gh-mirror/mikepaul-LuaJIT
1
0
Fork 0
mirror of https://github.com/LuaJIT/LuaJIT.git synced 2025-02-07 15:14:08 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

x64: Don't fuse implicitly 32-to-64 extended operands.

Browse Source
This commit is contained in:
Mike Pall 2012-11-13 19:20:52 +01:00
parent 87d74a8f3d
commit 1cd13f6b33
1 changed files with 18 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
src/lj_asm_x86.h
View File

@ -367,6 +367,18 @@ static Reg asm_fuseload(ASMState *as, IRRef ref, RegSet allow)
return ra_allocref(as, ref, allow);
}
#if LJ_64
/* Don't fuse a 32 bit load into a 64 bit operation. */
static Reg asm_fuseloadm(ASMState *as, IRRef ref, RegSet allow, int is64)
{
if (is64 && !irt_is64(IR(ref)->t))
return ra_alloc1(as, ref, allow);
return asm_fuseload(as, ref, allow);
}
#else
#define asm_fuseloadm(as, ref, allow, is64) asm_fuseload(as, (ref), (allow))
#endif
/* -- Calls --------------------------------------------------------------- */
/* Count the required number of stack slots for a call. */
@ -696,7 +708,7 @@ static void asm_conv(ASMState *as, IRIns *ir)
} else { /* Integer to FP conversion. */
Reg left = (LJ_64 && (st == IRT_U32 || st == IRT_U64)) ?
ra_alloc1(as, lref, RSET_GPR) :
asm_fuseload(as, lref, RSET_GPR);
asm_fuseloadm(as, lref, RSET_GPR, st64);
if (LJ_64 && st == IRT_U64) {
MCLabel l_end = emit_label(as);
const void *k = lj_ir_k64_find(as->J, U64x(43f00000,00000000));
@ -1829,7 +1841,7 @@ static void asm_intarith(ASMState *as, IRIns *ir, x86Arith xa)
if (asm_swapops(as, ir)) {
IRRef tmp = lref; lref = rref; rref = tmp;
}
right = asm_fuseload(as, rref, rset_clear(allow, dest));
right = asm_fuseloadm(as, rref, rset_clear(allow, dest), irt_is64(ir->t));
}
if (irt_isguard(ir->t)) /* For IR_ADDOV etc. */
asm_guardcc(as, CC_O);
@ -1842,7 +1854,7 @@ static void asm_intarith(ASMState *as, IRIns *ir, x86Arith xa)
emit_mrm(as, XO_IMUL, REX_64IR(ir, dest), right);
} else { /* IMUL r, r, k. */
/* NYI: use lea/shl/add/sub (FOLD only does 2^k) depending on CPU. */
Reg left = asm_fuseload(as, lref, RSET_GPR);
Reg left = asm_fuseloadm(as, lref, RSET_GPR, irt_is64(ir->t));
x86Op xo;
if (checki8(k)) { emit_i8(as, k); xo = XO_IMULi8;
} else { emit_i32(as, k); xo = XO_IMULi; }
@ -2109,7 +2121,7 @@ static void asm_comp(ASMState *as, IRIns *ir, uint32_t cc)
}
}
as->curins--; /* Skip to BAND to avoid failing in noconflict(). */
right = asm_fuseload(as, irl->op1, allow);
right = asm_fuseloadm(as, irl->op1, allow, r64);
as->curins++; /* Undo the above. */
test_nofuse:
asm_guardcc(as, cc);
@ -2146,7 +2158,7 @@ static void asm_comp(ASMState *as, IRIns *ir, uint32_t cc)
return;
} /* Otherwise handle register case as usual. */
} else {
left = asm_fuseload(as, lref, RSET_GPR);
left = asm_fuseloadm(as, lref, RSET_GPR, r64);
}
asm_guardcc(as, cc);
if (usetest && left != RID_MRM) {
@ -2160,7 +2172,7 @@ static void asm_comp(ASMState *as, IRIns *ir, uint32_t cc)
}
} else {
Reg left = ra_alloc1(as, lref, RSET_GPR);
Reg right = asm_fuseload(as, rref, rset_exclude(RSET_GPR, left));
Reg right = asm_fuseloadm(as, rref, rset_exclude(RSET_GPR, left), r64);
asm_guardcc(as, cc);
emit_mrm(as, XO_CMP, r64 + left, right);
}