be fixed in IDA
+  for (i = 0; i < opcnt; i++)
+  {
+    if (cbits[i >> 3] & (1 << (i & 7)))
+      continue;
+
+    if (g_labels[i] == NULL && i > 0 && ops[i - 1].op == OP_CALL
+      && ops[i - 1].pp != NULL && ops[i - 1].pp->is_osinc)
+    {
+      // the compiler sometimes still generates code after
+      // noreturn OS functions
+      break;
+    }
+    if (ops[i].op != OP_NOP)
+      ferr(&ops[i], "unreachable code\n");
+  }
+
+  for (i = 0; i < g_eqcnt; i++) {
+    if (g_eqs[i].offset > max_bp_offset && g_eqs[i].offset < 4*32)
+      max_bp_offset = g_eqs[i].offset;
+  }
+
+  if (fp->argc_stack < 0) {
+    max_bp_offset = (max_bp_offset + 3) & ~3;
+    fp->argc_stack = max_bp_offset / 4;
+    if ((g_ida_func_attr & IDAFA_BP_FRAME) && fp->argc_stack > 0)
+      fp->argc_stack--;
+  }
+
+  fp->regmask_dep = regmask_dep & ~(1 << xSP);
+  fp->has_ret = has_ret;
+#if 0
+  printf("// has_ret %d, regmask_dep %x\n",
+    fp->has_ret, fp->regmask_dep);
+  output_hdr_fp(stdout, fp, 1);
+  if (IS(funcn, "sub_10007F72")) exit(1);
+#endif
+
+  gen_x_cleanup(opcnt);
+}
+
+static void hg_fp_resolve_deps(struct func_prototype *fp)
+{
+  struct func_prototype fp_s;
+  int dep;
+  int i;
+
+  // this thing is recursive, so mark first..
+  fp->dep_resolved = 1;
+
+  for (i = 0; i < fp->dep_func_cnt; i++) {
+    strcpy(fp_s.name, fp->dep_func[i].name);
+    fp->dep_func[i].proto = bsearch(&fp_s, hg_fp, hg_fp_cnt,
+      sizeof(hg_fp[0]), hg_fp_cmp_name);
+    if (fp->dep_func[i].proto != NULL) {
+      if (!fp->dep_func[i].proto->dep_resolved)
+        hg_fp_resolve_deps(fp->dep_func[i].proto);
+
+      dep = ~fp->dep_func[i].regmask_live
+           & fp->dep_func[i].proto->regmask_dep;
+      fp->regmask_dep |= dep;
+      // printf("dep %s %s |= %x\n", fp->name,
+      //   fp->dep_func[i].name, dep);
+
+      if (fp->has_ret == -1 && fp->dep_func[i].ret_dep)
+        fp->has_ret = fp->dep_func[i].proto->has_ret;
+    }
+  }
+}
+
+// make all thiscall/edx arg functions referenced from .data fastcall
+static void do_func_refs_from_data(void)
+{
+  struct func_prototype *fp, fp_s;
+  int i;
+
+  for (i = 0; i < hg_ref_cnt; i++) {
+    strcpy(fp_s.name, hg_refs[i]);
+    fp = bsearch(&fp_s, hg_fp, hg_fp_cnt,
+      sizeof(hg_fp[0]), hg_fp_cmp_name);
+    if (fp == NULL)
+      continue;
+
+    if (fp->argc_stack != 0 && (fp->regmask_dep & (mxCX | mxDX)))
+      fp->regmask_dep |= mxCX | mxDX;
+  }
+}
+
+static void output_hdr_fp(FILE *fout, const struct func_prototype *fp,
+  int count)
+{
+  const struct parsed_proto *pp;
+  char *p, namebuf[NAMELEN];
+  const char *name;
+  int regmask_dep;
+  int argc_normal;
+  int j, arg;
+
+  for (; count > 0; count--, fp++) {
+    if (fp->has_ret == -1)
+      fprintf(fout, "// ret unresolved\n");
+#if 0
+    fprintf(fout, "// dep:");
+    for (j = 0; j < fp->dep_func_cnt; j++) {
+      fprintf(fout, " %s/", fp->dep_func[j].name);
+      if (fp->dep_func[j].proto != NULL)
+        fprintf(fout, "%04x/%d", fp->dep_func[j].proto->regmask_dep,
+          fp->dep_func[j].proto->has_ret);
+    }
+    fprintf(fout, "\n");
+#endif
+
+    p = strchr(fp->name, '@');
+    if (p != NULL) {
+      memcpy(namebuf, fp->name, p - fp->name);
+      namebuf[p - fp->name] = 0;
+      name = namebuf;
+    }
+    else
+      name = fp->name;
+    if (name[0] == '_')
+      name++;
+
+    pp = proto_parse(g_fhdr, name, 1);
+    if (pp != NULL && pp->is_include)
+      continue;
+
+    if (fp->pp != NULL) {
+      // part of seed, output later
+      continue;
+    }
+
+    regmask_dep = fp->regmask_dep;
+    argc_normal = fp->argc_stack;
+
+    fprintf(fout, "%-5s", fp->pp ? fp->pp->ret_type.name :
+      (fp->has_ret ? "int" : "void"));
+    if (regmask_dep && (fp->is_stdcall || fp->argc_stack > 0)
+      && (regmask_dep & ~mxCX) == 0)
+    {
+      fprintf(fout, "/*__thiscall*/  ");
+      argc_normal++;
+      regmask_dep = 0;
+    }
+    else if (regmask_dep && (fp->is_stdcall || fp->argc_stack == 0)
+      && (regmask_dep & ~(mxCX | mxDX)) == 0)
+    {
+      fprintf(fout, "  __fastcall    ");
+      if (!(regmask_dep & (1 << xDX)) && fp->argc_stack == 0)
+        argc_normal = 1;
+      else
+        argc_normal += 2;
+      regmask_dep = 0;
+    }
+    else if (regmask_dep && !fp->is_stdcall) {
+      fprintf(fout, "/*__usercall*/  ");
+    }
+    else if (regmask_dep) {
+      fprintf(fout, "/*__userpurge*/ ");
+    }
+    else if (fp->is_stdcall)
+      fprintf(fout, "  __stdcall     ");
+    else
+      fprintf(fout, "  __cdecl       ");
+
+    fprintf(fout, "%s(", name);
+
+    arg = 0;
+    for (j = 0; j < xSP; j++) {
+      if (regmask_dep & (1 << j)) {
+        arg++;
+        if (arg != 1)
+          fprintf(fout, ", ");
+        if (fp->pp != NULL)
+          fprintf(fout, "%s", fp->pp->arg[arg - 1].type.name);
         else
-          no_output = 1;
-        last_arith_dst = NULL;
-        delayed_flag_op = po;
-        break;
+          fprintf(fout, "int");
+        fprintf(fout, " a%d/*<%s>*/", arg, regs_r32[j]);
+      }
+    }
 
