[wasm-simd][x64] Optimize f64x2.extract_lane

pextrq + movq crosses register files twice, which is not efficient.

Optimize this by:
- checking if lane 0, do nothing if dst == src (macro-assembler helper)
- use vmovhlps on AVX, with src as the operands to avoid false
dependency on dst
- use movhlps otherwise, this is shorter than shufpd, and faster on
older system

Change-Id: I3486d87224c048b3229c2f92359b8b8e6d5fd025
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2589056
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Reviewed-by: Bill Budge <bbudge@chromium.org>
Cr-Commit-Position: refs/heads/master@{#71751}

src/codegen/x64/assembler-x64.h

@@ -1284,6 +1284,9 @@ class V8_EXPORT_PRIVATE Assembler : public AssemblerBase {
   void pshuflw(XMMRegister dst, XMMRegister src, uint8_t shuffle);
   void pshuflw(XMMRegister dst, Operand src, uint8_t shuffle);
 
+  void movhlps(XMMRegister dst, XMMRegister src) {
+    sse_instr(dst, src, 0x0F, 0x12);
+  }
   void movlhps(XMMRegister dst, XMMRegister src) {
     sse_instr(dst, src, 0x0F, 0x16);
   }
@@ -1386,6 +1389,9 @@ class V8_EXPORT_PRIVATE Assembler : public AssemblerBase {
   void vmovlhps(XMMRegister dst, XMMRegister src1, XMMRegister src2) {
     vinstr(0x16, dst, src1, src2, kNone, k0F, kWIG);
   }
+  void vmovhlps(XMMRegister dst, XMMRegister src1, XMMRegister src2) {
+    vinstr(0x12, dst, src1, src2, kNone, k0F, kWIG);
+  }
   void vcvtss2sd(XMMRegister dst, XMMRegister src1, XMMRegister src2) {
     vinstr(0x5a, dst, src1, src2, kF3, k0F, kWIG);
   }

src/codegen/x64/macro-assembler-x64.cc

@@ -1293,6 +1293,12 @@ void TurboAssembler::Move(Register dst, Register src) {
   }
 }
 
+void TurboAssembler::Move(XMMRegister dst, XMMRegister src) {
+  if (dst != src) {
+    movaps(dst, src);
+  }
+}
+
 void TurboAssembler::MovePair(Register dst0, Register src0, Register dst1,
                               Register src1) {
   if (dst0 != src1) {

src/codegen/x64/macro-assembler-x64.h

@@ -442,6 +442,7 @@ class V8_EXPORT_PRIVATE TurboAssembler : public TurboAssemblerBase {
 
   // Move if the registers are not identical.
   void Move(Register target, Register source);
+  void Move(XMMRegister target, XMMRegister source);
 
   void Move(Register dst, Handle<HeapObject> source,
             RelocInfo::Mode rmode = RelocInfo::FULL_EMBEDDED_OBJECT);

src/compiler/backend/x64/code-generator-x64.cc

@@ -2453,8 +2453,21 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       break;
     }
     case kX64F64x2ExtractLane: {
-      __ Pextrq(kScratchRegister, i.InputSimd128Register(0), i.InputInt8(1));
-      __ Movq(i.OutputDoubleRegister(), kScratchRegister);
+      DoubleRegister dst = i.OutputDoubleRegister();
+      XMMRegister src = i.InputSimd128Register(0);
+      uint8_t lane = i.InputUint8(1);
+      if (lane == 0) {
+        __ Move(dst, src);
+      } else {
+        DCHECK_EQ(1, lane);
+        if (CpuFeatures::IsSupported(AVX)) {
+          CpuFeatureScope avx_scope(tasm(), AVX);
+          // Pass src as operand to avoid false-dependency on dst.
+          __ vmovhlps(dst, src, src);
+        } else {
+          __ movhlps(dst, src);
+        }
+      }
       break;
     }
     case kX64F64x2Sqrt: {

