Merge pull request #17569 from hrydgard/arm64dec-optimize-saved-regs

ARM64: Optimize saved registers in vertex decoder.

Common/Arm64Emitter.h

@@ -94,7 +94,7 @@ enum ARM64Reg
 
 // R19-R28. R29 (FP), R30 (LR) are always saved and FP updated appropriately.
 const u32 ALL_CALLEE_SAVED = 0x1FF80000;
-const u32 ALL_CALLEE_SAVED_FP = 0x0000FF00;  // d8-d15
+const u32 ALL_CALLEE_SAVED_FP = 0x0000FF00;  // q8-q15
 
 inline bool Is64Bit(ARM64Reg reg) { return (reg & 0x20) != 0; }
 inline bool IsSingle(ARM64Reg reg) { return (reg & 0xC0) == 0x40; }

GPU/Common/VertexDecoderArm64.cpp

@@ -64,9 +64,9 @@ static const ARM64Reg neonScratchRegQ = Q2;
 static const ARM64Reg neonUVScaleReg = D0;
 static const ARM64Reg neonUVOffsetReg = D1;
 
-static const ARM64Reg src[3] = {S2, S3, S8};
-static const ARM64Reg srcD[3] = {D2, D3, D8};
-static const ARM64Reg srcQ[3] = {Q2, Q3, Q8};
+static const ARM64Reg src[2] = {S2, S3};
+static const ARM64Reg srcD = D2;
+static const ARM64Reg srcQ = Q2;
 
 static const ARM64Reg srcNEON = Q8;
 static const ARM64Reg accNEON = Q9;
@@ -169,8 +169,11 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec, int
 
 	// if (skinning) log = true;
 
-	uint64_t regs_to_save = Arm64Gen::ALL_CALLEE_SAVED;
-	uint64_t regs_to_save_fp = Arm64Gen::ALL_CALLEE_SAVED_FP;
+	// GPRs 0-15 do not need to be saved.
+	// We don't use any higher GPRs than 16. So:
+	uint64_t regs_to_save = 1 << 16; // Arm64Gen::ALL_CALLEE_SAVED;
+	// We only need to save Q8-Q15 if skinning is used.
+	uint64_t regs_to_save_fp = dec.skinInDecode ? Arm64Gen::ALL_CALLEE_SAVED_FP : 0;
 	fp.ABI_PushRegisters(regs_to_save, regs_to_save_fp);
 
 	// Keep the scale/offset in a few fp registers if we need it.
@@ -645,12 +648,12 @@ void VertexDecoderJitCache::Jit_TcFloatPrescale() {
 
 void VertexDecoderJitCache::Jit_PosS8() {
 	Jit_AnyS8ToFloat(dec_->posoff);
-	fp.STUR(128, srcQ[0], dstReg, dec_->decFmt.posoff);
+	fp.STUR(128, srcQ, dstReg, dec_->decFmt.posoff);
 }
 
 void VertexDecoderJitCache::Jit_PosS16() {
 	Jit_AnyS16ToFloat(dec_->posoff);
-	fp.STUR(128, srcQ[0], dstReg, dec_->decFmt.posoff);
+	fp.STUR(128, srcQ, dstReg, dec_->decFmt.posoff);
 }
 
 void VertexDecoderJitCache::Jit_PosFloat() {
@@ -677,8 +680,8 @@ void VertexDecoderJitCache::Jit_PosS8Through() {
 void VertexDecoderJitCache::Jit_PosS16Through() {
 	// Start with X and Y (which is signed.)
 	fp.LDUR(32, src[0], srcReg, dec_->posoff);
-	fp.SXTL(16, srcD[0], src[0]);
-	fp.SCVTF(32, srcD[0], srcD[0]);
+	fp.SXTL(16, srcD, src[0]);
+	fp.SCVTF(32, srcD, srcD);
 	fp.STUR(64, src[0], dstReg, dec_->decFmt.posoff);
 	// Now load in Z (which is unsigned.)
 	LDRH(INDEX_UNSIGNED, tempReg3, srcReg, dec_->posoff + 4);
@@ -744,7 +747,7 @@ void VertexDecoderJitCache::Jit_NormalS16Skin() {
 }
 
 void VertexDecoderJitCache::Jit_NormalFloatSkin() {
-	fp.LDUR(128, srcQ[0], srcReg, dec_->nrmoff);
+	fp.LDUR(128, srcQ, srcReg, dec_->nrmoff);
 	Jit_WriteMatrixMul(dec_->decFmt.nrmoff, false);
 }
 
@@ -759,28 +762,28 @@ void VertexDecoderJitCache::Jit_PosS16Skin() {
 }
 
 void VertexDecoderJitCache::Jit_PosFloatSkin() {
-	fp.LDUR(128, srcQ[0], srcReg, dec_->posoff);
+	fp.LDUR(128, srcQ, srcReg, dec_->posoff);
 	Jit_WriteMatrixMul(dec_->decFmt.posoff, true);
 }
 
 void VertexDecoderJitCache::Jit_AnyS8ToFloat(int srcoff) {
 	fp.LDUR(32, src[0], srcReg, srcoff);
-	fp.SXTL(8, srcD[0], src[0]);
-	fp.SXTL(16, srcQ[0], srcD[0]);
-	fp.SCVTF(32, srcQ[0], srcQ[0], 7);
+	fp.SXTL(8, srcD, src[0]);
+	fp.SXTL(16, srcQ, srcD);
+	fp.SCVTF(32, srcQ, srcQ, 7);
 }
 
 void VertexDecoderJitCache::Jit_AnyS16ToFloat(int srcoff) {
 	fp.LDUR(64, src[0], srcReg, srcoff);
-	fp.SXTL(16, srcQ[0], srcD[0]);
-	fp.SCVTF(32, srcQ[0], srcQ[0], 15);
+	fp.SXTL(16, srcQ, srcD);
+	fp.SCVTF(32, srcQ, srcQ, 15);
 }
 
 void VertexDecoderJitCache::Jit_WriteMatrixMul(int outOff, bool pos) {
-	// Multiply with the matrix sitting in Q4-Q7.
-	fp.FMUL(32, accNEON, Q4, srcQ[0], 0);
-	fp.FMLA(32, accNEON, Q5, srcQ[0], 1);
-	fp.FMLA(32, accNEON, Q6, srcQ[0], 2);
+	// Multiply srcQ with the matrix sitting in Q4-Q7.
+	fp.FMUL(32, accNEON, Q4, srcQ, 0);
+	fp.FMLA(32, accNEON, Q5, srcQ, 1);
+	fp.FMLA(32, accNEON, Q6, srcQ, 2);
 	if (pos) {
 		fp.FADD(32, accNEON, accNEON, Q7);
 	}