-      // note: we reuse OP_Jcc for SETcc, only flags differ
-      case OP_JO ... OP_JG:
-        if (po->flags & OPF_JMP)
-          fprintf(fout, "    goto %s;", po->operand[0].name);
-        // else SETcc - should already be handled
+    for (j = 0; j < argc_normal; j++) {
+      arg++;
+      if (arg != 1)
+        fprintf(fout, ", ");
+      if (fp->pp != NULL) {
+        fprintf(fout, "%s", fp->pp->arg[arg - 1].type.name);
+        if (!fp->pp->arg[arg - 1].type.is_ptr)
+          fprintf(fout, " ");
+      }
+      else
+        fprintf(fout, "int ");
+      fprintf(fout, "a%d", arg);
+    }
+
+    fprintf(fout, ");\n");
+  }
+}
+
+static void output_hdr(FILE *fout)
+{
+  static const char *lmod_c_names[] = {
+    [OPLM_UNSPEC] = "???",
+    [OPLM_BYTE]  = "uint8_t",
+    [OPLM_WORD]  = "uint16_t",
+    [OPLM_DWORD] = "uint32_t",
+    [OPLM_QWORD] = "uint64_t",
+  };
+  const struct scanned_var *var;
+  struct func_prototype *fp;
+  char line[256] = { 0, };
+  char name[256];
+  int i;
+
+  // add stuff from headers
+  for (i = 0; i < pp_cache_size; i++) {
+    if (pp_cache[i].is_cinc && !pp_cache[i].is_stdcall)
+      snprintf(name, sizeof(name), "_%s", pp_cache[i].name);
+    else
+      snprintf(name, sizeof(name), "%s", pp_cache[i].name);
+    fp = hg_fp_add(name);
+    fp->pp = &pp_cache[i];
+    fp->argc_stack = fp->pp->argc_stack;
+    fp->is_stdcall = fp->pp->is_stdcall;
+    fp->regmask_dep = get_pp_arg_regmask_src(fp->pp);
+    fp->has_ret = !IS(fp->pp->ret_type.name, "void");
+  }
+
+  // resolve deps
+  qsort(hg_fp, hg_fp_cnt, sizeof(hg_fp[0]), hg_fp_cmp_name);
+  for (i = 0; i < hg_fp_cnt; i++)
+    hg_fp_resolve_deps(&hg_fp[i]);
+
+  // adjust functions referenced from data segment
+  do_func_refs_from_data();
+
+  // note: messes up .proto ptr, don't use
+  //qsort(hg_fp, hg_fp_cnt, sizeof(hg_fp[0]), hg_fp_cmp_id);
+
+  // output variables
+  for (i = 0; i < hg_var_cnt; i++) {
+    var = &hg_vars[i];
+
+    if (var->pp != NULL)
+      // part of seed
+      continue;
+    else if (var->is_c_str)
+      fprintf(fout, "extern %-8s %s[];", "char", var->name);
+    else
+      fprintf(fout, "extern %-8s %s;",
+        lmod_c_names[var->lmod], var->name);
+
+    if (var->is_seeded)
+      fprintf(fout, " // seeded");
+    fprintf(fout, "\n");
+  }
+
+  fprintf(fout, "\n");
+
+  // output function prototypes
+  output_hdr_fp(fout, hg_fp, hg_fp_cnt);
+
+  // seed passthrough
+  fprintf(fout, "\n// - seed -\n");
+
+  rewind(g_fhdr);
+  while (fgets(line, sizeof(line), g_fhdr))
+    fwrite(line, 1, strlen(line), fout);
+}
+
+// '=' needs special treatment
+// also ' quote
+static char *next_word_s(char *w, size_t wsize, char *s)
+{
+  size_t i;
+
+  s = sskip(s);
+
+  i = 0;
+  if (*s == '\'' && s[1] != '\r' && s[1] != '\n') {
+    w[0] = s[0];
+    for (i = 1; i < wsize - 1; i++) {
+      if (s[i] == 0) {
+        printf("warning: missing closing quote: \"%s\"\n", s);
+        break;
+      }
+      if (s[i] == '\'')
         break;
+      w[i] = s[i];
+    }
+  }
 
