[wasm-simd] Share f64x2.convert_low_i32x4_u implementation

We create a ExternalReferenceAsOperand helper function in
SharedTurboAssemblerBase that delegates to the actual arch specific
implementation of TurboAssembler, because the ia32 and x64
ExternalReferenceAsOperand differs slightly in their implementation.

Bug: v8:11589
Change-Id: I378ea6b72fb2bba1a37482cc31cd58db0ba35721
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3114604Reviewed-by: Adam Klein <adamk@chromium.org>
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Cr-Commit-Position: refs/heads/main@{#76494}

src/codegen/ia32/macro-assembler-ia32.cc

@@ -701,25 +701,6 @@ void TurboAssembler::I8x16Popcnt(XMMRegister dst, XMMRegister src,
   }
 }
 
-void TurboAssembler::F64x2ConvertLowI32x4U(XMMRegister dst, XMMRegister src,
-                                           Register tmp) {
-  // dst = [ src_low, 0x43300000, src_high, 0x4330000 ];
-  // 0x43300000'00000000 is a special double where the significand bits
-  // precisely represents all uint32 numbers.
-  if (!CpuFeatures::IsSupported(AVX) && dst != src) {
-    movaps(dst, src);
-    src = dst;
-  }
-  Unpcklps(dst, src,
-           ExternalReferenceAsOperand(
-               ExternalReference::
-                   address_of_wasm_f64x2_convert_low_i32x4_u_int_mask(),
-               tmp));
-  Subpd(dst, dst,
-        ExternalReferenceAsOperand(
-            ExternalReference::address_of_wasm_double_2_power_52(), tmp));
-}
-
 void TurboAssembler::I32x4TruncSatF64x2SZero(XMMRegister dst, XMMRegister src,
                                              XMMRegister scratch,
                                              Register tmp) {

src/codegen/ia32/macro-assembler-ia32.h

@@ -397,7 +397,6 @@ class V8_EXPORT_PRIVATE TurboAssembler
   // Defined here to allow usage on both TurboFan and Liftoff.
   void I8x16Popcnt(XMMRegister dst, XMMRegister src, XMMRegister tmp1,
                    XMMRegister tmp2, Register scratch);
-  void F64x2ConvertLowI32x4U(XMMRegister dst, XMMRegister src, Register tmp);
   void I32x4TruncSatF64x2SZero(XMMRegister dst, XMMRegister src,
                                XMMRegister scratch, Register tmp);
   void I32x4TruncSatF64x2UZero(XMMRegister dst, XMMRegister src,

src/codegen/shared-ia32-x64/macro-assembler-shared-ia32-x64.h

@@ -418,8 +418,36 @@ class V8_EXPORT_PRIVATE SharedTurboAssembler : public TurboAssemblerBase {
 template <typename Impl>
 class V8_EXPORT_PRIVATE SharedTurboAssemblerBase : public SharedTurboAssembler {
   using SharedTurboAssembler::SharedTurboAssembler;
-  // TODO(zhin): intentionally empty for now, will move polymorphic functions
-  // here in future changes.
+
+ public:
+  void F64x2ConvertLowI32x4U(XMMRegister dst, XMMRegister src,
+                             Register scratch) {
+    ASM_CODE_COMMENT(this);
+    // dst = [ src_low, 0x43300000, src_high, 0x4330000 ];
+    // 0x43300000'00000000 is a special double where the significand bits
+    // precisely represents all uint32 numbers.
+    if (!CpuFeatures::IsSupported(AVX) && dst != src) {
+      movaps(dst, src);
+      src = dst;
+    }
+    Unpcklps(dst, src,
+             ExternalReferenceAsOperand(
+                 ExternalReference::
+                     address_of_wasm_f64x2_convert_low_i32x4_u_int_mask(),
+                 scratch));
+    Subpd(dst,
+          ExternalReferenceAsOperand(
+              ExternalReference::address_of_wasm_double_2_power_52(), scratch));
+  }
+
+ private:
+  // All implementation-specific methods must be called through this.
+  Impl* impl() { return static_cast<Impl*>(this); }
+
+  Operand ExternalReferenceAsOperand(ExternalReference reference,
+                                     Register scratch) {
+    return impl()->ExternalReferenceAsOperand(reference, scratch);
+  }
 };
 
 }  // namespace internal

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

@@ -2302,22 +2302,6 @@ void TurboAssembler::I8x16Popcnt(XMMRegister dst, XMMRegister src,
   }
 }
 
-void TurboAssembler::F64x2ConvertLowI32x4U(XMMRegister dst, XMMRegister src) {
-  // dst = [ src_low, 0x43300000, src_high, 0x4330000 ];
-  // 0x43300000'00000000 is a special double where the significand bits
-  // precisely represents all uint32 numbers.
-  if (!CpuFeatures::IsSupported(AVX) && dst != src) {
-    movaps(dst, src);
-    src = dst;
-  }
-  Unpcklps(dst, src,
-           ExternalReferenceAsOperand(
-               ExternalReference::
-                   address_of_wasm_f64x2_convert_low_i32x4_u_int_mask()));
-  Subpd(dst, ExternalReferenceAsOperand(
-                 ExternalReference::address_of_wasm_double_2_power_52()));
-}
-
 void TurboAssembler::I32x4TruncSatF64x2SZero(XMMRegister dst, XMMRegister src) {
   if (CpuFeatures::IsSupported(AVX)) {
     CpuFeatureScope avx_scope(this, AVX);

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

@@ -481,7 +481,6 @@ class V8_EXPORT_PRIVATE TurboAssembler
   // Defined here to allow usage on both TurboFan and Liftoff.
   void I8x16Popcnt(XMMRegister dst, XMMRegister src, XMMRegister tmp);
 
-  void F64x2ConvertLowI32x4U(XMMRegister dst, XMMRegister src);
   void I32x4TruncSatF64x2SZero(XMMRegister dst, XMMRegister src);
   void I32x4TruncSatF64x2UZero(XMMRegister dst, XMMRegister src);
 

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

@@ -2622,7 +2622,7 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
     }
     case kX64F64x2ConvertLowI32x4U: {
       __ F64x2ConvertLowI32x4U(i.OutputSimd128Register(),
-                               i.InputSimd128Register(0));
+                               i.InputSimd128Register(0), kScratchRegister);
       break;
     }
     case kX64F64x2PromoteLowF32x4: {

src/wasm/baseline/x64/liftoff-assembler-x64.h

@@ -3753,7 +3753,7 @@ void LiftoffAssembler::emit_f64x2_convert_low_i32x4_s(LiftoffRegister dst,
 
 void LiftoffAssembler::emit_f64x2_convert_low_i32x4_u(LiftoffRegister dst,
                                                       LiftoffRegister src) {
-  F64x2ConvertLowI32x4U(dst.fp(), src.fp());
+  F64x2ConvertLowI32x4U(dst.fp(), src.fp(), kScratchRegister);
 }
 
 void LiftoffAssembler::emit_f64x2_promote_low_f32x4(LiftoffRegister dst,