#include "my_str.h"
#include "common.h"
-#define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
-#define IS(w, y) !strcmp(w, y)
-#define IS_START(w, y) !strncmp(w, y, strlen(y))
-
#include "protoparse.h"
static const char *asmfn;
OPF_FPUSH = (1 << 22), /* pushes x87 stack */
OPF_FPOP = (1 << 23), /* pops x87 stack */
OPF_FSHIFT = (1 << 24), /* x87 stack shift is actually needed */
+ OPF_FINT = (1 << 25), /* integer float op arg */
};
enum op_op {
OP_ADC,
OP_SBB,
OP_BSF,
+ OP_BSR,
OP_INC,
OP_DEC,
OP_NEG,
#define mxAX (1 << xAX)
#define mxCX (1 << xCX)
#define mxDX (1 << xDX)
+#define mxSP (1 << xSP)
#define mxST0 (1 << xST0)
#define mxST1 (1 << xST1)
#define mxST1_0 (mxST1 | mxST0)
{ "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 },
+ { "bsr", OP_BSR, 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 },
{ "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 },
+ { "fild", OP_FILD, 1, 1, OPF_FPUSH|OPF_FINT },
{ "fld1", OP_FLDc, 0, 0, OPF_FPUSH },
{ "fldln2", OP_FLDc, 0, 0, OPF_FPUSH },
{ "fldz", OP_FLDc, 0, 0, OPF_FPUSH },
{ "fst", OP_FST, 1, 1, 0 },
{ "fstp", OP_FST, 1, 1, OPF_FPOP },
- { "fist", OP_FIST, 1, 1, 0 },
- { "fistp", OP_FIST, 1, 1, OPF_FPOP },
+ { "fist", OP_FIST, 1, 1, OPF_FINT },
+ { "fistp", OP_FIST, 1, 1, OPF_FPOP|OPF_FINT },
{ "fadd", OP_FADD, 0, 2, 0 },
{ "faddp", OP_FADD, 0, 2, OPF_FPOP },
{ "fdiv", OP_FDIV, 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 },
+ { "fiadd", OP_FIADD, 1, 1, OPF_FINT },
+ { "fidiv", OP_FIDIV, 1, 1, OPF_FINT },
+ { "fimul", OP_FIMUL, 1, 1, OPF_FINT },
+ { "fisub", OP_FISUB, 1, 1, OPF_FINT },
+ { "fidivr", OP_FIDIVR, 1, 1, OPF_FINT },
+ { "fisubr", OP_FISUBR, 1, 1, OPF_FINT },
{ "fcom", OP_FCOM, 0, 1, 0 },
{ "fcomp", OP_FCOM, 0, 1, OPF_FPOP },
{ "fnstsw", OP_FNSTSW, 1, 1, OPF_DATA },
{ "_allshr",OPP_ALLSHR },
{ "_ftol", OPP_FTOL },
{ "_CIpow", OPP_CIPOW },
+ { "abort", OPP_ABORT },
// must be last
{ "ud2", OP_UD2 },
};
int retval = -1;
int sf_ofs;
- if (po->flags & OPF_EBP_S)
+ if (g_bp_frame && (po->flags & OPF_EBP_S)
+ && !(po->regmask_src & mxSP))
ferr(po, "stack_frame_access while ebp is scratch\n");
parse_stack_access(po, name, ofs_reg, &offset,
}
break;
+ case OPLM_QWORD:
+ ferr_assert(po, !(offset & 7));
+ if (cast[0])
+ prefix = cast;
+ snprintf(buf, buf_size, "%s%sa%d",
+ prefix, is_lea ? "&" : "", i + 1);
+ break;
+
default:
ferr(po, "bp_arg bad lmod: %d\n", popr->lmod);
}
ferr_assert(po, !(sf_ofs & 7));
ferr_assert(po, ofs_reg[0] == 0);
// only used for x87 int64/float, float sets is_lea
- if (is_lea)
- snprintf(buf, buf_size, "%ssf.q[%d]", prefix, sf_ofs / 8);
- else
- snprintf(buf, buf_size, "*(s64 *)&sf.q[%d]", sf_ofs / 8);
+ if (!is_lea && (po->flags & OPF_FINT))
+ prefix = "*(s64 *)&";
+ snprintf(buf, buf_size, "%ssf.q[%d]", prefix, sf_ofs / 8);
break;
default:
return out_src_opr(buf, buf_size, po, popr, NULL, 0);
}
-static char *out_src_opr_float(char *buf, size_t buf_size,
- struct parsed_op *po, struct parsed_opr *popr, int need_float_stack)
+static char *out_opr_float(char *buf, size_t buf_size,
+ struct parsed_op *po, struct parsed_opr *popr, int is_src,
+ int need_float_stack)
{
const char *cast = NULL;
char tmp[256];
break;
case OPT_REGMEM:
+ if (popr->lmod == OPLM_QWORD && is_stack_access(po, popr)) {
+ stack_frame_access(po, popr, buf, buf_size,
+ popr->name, "", is_src, 0);
+ break;
+ }
+ // fallthrough
case OPT_LABEL:
case OPT_OFFSET:
switch (popr->lmod) {
return buf;
}
+static char *out_src_opr_float(char *buf, size_t buf_size,
+ struct parsed_op *po, struct parsed_opr *popr, int need_float_stack)
+{
+ return out_opr_float(buf, buf_size, po, popr, 1, need_float_stack);
+}
+
static char *out_dst_opr_float(char *buf, size_t buf_size,
struct parsed_op *po, struct parsed_opr *popr, int need_float_stack)
{
- // same?
