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

Bug: v8:9909
Change-Id: Ib744bcbfe3df66d6904421bcbd2581dd6a95b75c
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2116878
Commit-Queue: Fanchen Kong <fanchen.kong@intel.com>
Reviewed-by: Zhi An Ng <zhin@chromium.org>
Reviewed-by: Clemens Backes <clemensb@chromium.org>
Cr-Commit-Position: refs/heads/master@{#66947}

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

@@ -1731,6 +1731,12 @@ void LiftoffAssembler::emit_i16x8_add(LiftoffRegister dst, LiftoffRegister lhs,
        liftoff::GetSimd128Register(rhs.low_fp()));
 }
 
+void LiftoffAssembler::emit_i16x8_add_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8addsaturate_s");
+}
+
 void LiftoffAssembler::emit_i16x8_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   vsub(Neon16, liftoff::GetSimd128Register(dst.low_fp()),
@@ -1743,6 +1749,12 @@ void LiftoffAssembler::emit_i16x8_mul(LiftoffRegister dst, LiftoffRegister lhs,
   bailout(kSimd, "i16x8mul");
 }
 
+void LiftoffAssembler::emit_i16x8_add_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8addsaturate_u");
+}
+
 void LiftoffAssembler::emit_i16x8_extract_lane_u(LiftoffRegister dst,
                                                  LiftoffRegister lhs,
                                                  uint8_t imm_lane_idx) {
@@ -1801,6 +1813,12 @@ void LiftoffAssembler::emit_i8x16_add(LiftoffRegister dst, LiftoffRegister lhs,
        liftoff::GetSimd128Register(rhs.low_fp()));
 }
 
+void LiftoffAssembler::emit_i8x16_add_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16addsaturate_s");
+}
+
 void LiftoffAssembler::emit_i8x16_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   vsub(Neon8, liftoff::GetSimd128Register(dst.low_fp()),
@@ -1813,6 +1831,12 @@ void LiftoffAssembler::emit_i8x16_mul(LiftoffRegister dst, LiftoffRegister lhs,
   bailout(kSimd, "i8x16mul");
 }
 
+void LiftoffAssembler::emit_i8x16_add_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16addsaturate_u");
+}
+
 void LiftoffAssembler::StackCheck(Label* ool_code, Register limit_address) {
   ldr(limit_address, MemOperand(limit_address));
   cmp(sp, limit_address);

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

@@ -1260,6 +1260,12 @@ void LiftoffAssembler::emit_i16x8_add(LiftoffRegister dst, LiftoffRegister lhs,
   Add(dst.fp().V8H(), lhs.fp().V8H(), rhs.fp().V8H());
 }
 
+void LiftoffAssembler::emit_i16x8_add_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8addsaturate_s");
+}
+
 void LiftoffAssembler::emit_i16x8_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   Sub(dst.fp().V8H(), lhs.fp().V8H(), rhs.fp().V8H());
@@ -1270,6 +1276,12 @@ void LiftoffAssembler::emit_i16x8_mul(LiftoffRegister dst, LiftoffRegister lhs,
   bailout(kSimd, "i16x8mul");
 }
 
+void LiftoffAssembler::emit_i16x8_add_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8addsaturate_u");
+}
+
 void LiftoffAssembler::emit_i8x16_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Dup(dst.fp().V16B(), src.gp().W());
@@ -1302,6 +1314,12 @@ void LiftoffAssembler::emit_i8x16_add(LiftoffRegister dst, LiftoffRegister lhs,
   Add(dst.fp().V16B(), lhs.fp().V16B(), rhs.fp().V16B());
 }
 
+void LiftoffAssembler::emit_i8x16_add_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16addsaturate_s");
+}
+
 void LiftoffAssembler::emit_i8x16_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   Sub(dst.fp().V16B(), lhs.fp().V16B(), rhs.fp().V16B());
@@ -1312,6 +1330,12 @@ void LiftoffAssembler::emit_i8x16_mul(LiftoffRegister dst, LiftoffRegister lhs,
   bailout(kSimd, "i8x16mul");
 }
 
+void LiftoffAssembler::emit_i8x16_add_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16addsaturate_u");
+}
+
 void LiftoffAssembler::StackCheck(Label* ool_code, Register limit_address) {
   Ldr(limit_address, MemOperand(limit_address));
   Cmp(sp, limit_address);