src/diagnostics/x64/disasm-x64.cc

@@ -1381,9 +1381,15 @@ int DisassemblerX64::AVXInstruction(byte* data) {
         AppendToBuffer(",%s", NameOfXMMRegister(regop));
         break;
       case 0x12:
-        AppendToBuffer("vmovlps %s,%s,", NameOfXMMRegister(regop),
-                       NameOfXMMRegister(vvvv));
-        current += PrintRightXMMOperand(current);
+        if (mod == 0b11) {
+          AppendToBuffer("vmovhlps %s,%s,", NameOfXMMRegister(regop),
+                         NameOfXMMRegister(vvvv));
+          current += PrintRightXMMOperand(current);
+        } else {
+          AppendToBuffer("vmovlps %s,%s,", NameOfXMMRegister(regop),
+                         NameOfXMMRegister(vvvv));
+          current += PrintRightXMMOperand(current);
+        }
         break;
       case 0x13:
         AppendToBuffer("vmovlps ");
@@ -2065,8 +2071,13 @@ int DisassemblerX64::TwoByteOpcodeInstruction(byte* data) {
     // movups xmm/m128, xmm
     current += PrintOperands("movups", XMMOPER_XMMREG_OP_ORDER, current);
   } else if (opcode == 0x12) {
+    // movhlps xmm1, xmm2
     // movlps xmm1, m64
-    current += PrintOperands("movlps", XMMREG_OPER_OP_ORDER, current);
+    if (mod == 0b11) {
+      current += PrintOperands("movhlps", XMMREG_XMMOPER_OP_ORDER, current);
+    } else {
+      current += PrintOperands("movlps", XMMREG_OPER_OP_ORDER, current);
+    }
   } else if (opcode == 0x13) {
     // movlps m64, xmm1
     AppendToBuffer("movlps ");

test/cctest/test-disasm-x64.cc

@@ -401,8 +401,10 @@ TEST(DisasmX64) {
     __ movdqu(xmm0, Operand(rsp, 12));
     __ movdqu(Operand(rsp, 12), xmm0);
     __ movdqu(xmm1, xmm0);
+    __ movhlps(xmm5, xmm1);
     __ movlps(xmm8, Operand(rbx, rcx, times_4, 10000));
     __ movlps(Operand(rbx, rcx, times_4, 10000), xmm9);
+    __ movlhps(xmm5, xmm1);
     __ movhps(xmm8, Operand(rbx, rcx, times_4, 10000));
     __ movhps(Operand(rbx, rcx, times_4, 10000), xmm9);
     __ shufps(xmm0, xmm9, 0x0);
@@ -577,7 +579,6 @@ TEST(DisasmX64) {
       __ movups(xmm5, xmm1);
       __ movups(xmm5, Operand(rdx, 4));
       __ movups(Operand(rdx, 4), xmm5);
-      __ movlhps(xmm5, xmm1);
       __ pmulld(xmm5, xmm1);
       __ pmulld(xmm5, Operand(rdx, 4));
       __ pmullw(xmm5, xmm1);
@@ -659,8 +660,10 @@ TEST(DisasmX64) {
       __ vmovdqu(xmm9, Operand(rbx, rcx, times_4, 10000));
       __ vmovdqu(Operand(rbx, rcx, times_4, 10000), xmm0);
 
+      __ vmovhlps(xmm1, xmm3, xmm5);
       __ vmovlps(xmm8, xmm9, Operand(rbx, rcx, times_4, 10000));
       __ vmovlps(Operand(rbx, rcx, times_4, 10000), xmm9);
+      __ vmovlhps(xmm1, xmm3, xmm5);
       __ vmovhps(xmm8, xmm9, Operand(rbx, rcx, times_4, 10000));
       __ vmovhps(Operand(rbx, rcx, times_4, 10000), xmm12);
 
@@ -693,7 +696,6 @@ TEST(DisasmX64) {
       __ vmovups(xmm5, xmm1);
       __ vmovups(xmm5, Operand(rdx, 4));
       __ vmovups(Operand(rdx, 4), xmm5);
-      __ vmovlhps(xmm1, xmm3, xmm5);
 
       __ vandps(xmm0, xmm9, xmm2);
       __ vandps(xmm9, xmm1, Operand(rbx, rcx, times_4, 10000));