X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=tools%2Ftranslate.c;h=1ae3cf76eb1bc376ee454fb6ed46d936c917565e;hb=622eb2eff8349fa0a6675ce3213ab8f8fb1c1798;hp=a7f629ee4dccea9daee3ec74244d93a334112aba;hpb=497a6d6b4c2992fc9cbd2591985d108bc8859f72;p=ia32rtools.git diff --git a/tools/translate.c b/tools/translate.c index a7f629e..1ae3cf7 100644 --- a/tools/translate.c +++ b/tools/translate.c @@ -79,6 +79,7 @@ enum op_op { OP_NOT, OP_XLAT, OP_CDQ, + OP_BSWAP, OP_LODS, OP_STOS, OP_MOVS, @@ -136,14 +137,19 @@ enum op_op { OP_FISUB, OP_FIDIVR, OP_FISUBR, + OP_FCHS, OP_FCOS, OP_FPATAN, OP_FPTAN, OP_FSIN, OP_FSQRT, + OP_FXCH, + OP_FYL2X, // mmx OP_EMMS, // pseudo-ops for lib calls + OPP_ALLSHL, + OPP_ALLSHR, OPP_FTOL, // undefined OP_UD2, @@ -337,6 +343,9 @@ enum x86_regs { #define mxDX (1 << xDX) #define mxST0 (1 << xST0) #define mxST1 (1 << xST1) +#define mxST1_0 (mxST1 | mxST0) +#define mxST7_2 (0xfc << xST0) +#define mxSTa (0xff << xST0) // possible basic comparison types (without inversion) enum parsed_flag_op { @@ -918,6 +927,7 @@ static const struct { { "not", OP_NOT, 1, 1, OPF_DATA }, { "xlat", OP_XLAT, 0, 0, OPF_DATA }, { "cdq", OP_CDQ, 0, 0, OPF_DATA }, + { "bswap",OP_BSWAP, 1, 1, OPF_DATA }, { "lodsb",OP_LODS, 0, 0, OPF_DATA }, { "lodsw",OP_LODS, 0, 0, OPF_DATA }, { "lodsd",OP_LODS, 0, 0, OPF_DATA }, @@ -1029,6 +1039,7 @@ static const struct { { "fld", OP_FLD, 1, 1, OPF_FPUSH }, { "fild", OP_FILD, 1, 1, OPF_FPUSH }, { "fld1", OP_FLDc, 0, 0, OPF_FPUSH }, + { "fldln2", 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 }, @@ -1050,15 +1061,20 @@ static const struct { { "fisub", OP_FISUB, 1, 1, 0 }, { "fidivr", OP_FIDIVR, 1, 1, 0 }, { "fisubr", OP_FISUBR, 1, 1, 0 }, + { "fchs", OP_FCHS, 0, 0, 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 }, + { "fxch", OP_FXCH, 1, 1, 0 }, + { "fyl2x", OP_FYL2X, 0, 0, OPF_FPOP }, // mmx { "emms", OP_EMMS, 0, 0, OPF_DATA }, { "movq", OP_MOV, 2, 2, OPF_DATA }, // pseudo-ops for lib calls + { "_allshl",OPP_ALLSHL }, + { "_allshr",OPP_ALLSHR }, { "_ftol", OPP_FTOL }, // must be last { "ud2", OP_UD2 }, @@ -1157,6 +1173,7 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc) case OP_INC: case OP_DEC: case OP_NEG: + case OP_BSWAP: // more below.. op->regmask_src |= op->regmask_dst; break; @@ -1345,6 +1362,8 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc) 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, "ln2")) + op->operand[0].val = X87_CONST_LN2; else if (IS(words[op_w] + 3, "z")) op->operand[0].val = X87_CONST_Z; else @@ -1382,14 +1401,17 @@ static void parse_op(struct parsed_op *op, char words[16][256], int wordc) case OP_FISUB: case OP_FIDIVR: case OP_FISUBR: + case OP_FCHS: case OP_FCOS: case OP_FSIN: case OP_FSQRT: + case OP_FXCH: op->regmask_src |= mxST0; op->regmask_dst |= mxST0; break; case OP_FPATAN: + case OP_FYL2X: op->regmask_src |= mxST0 | mxST1; op->regmask_dst |= mxST0; break; @@ -2190,7 +2212,7 @@ static char *out_src_opr_u32(char *buf, size_t buf_size, } static char *out_src_opr_float(char *buf, size_t buf_size, - struct parsed_op *po, struct parsed_opr *popr) + struct parsed_op *po, struct parsed_opr *popr, int need_float_stack) { const char *cast = NULL; char tmp[256]; @@ -2200,7 +2222,15 @@ static char *out_src_opr_float(char *buf, size_t buf_size, 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); + if (need_float_stack) { + if (popr->reg == xST0) + snprintf(buf, buf_size, "f_st[f_stp & 7]"); + else + snprintf(buf, buf_size, "f_st[(f_stp + %d) & 7]", + popr->reg - xST0); + } + else + snprintf(buf, buf_size, "f_st%d", popr->reg - xST0); break; case OPT_REGMEM: @@ -2229,10 +2259,10 @@ static char *out_src_opr_float(char *buf, size_t buf_size, } static char *out_dst_opr_float(char *buf, size_t buf_size, - struct parsed_op *po, struct parsed_opr *popr) + struct parsed_op *po, struct parsed_opr *popr, int need_float_stack) { // same? - return out_src_opr_float(buf, buf_size, po, popr); + return out_src_opr_float(buf, buf_size, po, popr, need_float_stack); } static void out_test_for_cc(char *buf, size_t buf_size, @@ -3519,7 +3549,9 @@ static void resolve_branches_parse_calls(int opcnt) unsigned int regmask_src; unsigned int regmask_dst; } pseudo_ops[] = { - { "__ftol", OPP_FTOL, OPF_FPOP, mxST0, mxAX | mxDX }, + { "__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 }, }; const struct parsed_proto *pp_c; struct parsed_proto *pp; @@ -3803,11 +3835,11 @@ static void scan_prologue_epilogue(int opcnt) { g_sp_frame = 1; - 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) @@ -3844,7 +3876,16 @@ static void scan_prologue_epilogue(int opcnt) || !IS(opr_name(&ops[j], 0), "esp") || ops[j].operand[1].type != OPT_CONST || ops[j].operand[1].val != g_stack_fsz) + { + if (ops[i].op == OP_CALL && ops[i].pp != NULL + && ops[i].pp->is_noreturn) + { + // noreturn tailcall with no epilogue + i++; + 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 @@ -4318,9 +4359,11 @@ static int collect_call_args_early(struct parsed_op *po, int i, 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 (!g_header_mode) { + 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; @@ -4663,7 +4706,6 @@ static void reg_use_pass(int i, int opcnt, unsigned char *cbits, int *regmask_init, int regmask_arg) { struct parsed_op *po; - unsigned int mask; int already_saved; int regmask_new; int regmask_op; @@ -4763,20 +4805,15 @@ static void reg_use_pass(int i, int opcnt, unsigned char *cbits, po->regmask_dst &= ~(1 << xAX); } } + + // not "full stack" mode and have something in stack + if (!(regmask_now & mxST7_2) && (regmask_now & mxST1_0)) + ferr(po, "float stack is not empty on func call\n"); } 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 @@ -4809,21 +4846,34 @@ static void reg_use_pass(int i, int opcnt, unsigned char *cbits, } } + if (po->flags & OPF_FPUSH) { + if (regmask_now & mxST1) + regmask_now |= mxSTa; // switch to "full stack" mode + if (regmask_now & mxSTa) + po->flags |= OPF_FSHIFT; + if (!(regmask_now & mxST7_2)) { + regmask_now = + (regmask_now & ~mxST1_0) | ((regmask_now & mxST0) << 1); + } + } + regmask_now |= regmask_op; *regmask |= regmask_now; // released regs if (po->flags & OPF_FPOP) { - mask = mxST0 | mxST1; - if (!(regmask_now & mask)) + if ((regmask_now & mxSTa) == 0) ferr(po, "float pop on empty stack?\n"); - if (regmask_now & mxST1) + if (regmask_now & (mxST7_2 | mxST1)) po->flags |= OPF_FSHIFT; - regmask_now = (regmask_now & ~mask) | ((regmask_now & mxST1) >> 1); + if (!(regmask_now & mxST7_2)) { + regmask_now = + (regmask_now & ~mxST1_0) | ((regmask_now & mxST1) >> 1); + } } if (po->flags & OPF_TAIL) { - if (regmask_now & (mxST0 | mxST1)) + if (!