- return out_src_opr_float(buf, buf_size, po, popr, need_float_stack);
+ return out_opr_float(buf, buf_size, po, popr, 0, need_float_stack);
}
static void out_test_for_cc(char *buf, size_t buf_size,
snprintf(buf, buf_size, "(%d)", !!is_inv);
break;
+ case PFO_P: // PF==1
+ snprintf(buf, buf_size, "(%sdo_parity(%s))",
+ is_inv ? "!" : "", expr);
+ break;
+
default:
ferr(po, "%s: unhandled parsed_flag_op: %d\n", __func__, pfo);
}
return -1;
}
- // nothing interesting on this path
- return 0;
+ // nothing interesting on this path,
+ // still return ret for something recursive calls could find
+ return ret;
}
static void find_reachable_exits(int i, int opcnt, int magic,
{
static int exits[MAX_EXITS];
static int exit_count;
- int j, ret;
+ int found = 0;
+ int e, j, ret;
if (!set_flags) {
exit_count = 0;
}
for (j = 0; j < exit_count; j++) {
- ret = scan_for_rsave_pop_reg(exits[j], i + opcnt * 16 + set_flags,
+ e = exits[j];
+ ret = scan_for_rsave_pop_reg(e, i + opcnt * 16 + set_flags,
reg, set_flags);
- if (ret == -1)
- return -1;
+ if (ret != -1) {
+ found |= ret;
+ continue;
+ }
+ if (ops[e].op == OP_CALL && ops[e].pp != NULL
+ && ops[e].pp->is_noreturn)
+ {
+ // assume stack cleanup was skipped
+ continue;
+ }
+ return -1;
}
- return 1;
+ return found;
}
// scan for one or more pop of push <const>
if (*pp_found != NULL && pp != NULL && *pp_found != pp) {
if (!IS((*pp_found)->ret_type.name, pp->ret_type.name)
|| (*pp_found)->is_stdcall != pp->is_stdcall
- || (*pp_found)->is_fptr != pp->is_fptr
+ //|| (*pp_found)->is_fptr != pp->is_fptr
|| (*pp_found)->argc != pp->argc
|| (*pp_found)->argc_reg != pp->argc_reg
|| (*pp_found)->argc_stack != pp->argc_stack)
{ "__allshl", OPP_ALLSHL, OPF_DATA, mxAX|mxDX|mxCX, mxAX|mxDX },
{ "__allshr", OPP_ALLSHR, OPF_DATA, mxAX|mxDX|mxCX, mxAX|mxDX },
{ "__ftol", OPP_FTOL, OPF_FPOP, mxST0, mxAX | mxDX },
+ // more precise? Wine gets away with just __ftol handler
+ { "__ftol2", OPP_FTOL, OPF_FPOP, mxST0, mxAX | mxDX },
{ "__CIpow", OPP_CIPOW, OPF_FPOP, mxST0|mxST1, mxST0 },
};
const struct parsed_proto *pp_c;
struct parsed_data *pd;
struct parsed_op *po;
const char *tmpname;
+ enum op_op prev_op;
int i, l;
int ret;
if (pp != NULL) {
if (pp->is_fptr)
check_func_pp(po, pp, "fptr var call");
- if (pp->is_noreturn)
+ if (pp->is_noreturn) {
po->flags |= OPF_TAIL;
+ po->flags &= ~OPF_ATAIL; // most likely...