-      case OP_JMP:
-        assert_operand_cnt(1);
-        last_arith_dst = NULL;
-        delayed_flag_op = NULL;
+  for (; i < wsize - 1; i++) {
+    if (s[i] == 0 || my_isblank(s[i]) || (s[i] == '=' && i > 0))
+      break;
+    w[i] = s[i];
+  }
+  w[i] = 0;
 
-        if (po->operand[0].type == OPT_REGMEM) {
-          ret = sscanf(po->operand[0].name, "%[^[][%[^*]*4]",
-                  buf1, buf2);
-          if (ret != 2)
-            ferr(po, "parse failure for jmp '%s'\n",
-              po->operand[0].name);
-          fprintf(fout, "  goto *jt_%s[%s];", buf1, buf2);
-          break;
-        }
-        else if (po->operand[0].type != OPT_LABEL)
-          ferr(po, "unhandled jmp type\n");
+  if (s[i] != 0 && !my_isblank(s[i]) && s[i] != '=')
+    printf("warning: '%s' truncated\n", w);
+
+  return s + i;
+}
+
+static int cmpstringp(const void *p1, const void *p2)
+{
+  return strcmp(*(char * const *)p1, *(char * const *)p2);
+}
+
+static int is_xref_needed(char *p, char **rlist, int rlist_len)
+{
+  char *p2;
+
+  p = sskip(p);
+  if (strstr(p, "..."))
+    // unable to determine, assume needed
+    return 1;
+
+  if (*p == '.') // .text, .data, ...
+    // ref from other data or non-function -> no
+    return 0;
+
+  p2 = strpbrk(p, "+:\r\n\x18");
+  if (p2 != NULL)
+    *p2 = 0;
+  if (bsearch(&p, rlist, rlist_len, sizeof(rlist[0]), cmpstringp))
+    // referenced from removed code
+    return 0;
+
+  return 1;
+}
 
-        fprintf(fout, "  goto %s;", po->operand[0].name);
-        break;
+static int ida_xrefs_show_need(FILE *fasm, char *p,
+  char **rlist, int rlist_len)
+{
+  int found_need = 0;
+  char line[256];
+  long pos;
 
-      case OP_CALL:
-        assert_operand_cnt(1);
-        pp = po->datap;
-        my_assert_not(pp, NULL);
+  p = strrchr(p, ';');
+  if (p != NULL && *p == ';' && IS_START(p + 2, "DATA XREF: ")) {
+    p += 13;
+    if (is_xref_needed(p, rlist, rlist_len))
+      return 1;
+  }
 
-        if (pp->is_fptr)
-          fprintf(fout, "  %s = %s;\n", pp->name,
-            out_src_opr(buf1, sizeof(buf1), po, &po->operand[0],
-              "(void *)", 0));
+  pos = ftell(fasm);
+  while (1)
+  {
+    if (!my_fgets(line, sizeof(line), fasm))
+      break;
+    // non-first line is always indented
+    if (!my_isblank(line[0]))
+      break;
 
