ShuriZma
/
pcsx2
mirror of https://github.com/PCSX2/pcsx2.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

EERec: Clean up [LS]D[LR] a bit

This commit is contained in:
TellowKrinkle 2021-09-17 17:10:19 -05:00 committed by refractionpcsx2
parent b7f1c65f7e
commit 7563f54e83
1 changed files with 14 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

35
pcsx2/x86/ix86-32/iR5900LoadStore.cpp
View File

@ -452,11 +452,6 @@ void recSWR()
//////////////////////////////////////////////////// ////////////////////////////////////////////////////
alignas(16) const u32 SHIFT_MASKS[2][4] = {
{ 0xffffffff, 0xffffffff, 0x00000000, 0x00000000 },
{ 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff }
};
void recLDL() void recLDL()
{ {
if (!_Rt_) if (!_Rt_)
@ -514,23 +509,22 @@ void recLDL()
xSUB(edx, eax); xSUB(edx, eax);
xMOVDZX(xRegisterSSE(t1reg), eax); xMOVDZX(xRegisterSSE(t1reg), eax);
xMOVQZX(xRegisterSSE(t0reg), ptr128[&SHIFT_MASKS[0][0]]); xPCMP.EQD(xRegisterSSE(t0reg), xRegisterSSE(t0reg));
xPSRL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg)); xPSRL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg));
xPAND(xRegisterSSE(t0reg), xRegisterSSE(rtreg)); xPAND(xRegisterSSE(t0reg), xRegisterSSE(rtreg));
xMOVDQA(xRegisterSSE(t2reg), xRegisterSSE(t0reg)); xMOVDQA(xRegisterSSE(t2reg), xRegisterSSE(t0reg));
xMOVDZX(xRegisterSSE(t1reg), edx); xMOVDZX(xRegisterSSE(t1reg), edx);
xMOVQZX(xRegisterSSE(t0reg), ptr128[&dummyValue[0]]); xMOVQZX(xRegisterSSE(t0reg), ptr64[&dummyValue[0]]);
xPSLL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg)); xPSLL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg));
xPOR(xRegisterSSE(t0reg), xRegisterSSE(t2reg)); xPOR(xRegisterSSE(t0reg), xRegisterSSE(t2reg));
xMOVSD(xRegisterSSE(rtreg), xRegisterSSE(t0reg));
xForwardJump32 full; xForwardJump32 full;
skip.SetTarget(); skip.SetTarget();
xMOVQZX(xRegisterSSE(t0reg), ptr128[&dummyValue[0]]); xMOVL.PS(xRegisterSSE(rtreg), ptr128[&dummyValue[0]]);
full.SetTarget(); full.SetTarget();
xBLEND.PS(xRegisterSSE(rtreg), xRegisterSSE(t0reg), 0x3);
_freeXMMreg(t0reg); _freeXMMreg(t0reg);
_freeXMMreg(t1reg); _freeXMMreg(t1reg);
_freeXMMreg(t2reg); _freeXMMreg(t2reg);
@ -602,23 +596,22 @@ void recLDR()
xSUB(edx, eax); xSUB(edx, eax);
xMOVDZX(xRegisterSSE(t1reg), edx); //64-shift*8 xMOVDZX(xRegisterSSE(t1reg), edx); //64-shift*8
xMOVQZX(xRegisterSSE(t0reg), ptr128[&SHIFT_MASKS[0][0]]); xPCMP.EQD(xRegisterSSE(t0reg), xRegisterSSE(t0reg));
xPSLL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg)); xPSLL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg));
xPAND(xRegisterSSE(t0reg), xRegisterSSE(rtreg)); xPAND(xRegisterSSE(t0reg), xRegisterSSE(rtreg));
xMOVQZX(xRegisterSSE(t2reg), xRegisterSSE(t0reg)); xMOVQZX(xRegisterSSE(t2reg), xRegisterSSE(t0reg));
xMOVDZX(xRegisterSSE(t1reg), eax); //shift*8 xMOVDZX(xRegisterSSE(t1reg), eax); //shift*8
xMOVQZX(xRegisterSSE(t0reg), ptr128[&dummyValue[0]]); xMOVQZX(xRegisterSSE(t0reg), ptr64[&dummyValue[0]]);
xPSRL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg)); xPSRL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg));
