Skip to content

Commit

Permalink
Browse files Browse the repository at this point in the history
[ARM] Implement ps_msub. Minor optimizations.
  • Loading branch information
Sonicadvance1 committed Sep 16, 2013
1 parent afdac22 commit 43f2313
Show file tree
Hide file tree
Showing 3 changed files with 43 additions and 22 deletions.
1 change: 1 addition & 0 deletions Source/Core/Core/Src/PowerPC/JitArm32/Jit.h
Expand Up @@ -220,6 +220,7 @@ class JitArm : public JitBase, public ArmGen::ARMXCodeBlock
void ps_sum0(UGeckoInstruction _inst);
void ps_sum1(UGeckoInstruction _inst);
void ps_madd(UGeckoInstruction _inst);
void ps_msub(UGeckoInstruction _inst);
void ps_madds0(UGeckoInstruction _inst);
void ps_madds1(UGeckoInstruction _inst);
void ps_sub(UGeckoInstruction _inst);
Expand Down
62 changes: 41 additions & 21 deletions Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Paired.cpp
Expand Up @@ -69,15 +69,42 @@ void JitArm::ps_madd(UGeckoInstruction inst)
ARMReg V0 = fpr.GetReg();
ARMReg V1 = fpr.GetReg();

VMOV(V0, vB0);
VMOV(V1, vB1);
VMUL(V0, vA0, vC0);
VMUL(V1, vA1, vC1);
VADD(vD0, V0, vB0);
VADD(vD1, V1, vB1);

fpr.Unlock(V0);
fpr.Unlock(V1);
}

void JitArm::ps_msub(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITPairedOff)

VMLA(V0, vA0, vC0);
VMLA(V1, vA1, vC1);
u32 a = inst.FA, b = inst.FB, c = inst.FC, d = inst.FD;

VMOV(vD0, V0);
VMOV(vD1, V1);
if (inst.Rc) {
Default(inst); return;
}
ARMReg vA0 = fpr.R0(a);
ARMReg vA1 = fpr.R1(a);
ARMReg vB0 = fpr.R0(b);
ARMReg vB1 = fpr.R1(b);
ARMReg vC0 = fpr.R0(c);
ARMReg vC1 = fpr.R1(c);
ARMReg vD0 = fpr.R0(d, false);
ARMReg vD1 = fpr.R1(d, false);

ARMReg V0 = fpr.GetReg();
ARMReg V1 = fpr.GetReg();

VMUL(V0, vA0, vC0);
VMUL(V1, vA1, vC1);
VSUB(vD0, V0, vB0);
VSUB(vD1, V1, vB1);

fpr.Unlock(V0);
fpr.Unlock(V1);
}
Expand All @@ -102,15 +129,12 @@ void JitArm::ps_madds0(UGeckoInstruction inst)

ARMReg V0 = fpr.GetReg();
ARMReg V1 = fpr.GetReg();

VMOV(V0, vB0);
VMOV(V1, vB1);

VMLA(V0, vA0, vC0);
VMLA(V1, vA1, vC0);
VMUL(V0, vA0, vC0);
VMUL(V1, vA1, vC0);

VMOV(vD0, V0);
VMOV(vD1, V1);
VADD(vD0, V0, vB0);
VADD(vD1, V1, vB1);

fpr.Unlock(V0);
fpr.Unlock(V1);
Expand All @@ -137,14 +161,10 @@ void JitArm::ps_madds1(UGeckoInstruction inst)
ARMReg V0 = fpr.GetReg();
ARMReg V1 = fpr.GetReg();

VMOV(V0, vB0);
VMOV(V1, vB1);

VMLA(V0, vA0, vC1);
VMLA(V1, vA1, vC1);

VMOV(vD0, V0);
VMOV(vD1, V1);
VMUL(V0, vA0, vC1);
VMUL(V1, vA1, vC1);
VADD(vD0, V0, vB0);
VADD(vD1, V1, vB1);

fpr.Unlock(V0);
fpr.Unlock(V1);
Expand Down
2 changes: 1 addition & 1 deletion Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Tables.cpp
Expand Up @@ -156,7 +156,7 @@ static GekkoOPTemplate table4_2[] =
{24, &JitArm::Default}, //"ps_res", OPTYPE_PS, 0}},
{25, &JitArm::ps_mul}, //"ps_mul", OPTYPE_PS, 0}},
{26, &JitArm::Default}, //"ps_rsqrte", OPTYPE_PS, 0, 1}},
{28, &JitArm::Default}, //"ps_msub", OPTYPE_PS, 0}},
{28, &JitArm::ps_msub}, //"ps_msub", OPTYPE_PS, 0}},
{29, &JitArm::ps_madd}, //"ps_madd", OPTYPE_PS, 0}},
{30, &JitArm::Default}, //"ps_nmsub", OPTYPE_PS, 0}},
{31, &JitArm::Default}, //"ps_nmadd", OPTYPE_PS, 0}},
Expand Down

0 comments on commit 43f2313

Please sign in to comment.