-        fprintf(fout, "  ");
-        if (strstr(pp->ret_type.name, "int64")) {
-          if (po->flags & OPF_TAIL)
-            ferr(po, "int64 and tail?\n");
-          fprintf(fout, "mul_tmp = ");
-        }
-        else if (!IS(pp->ret_type.name, "void")) {
-          if (po->flags & OPF_TAIL) {
-            if (!IS(g_func_pp->ret_type.name, "void")) {
-              fprintf(fout, "return ");
-              if (g_func_pp->ret_type.is_ptr != pp->ret_type.is_ptr)
-                fprintf(fout, "(%s)", g_func_pp->ret_type.name);
-            }
-          }
-          else if (regmask & (1 << xAX)) {
-            fprintf(fout, "eax = ");
-            if (pp->ret_type.is_ptr)
-              fprintf(fout, "(u32)");
-          }
-        }
+    // should be no content, just comment
+    p = sskip(line);
+    if (*p != ';')
+      break;
 
-        if (pp->name[0] == 0)
-          ferr(po, "missing pp->name\n");
-        fprintf(fout, "%s%s(", pp->name,
-          pp->has_structarg ? "_sa" : "");
+    p = strrchr(p, ';');
+    p += 2;
+    // it's printed once, but no harm to check again
+    if (IS_START(p, "DATA XREF: "))
+      p += 11;
 
-        for (arg = 0; arg < pp->argc; arg++) {
-          if (arg > 0)
-            fprintf(fout, ", ");
+    if (is_xref_needed(p, rlist, rlist_len)) {
+      found_need = 1;
+      break;
+    }
+  }
+  fseek(fasm, pos, SEEK_SET);
+  return found_need;
+}
 
-          cast[0] = 0;
-          if (pp->arg[arg].type.is_ptr)
-            snprintf(cast, sizeof(cast), "(%s)", pp->arg[arg].type.name);
+static void scan_variables(FILE *fasm, char **rlist, int rlist_len)
+{
+  struct scanned_var *var;
+  char line[256] = { 0, };
+  char words[4][256];
+  int no_identifier;
+  char *p = NULL;
+  int wordc;
+  int l;
 
-          if (pp->arg[arg].reg != NULL) {
-            fprintf(fout, "%s%s", cast, pp->arg[arg].reg);
-            continue;
-          }
+  while (!feof(fasm))
+  {
+    // skip to next data section
+    while (my_fgets(line, sizeof(line), fasm))
+    {
+      asmln++;
 
-          // stack arg
-          tmp_op = pp->arg[arg].datap;
-          if (tmp_op == NULL)
-            ferr(po, "parsed_op missing for arg%d\n", arg);
-          if (tmp_op->argnum != 0) {
-            fprintf(fout, "%ss_a%d", cast, tmp_op->argnum);
-          }
-          else {
-            fprintf(fout, "%s",
-              out_src_opr(buf1, sizeof(buf1),
-                tmp_op, &tmp_op->operand[0], cast, 0));
-          }
-        }
-        fprintf(fout, ");");
+      p = sskip(line);
+      if (*p == 0 || *p == ';')
+        continue;
 
-        if (strstr(pp->ret_type.name, "int64")) {
-          fprintf(fout, "\n");
-          fprintf(fout, "  edx = mul_tmp >> 32;\n");
-          fprintf(fout, "  eax = mul_tmp;");
-        }
+      p = sskip(next_word_s(words[0], sizeof(words[0]), p));
+      if (*p == 0 || *p == ';')
+        continue;
 
-        if (pp->is_unresolved) {
-          snprintf(buf3, sizeof(buf3), " unresoved %dreg",
-            pp->argc_reg);
-          strcat(g_comment, buf3);
-        }
+      if (*p != 's' || !IS_START(p, "segment para public"))
+        continue;
 
-        if (po->flags & OPF_TAIL) {
-          ret = 0;
-          if (i == opcnt - 1 || pp->is_noreturn)
-            ret = 0;
-          else if (IS(pp->ret_type.name, "void"))
-            ret = 1;
-          else if (IS(g_func_pp->ret_type.name, "void"))
-            ret = 1;
-          // else already handled as 'return f()'
+      break;
+    }
 
-          if (ret) {
-            if (!IS(g_func_pp->ret_type.name, "void")) {
-              ferr(po, "int func -> void func tailcall?\n");
-            }
-            else {
-              fprintf(fout, "\n  return;");
-              strcat(g_comment, " ^ tailcall");
-            }
-          }
-          else
-            strcat(g_comment, " tailcall");
-        }
-        if (pp->is_noreturn)
-          strcat(g_comment, " noreturn");
-        delayed_flag_op = NULL;
-        last_arith_dst = NULL;
-        break;
+    if (p == NULL || !IS_START(p, "segment para public"))
+      break;
+    p = sskip(p + 19);
 