(regmask_now & mxST7_2) && (regmask_now & mxST1_0)) ferr(po, "float regs on tail: %x\n", regmask_now); // there is support for "conditional tailcall", sort of @@ -4954,9 +5004,12 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) int save_arg_vars[MAX_ARG_GRP] = { 0, }; unsigned char cbits[MAX_OPS / 8]; const char *float_type; - int cond_vars = 0; + const char *float_st0; + const char *float_st1; + int need_float_stack = 0; int need_tmp_var = 0; int need_tmp64 = 0; + int cond_vars = 0; int had_decl = 0; int label_pending = 0; int need_double = 0; @@ -4969,6 +5022,7 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) int regmask = 0; // used regs int pfomask = 0; int found = 0; + int dead_dst; int no_output; int i, j, l; int arg; @@ -5176,16 +5230,23 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) cond_vars |= 1 << PFO_C; } - if (po->op == OP_CMPS || po->op == OP_SCAS) { + switch (po->op) { + case OP_CMPS: + case OP_SCAS: cond_vars |= 1 << PFO_Z; - } - else if (po->op == OP_MUL - || (po->op == OP_IMUL && po->operand_cnt == 1)) - { + break; + + case OP_MUL: if (po->operand[0].lmod == OPLM_DWORD) need_tmp64 = 1; - } - else if (po->op == OP_CALL) { + break; + + case OP_IMUL: + if (po->operand_cnt == 1 && po->operand[0].lmod == OPLM_DWORD) + need_tmp64 = 1; + break; + + case OP_CALL: // note: resolved non-reg calls are OPF_DONE already pp = po->pp; ferr_assert(po, pp != NULL); @@ -5232,27 +5293,31 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) 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) - { - // = offset - if ((po->operand[1].pp->is_func || po->operand[1].pp->is_fptr) - && !IS_START(po->operand[1].name, "off_")) + break; + + case OP_MOV: + if (po->operand[0].pp != NULL && po->operand[1].pp != NULL) { - 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"); + // = offset + 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) { + break; + + case OP_DIV: + case OP_IDIV: if (po->operand[0].lmod == OPLM_DWORD) { // 32bit division is common, look for it if (po->op == OP_DIV) @@ -5266,24 +5331,46 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) } else need_tmp_var = 1; - } - else if (po->op == OP_CLD) + break; + + case OP_CLD: po->flags |= OPF_RMD | OPF_DONE; - else if (po->op == OPP_FTOL) { + break; + + case OP_RCL: + case OP_RCR: + case OP_XCHG: + need_tmp_var = 1; + break; + + case OP_FLD: + if (po->operand[0].lmod == OPLM_QWORD) + need_double = 1; + break; + + case OPP_ALLSHL: + case OPP_ALLSHR: + need_tmp64 = 1; + break; + + case 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; + break; } - 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; + default: + break; + } } float_type = need_double ? "double" : "float"; + need_float_stack = !!(regmask & mxST7_2); + float_st0 = need_float_stack ? "f_st[f_stp & 7]" : "f_st0"; + float_st1 = need_float_stack ? "f_st[(f_stp + 1) & 7]" : "f_st1"; // output starts here @@ -5441,14 +5528,21 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) } } // ... 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 (need_float_stack) { + fprintf(fout, " %s f_st[8];\n", float_type); + fprintf(fout, " int f_stp = 0;\n"); + had_decl = 1; + } + else { + 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; + } } } } @@ -5719,6 +5813,12 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) strcpy(g_comment, "cdq"); break; + case OP_BSWAP: + assert_operand_cnt(1); + out_dst_opr(buf1, sizeof(buf1), po, &po->operand[0]); + fprintf(fout, " %s = __builtin_bswap32(%s);", buf1, buf1); + break; + case OP_LODS: if (po->flags & OPF_REP) { assert_operand_cnt(3); @@ -6350,7 +6450,11 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) fprintf(fout, "(u32)"); } else if (po->regmask_dst & mxST0) { - fprintf(fout, "f_st0 = "); + ferr_assert(po, po->flags & OPF_FPUSH); + if (need_float_stack) + fprintf(fout, "f_st[--f_stp & 7] = "); + else + fprintf(fout, "f_st0 = "); } } @@ -6569,59 +6673,110 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) no_output = 1; break; + // pseudo ops + case OPP_ALLSHL: + case OPP_ALLSHR: + fprintf(fout, " tmp64 = ((u64)edx << 32) | eax;\n"); + fprintf(fout, " tmp64 = (s64)tmp64 %s= LOBYTE(ecx);\n", + po->op == OPP_ALLSHL ? "<<" : ">>"); + fprintf(fout, " edx = tmp64 >> 32; eax = tmp64;"); + strcat(g_comment, po->op == OPP_ALLSHL + ? " allshl" : " allshr"); + 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 (need_float_stack) { + out_src_opr_float(buf1, sizeof(buf1), + po, &po->operand[0], 1); + if (po->regmask_src & mxSTa) { + fprintf(fout, " f_st[(f_stp - 1) & 7] = %s; f_stp--;", + buf1); + } + else + fprintf(fout, " f_st[--f_stp & 7] = %s;", buf1); + } + else { + if (po->flags & OPF_FSHIFT) + fprintf(fout, " f_st1 = f_st0;"); + if (po->operand[0].type == OPT_REG + && po->operand[0].reg == xST0) + { + strcat(g_comment, " fld st"); + break; + } + fprintf(fout, " f_st0 = %s;", + out_src_opr_float(buf1, sizeof(buf1), + po, &po->operand[0], 0)); } - 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)); + out_src_opr(buf1, sizeof(buf1), po, &po->operand[0], + lmod_cast(po, po->operand[0].lmod, 1), 0); + snprintf(buf2, sizeof(buf2), "(%s)%s", float_type, buf1); + if (need_float_stack) { + fprintf(fout, " f_st[--f_stp & 7] = %s;", buf2); + } + else { + if (po->flags & OPF_FSHIFT) + fprintf(fout, " f_st1 = f_st0;"); + fprintf(fout, " f_st0 = %s;", buf2); + } strcat(g_comment, " fild"); break; case OP_FLDc: - if (po->flags & OPF_FSHIFT) - fprintf(fout, " f_st1 = f_st0;\n"); - fprintf(fout, " f_st0 = "); + if (need_float_stack) + fprintf(fout, " f_st[--f_stp & 7] = "); + else { + if (po->flags & OPF_FSHIFT) + fprintf(fout, " f_st1 = f_st0;"); + 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; + case X87_CONST_1: fprintf(fout, "1.0;"); break; + case X87_CONST_LN2: fprintf(fout, "0.693147180559945;"); 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; + out_dst_opr_float(buf1, sizeof(buf1), po, &po->operand[0], + need_float_stack); + dead_dst = po->operand[0].type == OPT_REG + && po->operand[0].reg == xST0; + if (need_float_stack) { + if (!dead_dst) + fprintf(fout, " %s = f_st[f_stp & 7];", buf1); + if (po->flags & OPF_FSHIFT) + fprintf(fout, " f_stp++;"); } - 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"); + else { + if (!dead_dst) + fprintf(fout, " %s = f_st0;", buf1); + if (po->flags & OPF_FSHIFT) + fprintf(fout, " f_st0 = f_st1;"); + } + if (dead_dst && !(po->flags & OPF_FSHIFT)) + no_output = 1; + else + strcat(g_comment, " fst"); break; case OP_FADD: case OP_FDIV: case OP_FMUL: case OP_FSUB: + out_dst_opr_float(buf1, sizeof(buf1), po, &po->operand[0], + need_float_stack); + out_src_opr_float(buf2, sizeof(buf2), po, &po->operand[1], + need_float_stack); + dead_dst = (po->flags & OPF_FPOP) + && po->operand[0].type == OPT_REG + && po->operand[0].reg == xST0; switch (po->op) { case OP_FADD: j = '+'; break; case OP_FDIV: j = '/'; break; @@ -6629,27 +6784,55 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) case OP_FSUB: j = '-'; break; default: j = 'x'; break; } - if (po->flags & OPF_FSHIFT) { - fprintf(fout, " f_st0 = f_st1 %c f_st0;", j); + if (need_float_stack) { + if (!dead_dst) + fprintf(fout, " %s %c= %s;", buf1, j, buf2); + if (po->flags & OPF_FSHIFT) + fprintf(fout, " f_stp++;"); } 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])); + if (po->flags & OPF_FSHIFT) { + // note: assumes only 2 regs handled + if (!dead_dst) + fprintf(fout, " f_st0 = f_st1 %c f_st0;", j); + else + fprintf(fout, " f_st0 = f_st1;"); + } + else if (!dead_dst) + fprintf(fout, " %s %c= %s;", buf1, j, buf2); } + no_output = (dead_dst && !