DrSocalkwe3n/luajit
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

x86/x64: Drop internal x87 math functions. Use libm functions.

Browse Source
This commit is contained in:
Mike Pall
2014-12-08 02:02:34 +01:00
parent e03df1e339
commit ad03eba715
7 changed files with 115 additions and 456 deletions
Download Patch File Download Diff File Expand all files Collapse all files

82
src/lj_asm_x86.h
View File

@@ -1593,26 +1593,9 @@ static void asm_x87load(ASMState *as, IRRef ref)
}
}
static void asm_fppow(ASMState *as, IRIns *ir, IRRef lref, IRRef rref)
{
/* The modified regs must match with the *.dasc implementation. */
RegSet drop = RSET_RANGE(RID_XMM0, RID_XMM2+1)|RID2RSET(RID_EAX);
IRIns *irx;
if (ra_hasreg(ir->r))
rset_clear(drop, ir->r); /* Dest reg handled below. */
ra_evictset(as, drop);
ra_destreg(as, ir, RID_XMM0);
emit_call(as, lj_vm_pow_sse);
irx = IR(lref);
if (ra_noreg(irx->r) && ra_gethint(irx->r) == RID_XMM1)
irx->r = RID_INIT; /* Avoid allocating xmm1 for x. */
ra_left(as, RID_XMM0, lref);
ra_left(as, RID_XMM1, rref);
}
static void asm_fpmath(ASMState *as, IRIns *ir)
{
IRFPMathOp fpm = ir->o == IR_FPMATH ? (IRFPMathOp)ir->op2 : IRFPM_OTHER;
IRFPMathOp fpm = (IRFPMathOp)ir->op2;
if (fpm == IRFPM_SQRT) {
Reg dest = ra_dest(as, ir, RSET_FPR);
Reg left = asm_fuseload(as, ir->op1, RSET_FPR);
@@ -1645,53 +1628,28 @@ static void asm_fpmath(ASMState *as, IRIns *ir)
}
} else if (fpm == IRFPM_EXP2 && asm_fpjoin_pow(as, ir)) {
/* Rejoined to pow(). */
} else { /* Handle x87 ops. */
int32_t ofs = sps_scale(ir->s); /* Use spill slot or temp slots. */
Reg dest = ir->r;
if (ra_hasreg(dest)) {
ra_free(as, dest);
ra_modified(as, dest);
emit_rmro(as, XO_MOVSD, dest, RID_ESP, ofs);
}
emit_rmro(as, XO_FSTPq, XOg_FSTPq, RID_ESP, ofs);
switch (fpm) { /* st0 = lj_vm_*(st0) */
case IRFPM_EXP: emit_call(as, lj_vm_exp_x87); break;
case IRFPM_EXP2: emit_call(as, lj_vm_exp2_x87); break;
case IRFPM_SIN: emit_x87op(as, XI_FSIN); break;
case IRFPM_COS: emit_x87op(as, XI_FCOS); break;
case IRFPM_TAN: emit_x87op(as, XI_FPOP); emit_x87op(as, XI_FPTAN); break;
case IRFPM_LOG: case IRFPM_LOG2: case IRFPM_LOG10:
/* Note: the use of fyl2xp1 would be pointless here. When computing
** log(1.0+eps) the precision is already lost after 1.0 is added.
** Subtracting 1.0 won't recover it. OTOH math.log1p would make sense.
*/
emit_x87op(as, XI_FYL2X); break;
case IRFPM_OTHER:
switch (ir->o) {
case IR_ATAN2:
emit_x87op(as, XI_FPATAN); asm_x87load(as, ir->op2); break;
case IR_LDEXP:
emit_x87op(as, XI_FPOP1); emit_x87op(as, XI_FSCALE); break;
default: lua_assert(0); break;
}
break;
default: lua_assert(0); break;
}
asm_x87load(as, ir->op1);
switch (fpm) {
case IRFPM_LOG: emit_x87op(as, XI_FLDLN2); break;
case IRFPM_LOG2: emit_x87op(as, XI_FLD1); break;
case IRFPM_LOG10: emit_x87op(as, XI_FLDLG2); break;
case IRFPM_OTHER:
if (ir->o == IR_LDEXP) asm_x87load(as, ir->op2);
break;
default: break;
}
} else {
asm_callid(as, ir, IRCALL_lj_vm_floor + fpm);
}
}
#define asm_atan2(as, ir) asm_fpmath(as, ir)
#define asm_ldexp(as, ir) asm_fpmath(as, ir)
#define asm_atan2(as, ir) asm_callid(as, ir, IRCALL_atan2)
static void asm_ldexp(ASMState *as, IRIns *ir)
{
int32_t ofs = sps_scale(ir->s); /* Use spill slot or temp slots. */
Reg dest = ir->r;
if (ra_hasreg(dest)) {
ra_free(as, dest);
ra_modified(as, dest);
emit_rmro(as, XO_MOVSD, dest, RID_ESP, ofs);
}
emit_rmro(as, XO_FSTPq, XOg_FSTPq, RID_ESP, ofs);
emit_x87op(as, XI_FPOP1);
emit_x87op(as, XI_FSCALE);
asm_x87load(as, ir->op1);
asm_x87load(as, ir->op2);
}
static void asm_fppowi(ASMState *as, IRIns *ir)
{