-      case OP_RET:
-        if (g_func_pp->is_vararg)
-          fprintf(fout, "  va_end(ap);\n");
- 
-        if (IS(g_func_pp->ret_type.name, "void")) {
-          if (i != opcnt - 1 || label_pending)
-            fprintf(fout, "  return;");
-        }
-        else if (g_func_pp->ret_type.is_ptr) {
-          fprintf(fout, "  return (%s)eax;",
-            g_func_pp->ret_type.name);
-        }
-        else
-          fprintf(fout, "  return eax;");
+    if (!IS_START(p, "'DATA'"))
+      continue;
 
-        last_arith_dst = NULL;
-        delayed_flag_op = NULL;
-        break;
+    // now process it
+    while (my_fgets(line, sizeof(line), fasm))
+    {
+      asmln++;
 
-      case OP_PUSH:
-        if (po->argnum != 0) {
-          // special case - saved func arg
-          out_src_opr_u32(buf1, sizeof(buf1), po, &po->operand[0]);
-          fprintf(fout, "  s_a%d = %s;", po->argnum, buf1);
-          break;
-        }
-        else if (po->flags & OPF_RSAVE) {
-          out_src_opr_u32(buf1, sizeof(buf1), po, &po->operand[0]);
-          fprintf(fout, "  s_%s = %s;", buf1, buf1);
-          break;
-        }
-        if (!(g_ida_func_attr & IDAFA_NORETURN))
-          ferr(po, "stray push encountered\n");
-        no_output = 1;
-        break;
+      p = line;
+      no_identifier = my_isblank(*p);
 
-      case OP_POP:
-        if (po->flags & OPF_RSAVE) {
-          out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]);
-          fprintf(fout, "  %s = s_%s;", buf1, buf1);
+      p = sskip(p);
+      if (*p == 0 || *p == ';')
+        continue;
+
+      for (wordc = 0; wordc < ARRAY_SIZE(words); wordc++) {
+        words[wordc][0] = 0;
+        p = sskip(next_word_s(words[wordc], sizeof(words[0]), p));
+        if (*p == 0 || *p == ';') {
+          wordc++;
           break;
         }
-        ferr(po, "stray pop encountered\n");
-        break;
-
-      case OP_NOP:
-        no_output = 1;
-        break;
+      }
 
-      default:
-        no_output = 1;
-        ferr(po, "unhandled op type %d, flags %x\n",
-          po->op, po->flags);
+      if (wordc == 2 && IS(words[1], "ends"))
         break;
-    }
-
-    if (g_comment[0] != 0) {
-      char *p = g_comment;
-      while (my_isblank(*p))
-        p++;
-      fprintf(fout, "  // %s", p);
-      g_comment[0] = 0;
-      no_output = 0;
-    }
-    if (!no_output)
-      fprintf(fout, "\n");
-
-    // some sanity checking
-    if ((po->flags & OPF_REP) && po->op != OP_STOS
-        && po->op != OP_MOVS && po->op != OP_CMPS)
-      ferr(po, "unexpected rep\n");
-    if ((po->flags & (OPF_REPZ|OPF_REPNZ)) && po->op != OP_CMPS)
-      ferr(po, "unexpected repz/repnz\n");
-
-    if (pfomask != 0)
-      ferr(po, "missed flag calc, pfomask=%x\n", pfomask);
+      if (wordc < 2)
+        continue;
 
-    // see is delayed flag stuff is still valid
-    if (delayed_flag_op != NULL && delayed_flag_op != po) {
-      if (is_any_opr_modified(delayed_flag_op, po, 0))
-        delayed_flag_op = NULL;
-    }
+      if (no_identifier) {
+        if (wordc >= 3 && IS(words[0], "dd") && IS(words[1], "offset"))
+          hg_ref_add(words[2]);
+        continue;
+      }
 
-    if (last_arith_dst != NULL && last_arith_dst != &po->operand[0]) {
-      if (is_opr_modified(last_arith_dst, po))
-        last_arith_dst = NULL;
-    }
+      if (IS_START(words[0], "__IMPORT_DESCRIPTOR_")) {
+        // when this starts, we don't need anything from this section
+        break;
+      }
 
-    label_pending = 0;
-  }
+      // check refs comment(s)
+      if (!ida_xrefs_show_need(fasm, p, rlist, rlist_len))
+        continue;
 
-  if (g_stack_fsz && !g_stack_frame_used)
-    fprintf(fout, "  (void)sf;\n");
+      if ((hg_var_cnt & 0xff) == 0) {
+        hg_vars = realloc(hg_vars, sizeof(hg_vars[0])
+                   * (hg_var_cnt + 0x100));
+        my_assert_not(hg_vars, NULL);
+        memset(hg_vars + hg_var_cnt, 0, sizeof(hg_vars[0]) * 0x100);
+      }
 
