Merge pull request #11305 from JosJuice/jitarm64-optimize-ps-merge

JitArm64: Optimize ps_mergeXX

Source/Core/Common/Arm64Emitter.cpp

@@ -2334,6 +2334,16 @@ void ARM64FloatEmitter::EmitPermute(u32 size, u32 op, ARM64Reg Rd, ARM64Reg Rn,
           (1 << 11) | (DecodeReg(Rn) << 5) | DecodeReg(Rd));
 }
 
+void ARM64FloatEmitter::EmitExtract(u32 imm4, u32 op, ARM64Reg Rd, ARM64Reg Rn, ARM64Reg Rm)
+{
+  ASSERT_MSG(DYNA_REC, !IsSingle(Rd), "Singles are not supported!");
+
+  bool quad = IsQuad(Rd);
+
+  Write32((quad << 30) | (23 << 25) | (op << 22) | (DecodeReg(Rm) << 16) | (imm4 << 11) |
+          (DecodeReg(Rn) << 5) | DecodeReg(Rd));
+}
+
 void ARM64FloatEmitter::EmitScalarImm(bool M, bool S, u32 type, u32 imm5, ARM64Reg Rd, u32 imm8)
 {
   ASSERT_MSG(DYNA_REC, !IsQuad(Rd), "Vector is not supported!");
@@ -3540,6 +3550,12 @@ void ARM64FloatEmitter::ZIP2(u8 size, ARM64Reg Rd, ARM64Reg Rn, ARM64Reg Rm)
   EmitPermute(size, 0b111, Rd, Rn, Rm);
 }
 
+// Extract
+void ARM64FloatEmitter::EXT(ARM64Reg Rd, ARM64Reg Rn, ARM64Reg Rm, u32 index)
+{
+  EmitExtract(index, 0, Rd, Rn, Rm);
+}
+
 // Scalar shift by immediate
 void ARM64FloatEmitter::SHL(ARM64Reg Rd, ARM64Reg Rn, u32 shift)
 {

Source/Core/Common/Arm64Emitter.h

@@ -1247,6 +1247,9 @@ public:
   void TRN2(u8 size, ARM64Reg Rd, ARM64Reg Rn, ARM64Reg Rm);
   void ZIP2(u8 size, ARM64Reg Rd, ARM64Reg Rn, ARM64Reg Rm);
 
+  // Extract
+  void EXT(ARM64Reg Rd, ARM64Reg Rn, ARM64Reg Rm, u32 index);
+
   // Scalar shift by immediate
   void SHL(ARM64Reg Rd, ARM64Reg Rn, u32 shift);
   void URSHR(ARM64Reg Rd, ARM64Reg Rn, u32 shift);
@@ -1305,6 +1308,7 @@ private:
   void EmitCompare(bool M, bool S, u32 op, u32 opcode2, ARM64Reg Rn, ARM64Reg Rm);
   void EmitCondSelect(bool M, bool S, CCFlags cond, ARM64Reg Rd, ARM64Reg Rn, ARM64Reg Rm);
   void EmitPermute(u32 size, u32 op, ARM64Reg Rd, ARM64Reg Rn, ARM64Reg Rm);
+  void EmitExtract(u32 imm4, u32 op, ARM64Reg Rd, ARM64Reg Rn, ARM64Reg Rm);
   void EmitScalarImm(bool M, bool S, u32 type, u32 imm5, ARM64Reg Rd, u32 imm8);
   void EmitShiftImm(bool Q, bool U, u32 imm, u32 opcode, ARM64Reg Rd, ARM64Reg Rn);
   void EmitScalarShiftImm(bool U, u32 imm, u32 opcode, ARM64Reg Rd, ARM64Reg Rn);

Source/Core/Core/PowerPC/JitArm64/JitArm64_Paired.cpp

@@ -33,9 +33,9 @@ void JitArm64::ps_mergeXX(UGeckoInstruction inst)
   const u8 size = singles ? 32 : 64;
   const auto reg_encoder = singles ? EncodeRegToDouble : EncodeRegToQuad;
 
-  const ARM64Reg VA = fpr.R(a, type);
+  const ARM64Reg VA = reg_encoder(fpr.R(a, type));
-  const ARM64Reg VB = fpr.R(b, type);
+  const ARM64Reg VB = reg_encoder(fpr.R(b, type));
-  const ARM64Reg VD = fpr.RW(d, type);
+  const ARM64Reg VD = reg_encoder(fpr.RW(d, type));
 
   switch (inst.SUBOP10)
   {
@@ -43,23 +43,20 @@ void JitArm64::ps_mergeXX(UGeckoInstruction inst)
     m_float_emit.TRN1(size, VD, VA, VB);
     break;
   case 560:  // 01
-    m_float_emit.INS(size, VD, 0, VA, 0);
+    if (d != b)
+    {
+      if (d != a)
+        m_float_emit.MOV(VD, VA);
+      if (a != b)
         m_float_emit.INS(size, VD, 1, VB, 1);
+    }
+    else if (d != a)
+    {
+      m_float_emit.INS(size, VD, 0, VA, 0);
+    }
     break;
   case 592:  // 10
-    if (d != a && d != b)
+    m_float_emit.EXT(VD, VA, VB, size >> 3);
-    {
-      m_float_emit.INS(size, VD, 0, VA, 1);
-      m_float_emit.INS(size, VD, 1, VB, 0);
-    }
-    else
-    {
-      ARM64Reg V0 = fpr.GetReg();
-      m_float_emit.INS(size, V0, 0, VA, 1);
-      m_float_emit.INS(size, V0, 1, VB, 0);
-      m_float_emit.MOV(reg_encoder(VD), reg_encoder(V0));
-      fpr.Unlock(V0);
-    }
     break;
   case 624:  // 11
     m_float_emit.TRN2(size, VD, VA, VB);