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

@@ -2239,6 +2239,13 @@ void LiftoffAssembler::emit_i16x8_add(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i16x8_add_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddsw, &Assembler::paddsw>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i16x8_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   liftoff::EmitSimdSub<&Assembler::vpsubw, &Assembler::psubw>(this, dst, lhs,
@@ -2251,6 +2258,13 @@ void LiftoffAssembler::emit_i16x8_mul(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i16x8_add_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddusw, &Assembler::paddusw>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i8x16_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Movd(dst.fp(), src.gp());
@@ -2291,6 +2305,13 @@ void LiftoffAssembler::emit_i8x16_add(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i8x16_add_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddsb, &Assembler::paddsb>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i8x16_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   liftoff::EmitSimdSub<&Assembler::vpsubb, &Assembler::psubb>(this, dst, lhs,
@@ -2359,6 +2380,13 @@ void LiftoffAssembler::emit_i8x16_mul(LiftoffRegister dst, LiftoffRegister lhs,
   }
 }
 
+void LiftoffAssembler::emit_i8x16_add_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddusb, &Assembler::paddusb>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::StackCheck(Label* ool_code, Register limit_address) {
   cmp(esp, Operand(limit_address, 0));
   j(below_equal, ool_code);

src/wasm/baseline/liftoff-assembler.h

@@ -771,10 +771,16 @@ class LiftoffAssembler : public TurboAssembler {
                                       uint8_t imm_lane_idx);
   inline void emit_i16x8_add(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
+  inline void emit_i16x8_add_saturate_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs);
   inline void emit_i16x8_sub(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
   inline void emit_i16x8_mul(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
+  inline void emit_i16x8_add_saturate_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs);
   inline void emit_i8x16_splat(LiftoffRegister dst, LiftoffRegister src);
   inline void emit_i8x16_extract_lane_u(LiftoffRegister dst,
                                         LiftoffRegister lhs,
@@ -787,10 +793,16 @@ class LiftoffAssembler : public TurboAssembler {
                                       uint8_t imm_lane_idx);
   inline void emit_i8x16_add(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
+  inline void emit_i8x16_add_saturate_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs);
   inline void emit_i8x16_sub(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
   inline void emit_i8x16_mul(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
+  inline void emit_i8x16_add_saturate_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs);
 
   inline void StackCheck(Label* ool_code, Register limit_address);
 

src/wasm/baseline/liftoff-compiler.cc

@@ -2320,18 +2320,30 @@ class LiftoffCompiler {
         return EmitUnOp<kI32, kS128>(&LiftoffAssembler::emit_i16x8_splat);
       case wasm::kExprI16x8Add:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_add);
+      case wasm::kExprI16x8AddSaturateS:
+        return EmitBinOp<kS128, kS128>(
+            &LiftoffAssembler::emit_i16x8_add_saturate_s);
       case wasm::kExprI16x8Sub:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_sub);
       case wasm::kExprI16x8Mul:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_mul);
+      case wasm::kExprI16x8AddSaturateU:
+        return EmitBinOp<kS128, kS128>(
+            &LiftoffAssembler::emit_i16x8_add_saturate_u);
       case wasm::kExprI8x16Splat:
         return EmitUnOp<kI32, kS128>(&LiftoffAssembler::emit_i8x16_splat);
       case wasm::kExprI8x16Add:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_add);
+      case wasm::kExprI8x16AddSaturateS:
+        return EmitBinOp<kS128, kS128>(
+            &LiftoffAssembler::emit_i8x16_add_saturate_s);
       case wasm::kExprI8x16Sub:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_sub);
       case wasm::kExprI8x16Mul:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_mul);
+      case wasm::kExprI8x16AddSaturateU:
+        return EmitBinOp<kS128, kS128>(
+            &LiftoffAssembler::emit_i8x16_add_saturate_u);
       default:
         unsupported(decoder, kSimd, "simd");
     }

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

@@ -2170,6 +2170,13 @@ void LiftoffAssembler::emit_i16x8_add(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i16x8_add_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddsw, &Assembler::paddsw>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i16x8_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   liftoff::EmitSimdSub<&Assembler::vpsubw, &Assembler::psubw>(this, dst, lhs,
@@ -2182,6 +2189,13 @@ void LiftoffAssembler::emit_i16x8_mul(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i16x8_add_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddusw, &Assembler::paddusw>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i8x16_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Movd(dst.fp(), src.gp());
@@ -2222,6 +2236,13 @@ void LiftoffAssembler::emit_i8x16_add(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i8x16_add_saturate_s(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddsb, &Assembler::paddsb>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i8x16_sub(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   liftoff::EmitSimdSub<&Assembler::vpsubb, &Assembler::psubb>(this, dst, lhs,
@@ -2290,6 +2311,13 @@ void LiftoffAssembler::emit_i8x16_mul(LiftoffRegister dst, LiftoffRegister lhs,
   }
 }
 
+void LiftoffAssembler::emit_i8x16_add_saturate_u(LiftoffRegister dst,
+                                                 LiftoffRegister lhs,
+                                                 LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddusb, &Assembler::paddusb>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::StackCheck(Label* ool_code, Register limit_address) {
   cmpq(rsp, Operand(limit_address, 0));
   j(below_equal, ool_code);