-  fprintf(fout, "}\n\n");
+      var = &hg_vars[hg_var_cnt++];
+      snprintf(var->name, sizeof(var->name), "%s", words[0]);
 
-  // cleanup
-  for (i = 0; i < opcnt; i++) {
-    struct label_ref *lr, *lr_del;
+      // maybe already in seed header?
+      var->pp = proto_parse(g_fhdr, var->name, 1);
+      if (var->pp != NULL) {
+        if (var->pp->is_fptr) {
+          var->lmod = OPLM_DWORD;
+          //var->is_ptr = 1;
+        }
+        else if (var->pp->is_func)
+          aerr("func?\n");
+        else if (!guess_lmod_from_c_type(&var->lmod, &var->pp->type))
+          aerr("unhandled C type '%s' for '%s'\n",
+            var->pp->type.name, var->name);
 
-    lr = g_label_refs[i].next;
-    while (lr != NULL) {
-      lr_del = lr;
-      lr = lr->next;
-      free(lr_del);
-    }
-    g_label_refs[i].i = -1;
-    g_label_refs[i].next = NULL;
+        var->is_seeded = 1;
+        continue;
+      }
 
-    if (ops[i].op == OP_CALL) {
-      pp = ops[i].datap;
-      if (pp)
-        proto_release(pp);
+      if      (IS(words[1], "dd")) {
+        var->lmod = OPLM_DWORD;
+        if (wordc >= 4 && IS(words[2], "offset"))
+          hg_ref_add(words[3]);
+      }
+      else if (IS(words[1], "dw"))
+        var->lmod = OPLM_WORD;
+      else if (IS(words[1], "db")) {
+        var->lmod = OPLM_BYTE;
+        if (wordc >= 3 && (l = strlen(words[2])) > 4) {
+          if (words[2][0] == '\'' && IS(words[2] + l - 2, ",0"))
+            var->is_c_str = 1;
+        }
+      }
+      else if (IS(words[1], "dq"))
+        var->lmod = OPLM_QWORD;
+      //else if (IS(words[1], "dt"))
+      else
+        aerr("type '%s' not known\n", words[1]);
     }
   }
-  g_func_pp = NULL;
+
+  rewind(fasm);
+  asmln = 0;
 }
 
 static void set_label(int i, const char *name)
@@ -4098,34 +7705,14 @@ static void set_label(int i, const char *name)
   if (p != NULL)
     len = p - name;
 
-  if (len > sizeof(g_labels[0]) - 1)
-    aerr("label '%s' too long: %d\n", name, len);
-  if (g_labels[i][0] != 0 && !IS_START(g_labels[i], "algn_"))
+  if (g_labels[i] != NULL && !IS_START(g_labels[i], "algn_"))
     aerr("dupe label '%s' vs '%s'?\n", name, g_labels[i]);
+  g_labels[i] = realloc(g_labels[i], len + 1);
+  my_assert_not(g_labels[i], NULL);
   memcpy(g_labels[i], name, len);
   g_labels[i][len] = 0;
 }
 
-// '=' needs special treatment..
-static char *next_word_s(char *w, size_t wsize, char *s)
-{
-	size_t i;
-
-	s = sskip(s);
-
-	for (i = 0; i < wsize - 1; i++) {
-		if (s[i] == 0 || my_isblank(s[i]) || (s[i] == '=' && i > 0))
-			break;
-		w[i] = s[i];
-	}
-	w[i] = 0;
-
-	if (s[i] != 0 && !my_isblank(s[i]) && s[i] != '=')
-		printf("warning: '%s' truncated\n", w);
-
-	return s + i;
-}
-
 struct chunk_item {
   char *name;
   long fptr;
@@ -4156,11 +7743,6 @@ static int cmp_chunks(const void *p1, const void *p2)
   return strcmp(c1->name, c2->name);
 }
 
-static int cmpstringp(const void *p1, const void *p2)
-{
-  return strcmp(*(char * const *)p1, *(char * const *)p2);
-}
-
 static void scan_ahead(FILE *fasm)
 {
   char words[2][256];
@@ -4174,7 +7756,7 @@ static void scan_ahead(FILE *fasm)
   oldpos = ftell(fasm);
   oldasmln = asmln;
 
-  while (fgets(line, sizeof(line), fasm))
+  while (my_fgets(line, sizeof(line), fasm))
   {
     wordc = 0;
     asmln++;
@@ -4199,6 +7781,9 @@ static void scan_ahead(FILE *fasm)
 
         add_func_chunk(fasm, words[0], asmln);
       }
+      else if (IS_START(p, "; sctend"))
+        break;
+
       continue;
     } // *p == ';'
 
