ShuriZma
/
dolphin
mirror of https://github.com/dolphin-emu/dolphin.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

[ARM] Implement ps_msub. Minor optimizations.

This commit is contained in:
Ryan Houdek 2013-09-16 11:21:04 +00:00
parent afdac224cb
commit 43f2313ef9
3 changed files with 43 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
Source/Core/Core/Src/PowerPC/JitArm32/Jit.h
View File

@ -220,6 +220,7 @@ public:
void ps_sum0(UGeckoInstruction _inst); void ps_sum0(UGeckoInstruction _inst);
void ps_sum1(UGeckoInstruction _inst); void ps_sum1(UGeckoInstruction _inst);
void ps_madd(UGeckoInstruction _inst); void ps_madd(UGeckoInstruction _inst);
void ps_msub(UGeckoInstruction _inst);
void ps_madds0(UGeckoInstruction _inst); void ps_madds0(UGeckoInstruction _inst);
void ps_madds1(UGeckoInstruction _inst); void ps_madds1(UGeckoInstruction _inst);
void ps_sub(UGeckoInstruction _inst); void ps_sub(UGeckoInstruction _inst);

62
Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Paired.cpp
View File

@ -69,14 +69,41 @@ void JitArm::ps_madd(UGeckoInstruction inst)
ARMReg V0 = fpr.GetReg(); ARMReg V0 = fpr.GetReg();
ARMReg V1 = fpr.GetReg(); ARMReg V1 = fpr.GetReg();
VMOV(V0, vB0); VMUL(V0, vA0, vC0);
VMOV(V1, vB1); VMUL(V1, vA1, vC1);
VADD(vD0, V0, vB0);
VADD(vD1, V1, vB1);
VMLA(V0, vA0, vC0); fpr.Unlock(V0);
VMLA(V1, vA1, vC1); fpr.Unlock(V1);
}
VMOV(vD0, V0); void JitArm::ps_msub(UGeckoInstruction inst)
VMOV(vD1, V1); {
INSTRUCTION_START
JITDISABLE(bJITPairedOff)
u32 a = inst.FA, b = inst.FB, c = inst.FC, d = inst.FD;
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(V0);
fpr.Unlock(V1); fpr.Unlock(V1);
@ -103,14 +130,11 @@ void JitArm::ps_madds0(UGeckoInstruction inst)
ARMReg V0 = fpr.GetReg(); ARMReg V0 = fpr.GetReg();
ARMReg V1 = fpr.GetReg(); ARMReg V1 = fpr.GetReg();
VMOV(V0, vB0); VMUL(V0, vA0, vC0);
VMOV(V1, vB1); VMUL(V1, vA1, vC0);
VMLA(V0, vA0, vC0); VADD(vD0, V0, vB0);
VMLA(V1, vA1, vC0); VADD(vD1, V1, vB1);
VMOV(vD0, V0);
VMOV(vD1, V1);
fpr.Unlock(V0); fpr.Unlock(V0);
fpr.Unlock(V1); fpr.Unlock(V1);
@ -137,14 +161,10 @@ void JitArm::ps_madds1(UGeckoInstruction inst)
ARMReg V0 = fpr.GetReg(); ARMReg V0 = fpr.GetReg();
ARMReg V1 = fpr.GetReg(); ARMReg V1 = fpr.GetReg();
VMOV(V0, vB0); VMUL(V0, vA0, vC1);
VMOV(V1, vB1); VMUL(V1, vA1, vC1);
VADD(vD0, V0, vB0);
VMLA(V0, vA0, vC1); VADD(vD1, V1, vB1);
VMLA(V1, vA1, vC1);
VMOV(vD0, V0);
VMOV(vD1, V1);
fpr.Unlock(V0); fpr.Unlock(V0);
fpr.Unlock(V1); fpr.Unlock(V1);

2
Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Tables.cpp
View File

@ -156,7 +156,7 @@ static GekkoOPTemplate table4_2[] =
{24, &JitArm::Default}, //"ps_res", OPTYPE_PS, 0}}, {24, &JitArm::Default}, //"ps_res", OPTYPE_PS, 0}},
{25, &JitArm::ps_mul}, //"ps_mul", OPTYPE_PS, 0}}, {25, &JitArm::ps_mul}, //"ps_mul", OPTYPE_PS, 0}},
{26, &JitArm::Default}, //"ps_rsqrte", OPTYPE_PS, 0, 1}}, {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}}, {29, &JitArm::ps_madd}, //"ps_madd", OPTYPE_PS, 0}},
{30, &JitArm::Default}, //"ps_nmsub", OPTYPE_PS, 0}}, {30, &JitArm::Default}, //"ps_nmsub", OPTYPE_PS, 0}},
{31, &JitArm::Default}, //"ps_nmadd", OPTYPE_PS, 0}}, {31, &JitArm::Default}, //"ps_nmadd", OPTYPE_PS, 0}},