+ }
}
po->pp = pp;
continue;
tailcall:
po->op = OP_CALL;
po->flags |= OPF_TAIL;
- if (i > 0 && ops[i - 1].op == OP_POP)
+ prev_op = i > 0 ? ops[i - 1].op : OP_UD2;
+ if (prev_op == OP_POP)
+ po->flags |= OPF_ATAIL;
+ if (g_stack_fsz + g_bp_frame == 0 && prev_op != OP_PUSH
+ && (g_func_pp == NULL || g_func_pp->argc_stack > 0))
+ {
po->flags |= OPF_ATAIL;
+ }
i--; // reprocess
}
}
-static void scan_prologue_epilogue(int opcnt)
+static void scan_prologue_epilogue(int opcnt, int *stack_align)
{
int ecx_push = 0, esp_sub = 0, pusha = 0;
int sandard_epilogue;
i++;
}
+ if (ops[i].op == OP_AND && ops[i].operand[0].reg == xSP
+ && ops[i].operand[1].type == OPT_CONST)
+ {
+ l = ops[i].operand[1].val;
+ j = ffs(l) - 1;
+ if (j == -1 || (l >> j) != -1)
+ ferr(&ops[i], "unhandled esp align: %x\n", l);
+ if (stack_align != NULL)
+ *stack_align = 1 << j;
+ ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
+ i++;
+ }
+
if (ops[i].op == OP_SUB && IS(opr_name(&ops[i], 0), "esp")) {
g_stack_fsz = opr_const(&ops[i], 1);
ops[i].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
l += ops[j].operand[1].val / 4 - 1;
}
else
- ferr(&ops[j], "'pop ecx' expected\n");
+ break;
ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
j--;
}
- if (l != ecx_push)
+ if (l != ecx_push) {
+ if (i < opcnt && ops[i].op == OP_CALL
+ && ops[i].pp != NULL && ops[i].pp->is_noreturn)
+ {
+ // noreturn tailcall with no epilogue
+ i++;
+ found = 1;
+ continue;
+ }
ferr(&ops[j], "epilogue scan failed\n");
+ }
found = 1;
}
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)
+ || ops[j].operand[1].type != OPT_CONST)
{
- if (ops[i].op == OP_CALL && ops[i].pp != NULL
- && ops[i].pp->is_noreturn)
+ if (i < opcnt && ops[i].op == OP_CALL
+ && ops[i].pp != NULL && ops[i].pp->is_noreturn)
{
// noreturn tailcall with no epilogue
i++;
+ found = 1;
continue;
}
ferr(&ops[j], "'add esp' expected\n");
}
- ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
- ops[j].operand[1].val = 0; // hack for stack arg scanner
+ if (ops[j].operand[1].val < g_stack_fsz)
+ ferr(&ops[j], "esp adj is too low (need %d)\n", g_stack_fsz);
+
+ ops[j].operand[1].val -= g_stack_fsz; // for stack arg scanner
+ if (ops[j].operand[1].val == 0)
+ ops[j].flags |= OPF_RMD | OPF_DONE | OPF_NOREGS;
found = 1;
}
ferr(po, "too many args for '%s'\n", tmpname);
}
if (pp->argc_stack > adj / 4) {
+ if (pp->is_noreturn)
+ // assume no stack adjust was emited
+ goto out;
fnote(po, "(this call)\n");
ferr(&ops[ret], "stack tracking failed for '%s': %x %x\n",
tmpname, pp->argc_stack * 4, adj);
ferr(po, "missing esp_adjust for vararg func '%s'\n",
pp->name);
+out:
return pp;
}
ret = parse_stack_esp_offset(po, po->operand[0].name, &offset);
if (ret != 0)
continue;
- if (offset < 0 || offset >= pp->argc_stack * 4 || (offset & 3))
- ferr(po, "bad offset %d (%d args)\n", offset, pp->argc_stack);
+ if (offset < 0 || offset >= pp->argc_stack * 4 || (offset & 3)) {
+ //ferr(po, "offset %d, %d args\n", offset, pp->argc_stack);
+ continue;
+ }
arg = base_arg + offset / 4;
po->p_argnext = -1;
pp_tmp = ops[j].pp;
if (pp_tmp == NULL)
- ferr(po, "arg collect hit unparsed call '%s'\n",
- ops[j].operand[0].name);
+ ferr(po, "arg collect %d/%d hit unparsed call '%s'\n",
+ arg, pp->argc, 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);
pp->argc_reg++;
}
-static void output_std_flags(FILE *fout, struct parsed_op *po,
+static void output_std_flag_z(FILE *fout, struct parsed_op *po,
int *pfomask, const char *dst_opr_text)
{
if (*pfomask & (1 << PFO_Z)) {