@@ -4243,6 +7828,8 @@ int main(int argc, char *argv[])
   int skip_warned = 0;
   int eq_alloc;
   int verbose = 0;
+  int multi_seg = 0;
+  int end = 0;
   int arg_out;
   int arg;
   int pi = 0;
@@ -4256,13 +7843,24 @@ int main(int argc, char *argv[])
       verbose = 1;
     else if (IS(argv[arg], "-rf"))
       g_allow_regfunc = 1;
+    else if (IS(argv[arg], "-m"))
+      multi_seg = 1;
+    else if (IS(argv[arg], "-hdr"))
+      g_header_mode = g_quiet_pp = g_allow_regfunc = 1;
     else
       break;
   }
 
   if (argc < arg + 3) {
-    printf("usage:\n%s [-v] [-rf] <.c> <.asm> <hdrf> [rlist]*\n",
-      argv[0]);
+    printf("usage:\n%s [-v] [-rf] [-m] <.c> <.asm> <hdr.h> [rlist]*\n"
+           "%s -hdr <out.h> <.asm> <seed.h> [rlist]*\n"
+           "options:\n"
+           "  -hdr - header generation mode\n"
+           "  -rf  - allow unannotated indirect calls\n"
+           "  -m   - allow multiple .text sections\n"
+           "[rlist] is a file with function names to skip,"
+           " one per line\n",
+      argv[0], argv[0]);
     return 1;
   }
 
@@ -4292,7 +7890,7 @@ int main(int argc, char *argv[])
     frlist = fopen(argv[arg], "r");
     my_assert_not(frlist, NULL);
 
-    while (fgets(line, sizeof(line), frlist)) {
+    while (my_fgets(line, sizeof(line), frlist)) {
       p = sskip(line);
       if (*p == 0 || *p == ';')
         continue;
@@ -4341,7 +7939,10 @@ int main(int argc, char *argv[])
     g_label_refs[i].next = NULL;
   }
 
-  while (fgets(line, sizeof(line), fasm))
+  if (g_header_mode)
+    scan_variables(fasm, rlist, rlist_len);
+
+  while (my_fgets(line, sizeof(line), fasm))
   {
     wordc = 0;
     asmln++;
@@ -4393,6 +7994,46 @@ int main(int argc, char *argv[])
           }
         }
       }
+      else if (p[2] == 's' && IS_START(p, "; sctattr:"))
+      {
+        static const char *attrs[] = {
+          "clear_sf",
+          "clear_regmask",
+        };
+
+        // parse manual attribute-list comment
+        g_sct_func_attr = 0;
+        p = sskip(p + 10);
+
+        for (; *p != 0; p = sskip(p)) {
+          for (i = 0; i < ARRAY_SIZE(attrs); i++) {
+            if (!strncmp(p, attrs[i], strlen(attrs[i]))) {
+              g_sct_func_attr |= 1 << i;
+              p += strlen(attrs[i]);
+              break;
+            }
+          }
+          if (*p == '=') {
+            j = ret = 0;
+            if (i == 0)
+              // clear_sf=start,len (in dwords)
+              ret = sscanf(p, "=%d,%d%n", &g_stack_clear_start,
+                      &g_stack_clear_len, &j);
+            else if (i == 1)
+              // clear_regmask=<mask>
+              ret = sscanf(p, "=%x%n", &g_regmask_init, &j) + 1;
+            if (ret < 2) {
+              anote("unparsed attr value: %s\n", p);
+              break;
+            }
+            p += j;
+          }
+          else if (i == ARRAY_SIZE(attrs)) {
+            anote("unparsed sct attr: %s\n", p);
+            break;
+          }
+        }
+      }
       else if (p[2] == 'S' && IS_START(p, "; START OF FUNCTION CHUNK FOR "))
       {
         p += 30;
@@ -4436,8 +8077,8 @@ int main(int argc, char *argv[])
           unsigned long addr = strtoul(p, NULL, 16);
           unsigned long f_addr = strtoul(g_func + 4, NULL, 16);
           if (addr > f_addr && !scanned_ahead) {
-            anote("scan_ahead caused by '%s', addr %lx\n",
-              g_func, addr);
+            //anote("scan_ahead caused by '%s', addr %lx\n",
+            //  g_func, addr);
             scan_ahead(fasm);
             scanned_ahead = 1;
             func_chunks_sorted = 0;
@@ -4471,6 +8112,11 @@ parse_words:
       if (IS_START(p, "; sctproto:")) {
         sctproto = strdup(p + 11);
       }
+      else if (IS_START(p, "; sctend")) {
+        end = 1;
+        if (!pending_endp)
+          break;
+      }
     }
 
     if (wordc == 0) {
@@ -4491,7 +8137,7 @@ parse_words:
 do_pending_endp:
     // do delayed endp processing to collect switch jumptables
     if (pending_endp) {
-      if (in_func && !skip_func && wordc >= 2
+      if (in_func && !g_skip_func && !end && wordc >= 2
           && ((words[0][0] == 'd' && words[0][2] == 0)
               || (words[1][0] == 'd' && words[1][2] == 0)))
       {
@@ -4546,20 +8192,28 @@ do_pending_endp:
         continue;
       }
 