(po->flags & OPF_FSHIFT)); 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])); + out_dst_opr_float(buf1, sizeof(buf1), po, &po->operand[0], + need_float_stack); + out_src_opr_float(buf2, sizeof(buf2), po, &po->operand[1], + need_float_stack); + out_src_opr_float(buf3, sizeof(buf3), po, &po->operand[0], + need_float_stack); + dead_dst = (po->flags & OPF_FPOP) + && po->operand[0].type == OPT_REG + && po->operand[0].reg == xST0; + j = po->op == OP_FDIVR ? '/' : '-'; + if (need_float_stack) { + if (!dead_dst) + fprintf(fout, " %s = %s %c %s;", buf1, buf2, j, buf3); + if (po->flags & OPF_FSHIFT) + fprintf(fout, " f_stp++;"); + } + else { + if (po->flags & OPF_FSHIFT) { + if (!dead_dst) + fprintf(fout, " f_st0 = f_st0 %c f_st1;", j); + else + fprintf(fout, " f_st0 = f_st1;"); + } + else if (!dead_dst) + fprintf(fout, " %s = %s %c %s;", buf1, buf2, j, buf3); + } + no_output = (dead_dst && !(po->flags & OPF_FSHIFT)); break; case OP_FIADD: @@ -6663,43 +6846,86 @@ static void gen_func(FILE *fout, FILE *fhdr, const char *funcn, int opcnt) case OP_FISUB: j = '-'; break; default: j = 'x'; break; } - fprintf(fout, " f_st0 %c= (%s)%s;", j, float_type, + fprintf(fout, " %s %c= (%s)%s;", float_st0, + 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 ? '/' : '-'); + fprintf(fout, " %s = %s %c %s;", float_st0, + out_src_opr_float(buf2, sizeof(buf2), po, &po->operand[1], + need_float_stack), + po->op == OP_FIDIVR ? '/' : '-', float_st0); + break; + + case OP_FCHS: + fprintf(fout, " %s = -%s;", float_st0, float_st0); break; case OP_FCOS: - fprintf(fout, " f_st0 = cos%s(f_st0);", - need_double ? "" : "f"); + fprintf(fout, " %s = cos%s(%s);", float_st0, + need_double ? "" : "f", float_st0); break; case OP_FPATAN: - fprintf(fout, " f_st0 = atan%s(f_st1 / f_st0);", - need_double ? "" : "f"); + if (need_float_stack) { + fprintf(fout, " %s = atan%s(%s / %s);", float_st1, + need_double ? "" : "f", float_st1, float_st0); + fprintf(fout, " f_stp++;"); + } + else { + fprintf(fout, " f_st0 = atan%s(f_st1 / f_st0);", + need_double ? "" : "f"); + } + break; + + case OP_FYL2X: + if (need_float_stack) { + fprintf(fout, " %s = %s * log2%s(%s);", float_st1, + float_st1, need_double ? "" : "f", float_st0); + fprintf(fout, " f_stp++;"); + } + else { + fprintf(fout, " f_st0 = f_st1 * log2%s(f_st0);", + need_double ? "" : "f"); + } break; case OP_FSIN: - fprintf(fout, " f_st0 = sin%s(f_st0);", - need_double ? "" : "f"); + fprintf(fout, " %s = sin%s(%s);", float_st0, + need_double ? "" : "f", float_st0); break; case OP_FSQRT: - fprintf(fout, " f_st0 = sqrt%s(f_st0);", - need_double ? "" : "f"); + fprintf(fout, " %s = sqrt%s(%s);", float_st0, + need_double ? "" : "f", float_st0); + break; + + case OP_FXCH: + dead_dst = po->operand[0].type == OPT_REG + && po->operand[0].reg == xST0; + if (!dead_dst) { + out_src_opr_float(buf1, sizeof(buf1), po, &po->operand[0], + need_float_stack); + fprintf(fout, " { %s t = %s; %s = %s; %s = t; }", float_type, + float_st0, float_st0, buf1, buf1); + strcat(g_comment, " fxch"); + } + else + no_output = 1; break; 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;"); + fprintf(fout, " eax = (s32)%s;", float_st0); + if (po->flags & OPF_FSHIFT) { + if (need_float_stack) + fprintf(fout, " f_stp++;"); + else + fprintf(fout, " f_st0 = f_st1;"); + } strcat(g_comment, " ftol"); break; @@ -7216,7 +7442,7 @@ static void gen_hdr(const char *funcn, int opcnt) fp->argc_stack--; } - fp->regmask_dep = regmask_dep & ~(1 << xSP); + fp->regmask_dep = regmask_dep & ~((1 << xSP) | mxSTa); fp->has_ret = has_ret; #if 0 printf("// has_ret %d, regmask_dep %x\n",