lmod_cast_u(po, po->operand[0].lmod), dst_opr_text);
*pfomask &= ~(1 << PFO_Z);
}
+}
+
+static void output_std_flag_s(FILE *fout, struct parsed_op *po,
+ int *pfomask, const char *dst_opr_text)
+{
if (*pfomask & (1 << PFO_S)) {
fprintf(fout, "\n cond_s = (%s%s < 0);",
lmod_cast_s(po, po->operand[0].lmod), dst_opr_text);
}
}
+static void output_std_flags(FILE *fout, struct parsed_op *po,
+ int *pfomask, const char *dst_opr_text)
+{
+ output_std_flag_z(fout, po, pfomask, dst_opr_text);
+ output_std_flag_s(fout, po, pfomask, dst_opr_text);
+}
+
enum {
OPP_FORCE_NORETURN = (1 << 0),
OPP_SIMPLE_ARGS = (1 << 1),
if (!pp->is_fptr)
fprintf(fout, " a%d", i + 1);
}
+
+ if (pp->arg[i].type.is_64bit)
+ i++;
}
if (pp->is_vararg) {
if (i > 0)
int had_decl = 0;
int label_pending = 0;
int need_double = 0;
+ int stack_align = 0;
+ int stack_fsz_adj = 0;
int regmask_save = 0; // used regs saved/restored in this func
int regmask_arg; // regs from this function args (fastcall, etc)
int regmask_ret; // regs needed on ret
// pass2:
// - handle ebp/esp frame, remove ops related to it
- scan_prologue_epilogue(opcnt);
+ scan_prologue_epilogue(opcnt, &stack_align);
+
+ // handle a case where sf size is unalignment, but is
+ // placed in a way that elements are still aligned
+ if (g_stack_fsz & 4) {
+ for (i = 0; i < g_eqcnt; i++) {
+ if (g_eqs[i].lmod != OPLM_QWORD)
+ continue;
+ if (!(g_eqs[i].offset & 4)) {
+ g_stack_fsz += 4;
+ stack_fsz_adj = 4;
+ }
+ break;
+ }
+ }
// pass3:
// - remove dead labels
// declare stack frame, va_arg
if (g_stack_fsz) {
+ if (stack_fsz_adj)
+ fprintf(fout, " // stack_fsz_adj %d\n", stack_fsz_adj);
+
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);
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);
+
+ if (stack_align > 8)
+ ferr(ops, "unhandled stack align of %d\n", stack_align);
+ else if (stack_align == 8)
+ fprintf(fout, " u64 align;");
fprintf(fout, " } sf;\n");
had_decl = 1;
}
break;
case OP_BSF:
+ case OP_BSR:
+ // on SKL, if src is 0, dst is left unchanged
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 = %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);
+ output_std_flag_z(fout, po, &pfomask, buf2);
+ if (po->op == OP_BSF)
+ snprintf(buf3, sizeof(buf3), "__builtin_ffs(%s) - 1", buf2);
+ else
+ snprintf(buf3, sizeof(buf3), "31 - __builtin_clz(%s)", buf2);
+ fprintf(fout, " if (%s) %s = %s;", buf2, buf1, buf3);
last_arith_dst = &po->operand[0];
delayed_flag_op = NULL;
- strcat(g_comment, " bsf");
+ strcat(g_comment, po->op == OP_BSF ? " bsf" : " bsr");
break;
case OP_DEC:
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");
+ int check_compat =
+ g_func_pp->is_stdcall && g_func_pp->argc_stack > 0;
+ check_compat |= pp->argc_stack > 0;
+ if (check_compat
+ && (pp->argc_stack != g_func_pp->argc_stack
+ || pp->is_stdcall != g_func_pp->is_stdcall))
+ ferr(po, "incompatible arg-reuse tailcall\n");
if (g_func_pp->has_retreg)
ferr(po, "TODO: retreg+tailcall\n");
z_check = ((long)po->datap >> 16) & 1;
out_src_opr_float(buf1, sizeof(buf1), po, &po->operand[0],
need_float_stack);
- if (mask == 0x0100) { // C0 -> <
+ if (mask == 0x0100 || mask == 0x0500) { // C0 -> <
fprintf(fout, " f_sw = %s < %s ? 0x0100 : 0;",
float_st0, buf1);
}
last_arith_dst = NULL;
}
- label_pending = 0;
+ if (!no_output)
+ label_pending = 0;
}
if (g_stack_fsz && !g_stack_frame_used)
// pass2:
// - handle ebp/esp frame, remove ops related to it
- scan_prologue_epilogue(opcnt);
+ scan_prologue_epilogue(opcnt, NULL);
// pass3:
// - remove dead labels
notaz.gp2x.de / ia32rtools.git / blobdiff