Bug 1115766 - SpiderMonkey: Eliminate a copy in shuffleFloat32 with AVX r=bbouvier

js/src/jit/shared/CodeGenerator-x86-shared.cpp

@@ -2145,9 +2145,8 @@ CodeGeneratorX86Shared::visitSimdSplatX4(LSimdSplatX4 *ins)
       }
       case MIRType_Float32x4: {
         FloatRegister r = ToFloatRegister(ins->getOperand(0));
-        if (r != output)
-            masm.moveFloat32x4(r, output);
-        masm.vshufps(0, output, output, output);
+        FloatRegister rCopy = masm.reusedInputFloat32x4(r, output);
+        masm.vshufps(0, rCopy, rCopy, output);
         break;
       }
       default:

js/src/jit/shared/MacroAssembler-x86-shared.h

@@ -1011,12 +1011,11 @@ class MacroAssemblerX86Shared : public Assembler
         // clobber the output with the input and apply the instruction
         // afterwards.
         // Note: this is useAtStart-safe because src isn't read afterwards.
-        if (src != dest)
-            moveFloat32x4(src, dest);
-        vshufps(mask, dest, dest, dest);
+        FloatRegister srcCopy = reusedInputFloat32x4(src, dest);
+        vshufps(mask, srcCopy, srcCopy, dest);
     }
     void shuffleMix(uint32_t mask, const Operand &src, FloatRegister dest) {
-        // Note this uses vshufps, which is a cross-domain penaly on CPU where it
+        // Note this uses vshufps, which is a cross-domain penalty on CPU where it
         // applies, but that's the way clang and gcc do it.
         vshufps(mask, src, dest, dest);
     }