[wasm-simd][liftoff][x64] Implement saturating rounding multiply high

Create a macro-assembler helper function to enable code sharing between
Liftoff and TurboFan.

Bug: v8:10971
Change-Id: I10228b502d959dbde670c38c44992fa7133dab7e
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2621859Reviewed-by: Clemens Backes <clemensb@chromium.org>
Reviewed-by: Bill Budge <bbudge@chromium.org>
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Cr-Commit-Position: refs/heads/master@{#72058}

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

@@ -2211,6 +2211,17 @@ void TurboAssembler::I16x8ExtMul(XMMRegister dst, XMMRegister src1,
   }
 }
 
+void TurboAssembler::I16x8Q15MulRSatS(XMMRegister dst, XMMRegister src1,
+                                      XMMRegister src2) {
+  // k = i16x8.splat(0x8000)
+  Pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
+  Psllw(kScratchDoubleReg, byte{15});
+
+  Pmulhrsw(dst, src1, src2);
+  Pcmpeqw(kScratchDoubleReg, dst);
+  Pxor(dst, kScratchDoubleReg);
+}
+
 void TurboAssembler::Abspd(XMMRegister dst) {
   Andps(dst, ExternalReferenceAsOperand(
                  ExternalReference::address_of_double_abs_constant()));

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

@@ -592,6 +592,8 @@ class V8_EXPORT_PRIVATE TurboAssembler : public TurboAssemblerBase {
   void I16x8ExtMul(XMMRegister dst, XMMRegister src1, XMMRegister src2,
                    bool low, bool is_signed);
 
+  void I16x8Q15MulRSatS(XMMRegister dst, XMMRegister src1, XMMRegister src2);
+
   void Abspd(XMMRegister dst);
   void Negpd(XMMRegister dst);
 

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

@@ -3351,16 +3351,8 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       break;
     }
     case kX64I16x8Q15MulRSatS: {
-      XMMRegister dst = i.OutputSimd128Register();
-      XMMRegister src0 = i.InputSimd128Register(0);
-      XMMRegister src1 = i.InputSimd128Register(1);
-      // k = i16x8.splat(0x8000)
-      __ Pcmpeqd(kScratchDoubleReg, kScratchDoubleReg);
-      __ Psllw(kScratchDoubleReg, byte{15});
-
-      __ Pmulhrsw(dst, src0, src1);
-      __ Pcmpeqw(kScratchDoubleReg, dst);
-      __ Pxor(dst, kScratchDoubleReg);
+      __ I16x8Q15MulRSatS(i.OutputSimd128Register(), i.InputSimd128Register(0),
+                          i.InputSimd128Register(1));
       break;
     }
     case kX64I8x16Splat: {

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

@@ -3317,6 +3317,12 @@ void LiftoffAssembler::emit_i16x8_extmul_high_i8x16_u(LiftoffRegister dst,
         src2.high_fp());
 }
 
+void LiftoffAssembler::emit_i16x8_q15mulr_sat_s(LiftoffRegister dst,
+                                                LiftoffRegister src1,
+                                                LiftoffRegister src2) {
+  bailout(kSimd, "i16x8_q15mulr_sat_s");
+}
+
 void LiftoffAssembler::emit_i8x16_shuffle(LiftoffRegister dst,
                                           LiftoffRegister lhs,
                                           LiftoffRegister rhs,

src/wasm/baseline/arm64/liftoff-assembler-arm64.h

@@ -2904,6 +2904,12 @@ void LiftoffAssembler::emit_i16x8_extmul_high_i8x16_u(LiftoffRegister dst,
   Umull2(dst.fp().V8H(), src1.fp().V16B(), src2.fp().V16B());
 }
 
+void LiftoffAssembler::emit_i16x8_q15mulr_sat_s(LiftoffRegister dst,
+                                                LiftoffRegister src1,
+                                                LiftoffRegister src2) {
+  bailout(kSimd, "i16x8_q15mulr_sat_s");
+}
+
 void LiftoffAssembler::emit_i32x4_abs(LiftoffRegister dst,
                                       LiftoffRegister src) {
   Abs(dst.fp().V4S(), src.fp().V4S());

src/wasm/baseline/ia32/liftoff-assembler-ia32.h

@@ -3574,6 +3574,12 @@ void LiftoffAssembler::emit_i16x8_extmul_high_i8x16_u(LiftoffRegister dst,
               /*low=*/false, /*is_signed=*/false);
 }
 
+void LiftoffAssembler::emit_i16x8_q15mulr_sat_s(LiftoffRegister dst,
+                                                LiftoffRegister src1,
+                                                LiftoffRegister src2) {
+  bailout(kSimd, "i16x8_q15mulr_sat_s");
+}
+
 void LiftoffAssembler::emit_i32x4_neg(LiftoffRegister dst,
                                       LiftoffRegister src) {
   if (dst.fp() == src.fp()) {

src/wasm/baseline/liftoff-assembler.h

@@ -1036,6 +1036,9 @@ class LiftoffAssembler : public TurboAssembler {
   inline void emit_i16x8_extmul_high_i8x16_u(LiftoffRegister dst,
                                              LiftoffRegister src1,
                                              LiftoffRegister src2);
+  inline void emit_i16x8_q15mulr_sat_s(LiftoffRegister dst,
+                                       LiftoffRegister src1,
+                                       LiftoffRegister src2);
   inline void emit_i32x4_neg(LiftoffRegister dst, LiftoffRegister src);
   inline void emit_v32x4_anytrue(LiftoffRegister dst, LiftoffRegister src);
   inline void emit_v32x4_alltrue(LiftoffRegister dst, LiftoffRegister src);

src/wasm/baseline/liftoff-compiler.cc

@@ -2935,6 +2935,9 @@ class LiftoffCompiler {
       case wasm::kExprI16x8ExtMulHighI8x16U:
         return EmitBinOp<kS128, kS128>(
             &LiftoffAssembler::emit_i16x8_extmul_high_i8x16_u);
+      case wasm::kExprI16x8Q15MulRSatS:
+        return EmitBinOp<kS128, kS128>(
+            &LiftoffAssembler::emit_i16x8_q15mulr_sat_s);
       case wasm::kExprI32x4Neg:
         return EmitUnOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_neg);
       case wasm::kExprV32x4AnyTrue:

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

@@ -3175,6 +3175,12 @@ void LiftoffAssembler::emit_i16x8_extmul_high_i8x16_u(LiftoffRegister dst,
               /*is_signed=*/false);
 }
 
+void LiftoffAssembler::emit_i16x8_q15mulr_sat_s(LiftoffRegister dst,
+                                                LiftoffRegister src1,
+                                                LiftoffRegister src2) {
+  I16x8Q15MulRSatS(dst.fp(), src1.fp(), src2.fp());
+}
+
 void LiftoffAssembler::emit_i32x4_neg(LiftoffRegister dst,
                                       LiftoffRegister src) {
   if (dst.fp() == src.fp()) {