xPOR(xRegisterSSE(t0reg), xRegisterSSE(t2reg)); xPOR(xRegisterSSE(t0reg), xRegisterSSE(t2reg));
xMOVSD(xRegisterSSE(rtreg), xRegisterSSE(t0reg));
xForwardJump32 full; xForwardJump32 full;
skip.SetTarget(); skip.SetTarget();
xMOVQZX(xRegisterSSE(t0reg), ptr128[&dummyValue[0]]); xMOVL.PS(xRegisterSSE(rtreg), ptr128[&dummyValue[0]]);
full.SetTarget(); full.SetTarget();
xBLEND.PS(xRegisterSSE(rtreg), xRegisterSSE(t0reg), 0x3);
_freeXMMreg(t0reg); _freeXMMreg(t0reg);
_freeXMMreg(t1reg); _freeXMMreg(t1reg);
_freeXMMreg(t2reg); _freeXMMreg(t2reg);
@ -689,9 +682,9 @@ void recSDL()
xSUB(edx, eax); xSUB(edx, eax);
// Generate mask 128-(shiftx8) xPSRA.W does bit for bit // Generate mask 128-(shiftx8) xPSRA.W does bit for bit
xMOVDZX(xRegisterSSE(t1reg), eax); xMOVDZX(xRegisterSSE(t1reg), eax);
xMOVQZX(xRegisterSSE(t0reg), ptr128[&SHIFT_MASKS[0][0]]); xPCMP.EQD(xRegisterSSE(t0reg), xRegisterSSE(t0reg));
xPSLL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg)); xPSLL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg));
xMOVQZX(xRegisterSSE(t1reg), ptr128[&dummyValue[0]]); // This line is super slow, but using MOVDQA/MOVAPS is even slower! xMOVQZX(xRegisterSSE(t1reg), ptr64[&dummyValue[0]]); // This line is super slow, but using MOVDQA/MOVAPS is even slower!
xPAND(xRegisterSSE(t0reg), xRegisterSSE(t1reg)); xPAND(xRegisterSSE(t0reg), xRegisterSSE(t1reg));
// Shift over reg value (shift, PSLL.Q multiplies by 8) // Shift over reg value (shift, PSLL.Q multiplies by 8)
@ -700,7 +693,7 @@ void recSDL()
xPOR(xRegisterSSE(rtreg), xRegisterSSE(t0reg)); xPOR(xRegisterSSE(rtreg), xRegisterSSE(t0reg));
skip.SetTarget(); skip.SetTarget();
xMOVQ(ptr128[&dummyValue[0]], xRegisterSSE(rtreg)); xMOVQ(ptr64[&dummyValue[0]], xRegisterSSE(rtreg));
_deleteGPRtoXMMreg(_Rt_, 3); _deleteGPRtoXMMreg(_Rt_, 3);
_freeXMMreg(t0reg); _freeXMMreg(t0reg);
@ -790,9 +783,9 @@ void recSDR()
xSUB(edx, eax); xSUB(edx, eax);
// Generate mask 128-(shiftx8) xPSRA.W does bit for bit // Generate mask 128-(shiftx8) xPSRA.W does bit for bit
xMOVDZX(xRegisterSSE(t1reg), edx); xMOVDZX(xRegisterSSE(t1reg), edx);
xMOVQZX(xRegisterSSE(t0reg), ptr128[&SHIFT_MASKS[0][0]]); xPCMP.EQD(xRegisterSSE(t0reg), xRegisterSSE(t0reg));
xPSRL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg)); xPSRL.Q(xRegisterSSE(t0reg), xRegisterSSE(t1reg));
xMOVQZX(xRegisterSSE(t1reg), ptr128[&dummyValue[0]]); // This line is super slow, but using MOVDQA/MOVAPS is even slower! xMOVQZX(xRegisterSSE(t1reg), ptr64[&dummyValue[0]]); // This line is super slow, but using MOVDQA/MOVAPS is even slower!
xPAND(xRegisterSSE(t0reg), xRegisterSSE(t1reg)); xPAND(xRegisterSSE(t0reg), xRegisterSSE(t1reg));
// Shift over reg value (shift, PSLL.Q multiplies by 8) // Shift over reg value (shift, PSLL.Q multiplies by 8)
@ -801,7 +794,7 @@ void recSDR()
xPOR(xRegisterSSE(rtreg), xRegisterSSE(t0reg)); xPOR(xRegisterSSE(rtreg), xRegisterSSE(t0reg));
skip.SetTarget(); skip.SetTarget();
xMOVQ(ptr128[&dummyValue[0]], xRegisterSSE(rtreg)); xMOVQ(ptr64[&dummyValue[0]], xRegisterSSE(rtreg));
_deleteGPRtoXMMreg(_Rt_, 3); _deleteGPRtoXMMreg(_Rt_, 3);
_freeXMMreg(t0reg); _freeXMMreg(t0reg);