[wasm-simd] [liftoff] Implement sub_saturate on x64 and ia32

Bug: v8:9909
Change-Id: I13d7ae7b304fe55645ef7bcc9ac404782bb55d4e
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2131323Reviewed-by: Zhi An Ng <zhin@chromium.org>
Commit-Queue: Fanchen Kong <fanchen.kong@intel.com>
Cr-Commit-Position: refs/heads/master@{#67099}

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

@@ -1834,6 +1834,18 @@ void LiftoffAssembler::emit_i16x8_sub(LiftoffRegister dst, LiftoffRegister lhs,
        liftoff::GetSimd128Register(lhs), liftoff::GetSimd128Register(rhs));
 }
 
+void LiftoffAssembler::emit_i16x8_sub_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8subsaturate_s");
+}
+
+void LiftoffAssembler::emit_i16x8_sub_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8subsaturate_u");
+}
+
 void LiftoffAssembler::emit_i16x8_mul(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   vmul(Neon16, liftoff::GetSimd128Register(dst),
@@ -1948,6 +1960,18 @@ void LiftoffAssembler::emit_i8x16_sub(LiftoffRegister dst, LiftoffRegister lhs,
        liftoff::GetSimd128Register(lhs), liftoff::GetSimd128Register(rhs));
 }
 
+void LiftoffAssembler::emit_i8x16_sub_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16subsaturate_s");
+}
+
+void LiftoffAssembler::emit_i8x16_sub_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16subsaturate_u");
+}
+
 void LiftoffAssembler::emit_i8x16_mul(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   vmul(Neon8, liftoff::GetSimd128Register(dst),

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

@@ -1336,6 +1336,18 @@ void LiftoffAssembler::emit_i16x8_sub(LiftoffRegister dst, LiftoffRegister lhs,
   Sub(dst.fp().V8H(), lhs.fp().V8H(), rhs.fp().V8H());
 }
 
+void LiftoffAssembler::emit_i16x8_sub_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8subsaturate_s");
+}
+
+void LiftoffAssembler::emit_i16x8_sub_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8subsaturate_u");
+}
+
 void LiftoffAssembler::emit_i16x8_mul(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   Mul(dst.fp().V8H(), lhs.fp().V8H(), rhs.fp().V8H());
@@ -1419,6 +1431,18 @@ void LiftoffAssembler::emit_i8x16_sub(LiftoffRegister dst, LiftoffRegister lhs,
   Sub(dst.fp().V16B(), lhs.fp().V16B(), rhs.fp().V16B());
 }
 
+void LiftoffAssembler::emit_i8x16_sub_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16subsaturate_s");
+}
+
+void LiftoffAssembler::emit_i8x16_sub_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16subsaturate_u");
+}
+
 void LiftoffAssembler::emit_i8x16_mul(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   Mul(dst.fp().V16B(), lhs.fp().V16B(), rhs.fp().V16B());

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

@@ -2050,6 +2050,20 @@ void LiftoffAssembler::emit_i8x16_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                                               rhs);
 }
 
+void LiftoffAssembler::emit_i8x16_sub_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdSub<&Assembler::vpsubsb, &Assembler::psubsb>(this, dst, lhs,
+                                                                rhs);
+}
+
+void LiftoffAssembler::emit_i8x16_sub_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdSub<&Assembler::vpsubusb, &Assembler::psubusb>(this, dst,
+                                                                  lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i8x16_mul(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   static constexpr RegClass tmp_rc = reg_class_for(ValueType::kS128);
@@ -2177,6 +2191,20 @@ void LiftoffAssembler::emit_i16x8_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                                               rhs);
 }
 
+void LiftoffAssembler::emit_i16x8_sub_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdSub<&Assembler::vpsubsw, &Assembler::psubsw>(this, dst, lhs,
+                                                                rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_sub_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdSub<&Assembler::vpsubusw, &Assembler::psubusw>(this, dst,
+                                                                  lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i16x8_mul(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vpmullw, &Assembler::pmullw>(

src/wasm/baseline/liftoff-assembler.h

@@ -732,6 +732,12 @@ class LiftoffAssembler : public TurboAssembler {
                                         LiftoffRegister rhs);
   inline void emit_i8x16_sub(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
+  inline void emit_i8x16_sub_saturate_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs);
+  inline void emit_i8x16_sub_saturate_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs);
   inline void emit_i8x16_mul(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
   inline void emit_i8x16_min_s(LiftoffRegister dst, LiftoffRegister lhs,
@@ -753,6 +759,12 @@ class LiftoffAssembler : public TurboAssembler {
                                         LiftoffRegister rhs);
   inline void emit_i16x8_sub(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
+  inline void emit_i16x8_sub_saturate_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs);
+  inline void emit_i16x8_sub_saturate_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs);
   inline void emit_i16x8_mul(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
   inline void emit_i16x8_min_s(LiftoffRegister dst, LiftoffRegister lhs,

src/wasm/baseline/liftoff-compiler.cc

@@ -2355,6 +2355,12 @@ class LiftoffCompiler {
             &LiftoffAssembler::emit_i8x16_add_saturate_u);
       case wasm::kExprI8x16Sub:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_sub);
+      case wasm::kExprI8x16SubSaturateS:
+        return EmitBinOp<kS128, kS128>(
+            &LiftoffAssembler::emit_i8x16_sub_saturate_s);
+      case wasm::kExprI8x16SubSaturateU:
+        return EmitBinOp<kS128, kS128>(
+            &LiftoffAssembler::emit_i8x16_sub_saturate_u);
       case wasm::kExprI8x16Mul:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_mul);
       case wasm::kExprI8x16MinS:
@@ -2377,6 +2383,12 @@ class LiftoffCompiler {
             &LiftoffAssembler::emit_i16x8_add_saturate_u);
       case wasm::kExprI16x8Sub:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_sub);
+      case wasm::kExprI16x8SubSaturateS:
+        return EmitBinOp<kS128, kS128>(
+            &LiftoffAssembler::emit_i16x8_sub_saturate_s);
+      case wasm::kExprI16x8SubSaturateU:
+        return EmitBinOp<kS128, kS128>(
+            &LiftoffAssembler::emit_i16x8_sub_saturate_u);
       case wasm::kExprI16x8Mul:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_mul);
       case wasm::kExprI16x8MinS:

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

@@ -1994,6 +1994,20 @@ void LiftoffAssembler::emit_i8x16_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                                               rhs);
 }
 
+void LiftoffAssembler::emit_i8x16_sub_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdSub<&Assembler::vpsubsb, &Assembler::psubsb>(this, dst, lhs,
+                                                                rhs);
+}
+
+void LiftoffAssembler::emit_i8x16_sub_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdSub<&Assembler::vpsubusb, &Assembler::psubusb>(this, dst,
+                                                                  lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i8x16_mul(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   static constexpr RegClass tmp_rc = reg_class_for(ValueType::kS128);
@@ -2121,6 +2135,20 @@ void LiftoffAssembler::emit_i16x8_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                                               rhs);
 }
 
+void LiftoffAssembler::emit_i16x8_sub_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdSub<&Assembler::vpsubsw, &Assembler::psubsw>(this, dst, lhs,
+                                                                rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_sub_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdSub<&Assembler::vpsubusw, &Assembler::psubusw>(this, dst,
+                                                                  lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i16x8_mul(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vpmullw, &Assembler::pmullw>(