-      if (in_func && !skip_func)
-        gen_func(fout, g_fhdr, g_func, pi);
+      if (in_func && !g_skip_func) {
+        if (g_header_mode)
+          gen_hdr(g_func, pi);
+        else
+          gen_func(fout, g_fhdr, g_func, pi);
+      }
 
       pending_endp = 0;
       in_func = 0;
       g_ida_func_attr = 0;
+      g_sct_func_attr = 0;
+      g_stack_clear_start = 0;
+      g_stack_clear_len = 0;
+      g_regmask_init = 0;
       skip_warned = 0;
-      skip_func = 0;
+      g_skip_func = 0;
       g_func[0] = 0;
       func_chunks_used = 0;
       func_chunk_i = -1;
       if (pi != 0) {
         memset(&ops, 0, pi * sizeof(ops[0]));
-        memset(g_labels, 0, pi * sizeof(g_labels[0]));
+        clear_labels(pi);
         pi = 0;
       }
       g_eqcnt = 0;
@@ -4573,7 +8227,11 @@ do_pending_endp:
         pd->d = NULL;
       }
       g_func_pd_cnt = 0;
+      g_func_lmods = 0;
       pd = NULL;
+
+      if (end)
+        break;
       if (wordc == 0)
         continue;
     }
@@ -4584,7 +8242,7 @@ do_pending_endp:
           words[0], g_func);
       p = words[0];
       if (bsearch(&p, rlist, rlist_len, sizeof(rlist[0]), cmpstringp))
-        skip_func = 1;
+        g_skip_func = 1;
       strcpy(g_func, words[0]);
       set_label(0, words[0]);
       in_func = 1;
@@ -4603,10 +8261,10 @@ do_pending_endp:
         && ops[0].op == OP_JMP && ops[0].operand[0].had_ds)
       {
         // import jump
-        skip_func = 1;
+        g_skip_func = 1;
       }
 
-      if (!skip_func && func_chunks_used) {
+      if (!g_skip_func && func_chunks_used) {
         // start processing chunks
         struct chunk_item *ci, key = { g_func, 0 };
 
@@ -4637,8 +8295,27 @@ do_pending_endp:
       continue;
     }
 
-    if (wordc == 2 && IS(words[1], "ends"))
-      break;
+    if (wordc == 2 && IS(words[1], "ends")) {
+      if (!multi_seg) {
+        end = 1;
+        if (pending_endp)
+          goto do_pending_endp;
+        break;
+      }
+
+      // scan for next text segment
+      while (my_fgets(line, sizeof(line), fasm)) {
+        asmln++;
+        p = sskip(line);
+        if (*p == 0 || *p == ';')
+          continue;
+
+        if (strstr(p, "segment para public 'CODE' use32"))
+          break;
+      }
+
+      continue;
+    }
 
     p = strchr(words[0], ':');
     if (p != NULL) {
@@ -4646,13 +8323,14 @@ do_pending_endp:
       continue;
     }
 
-    if (!in_func || skip_func) {
-      if (!skip_warned && !skip_func && g_labels[pi][0] != 0) {
+    if (!in_func || g_skip_func) {
+      if (!skip_warned && !g_skip_func && g_labels[pi] != NULL) {
         if (verbose)
           anote("skipping from '%s'\n", g_labels[pi]);
         skip_warned = 1;
       }
-      g_labels[pi][0] = 0;
+      free(g_labels[pi]);
+      g_labels[pi] = NULL;
       continue;
     }
 
@@ -4679,6 +8357,8 @@ do_pending_endp:
         g_eqs[g_eqcnt].lmod = OPLM_WORD;
       else if (IS(words[2], "byte"))
         g_eqs[g_eqcnt].lmod = OPLM_BYTE;
+      else if (IS(words[2], "qword"))
+        g_eqs[g_eqcnt].lmod = OPLM_QWORD;
       else
         aerr("bad lmod: '%s'\n", words[2]);
 
@@ -4692,14 +8372,14 @@ do_pending_endp:
 
     parse_op(&ops[pi], words, wordc);
 
-    if (sctproto != NULL) {
-      if (ops[pi].op == OP_CALL)
-        ops[pi].datap = sctproto;
-      sctproto = NULL;
-    }
+    ops[pi].datap = sctproto;
+    sctproto = NULL;
     pi++;
   }
 
+  if (g_header_mode)
+    output_hdr(fout);
+
   fclose(fout);
   fclose(fasm);
   fclose(g_fhdr);