[wasm-simd][liftoff] Implement i16x8 shr

Implement for x64, ia32, arm64, and arm.

Bug: v8:9909
Change-Id: Ia9499b211e915f8763b4f593824099391dca7e1f
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2202718
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Reviewed-by: Clemens Backes <clemensb@chromium.org>
Cr-Commit-Position: refs/heads/master@{#67836}

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

@@ -2543,6 +2543,30 @@ void LiftoffAssembler::emit_i16x8_shli(LiftoffRegister dst, LiftoffRegister lhs,
        liftoff::GetSimd128Register(lhs), rhs & 15);
 }
 
+void LiftoffAssembler::emit_i16x8_shr_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShift<liftoff::kRight, NeonS16, Neon16>(this, dst, lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_shri_s(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftImmediate<liftoff::kRight, NeonS16>(this, dst, lhs,
+                                                            rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_shr_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShift<liftoff::kRight, NeonU16, Neon16>(this, dst, lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_shri_u(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftImmediate<liftoff::kRight, NeonU16>(this, dst, lhs,
+                                                            rhs);
+}
+
 void LiftoffAssembler::emit_i16x8_add(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   vadd(Neon16, liftoff::GetSimd128Register(dst),

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

@@ -1523,6 +1523,35 @@ void LiftoffAssembler::emit_i16x8_shli(LiftoffRegister dst, LiftoffRegister lhs,
   Shl(dst.fp().V8H(), lhs.fp().V8H(), rhs & 15);
 }
 
+void LiftoffAssembler::emit_i16x8_shr_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShift<liftoff::ShiftDirection::kRight,
+                         liftoff::ShiftSign::kSigned>(
+      this, dst.fp().V8H(), lhs.fp().V8H(), rhs.gp(), kFormat8H);
+}
+
+void LiftoffAssembler::emit_i16x8_shri_s(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftRightImmediate<kFormat8H, liftoff::ShiftSign::kSigned>(
+      this, dst.fp().V8H(), lhs.fp().V8H(), rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_shr_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShift<liftoff::ShiftDirection::kRight,
+                         liftoff::ShiftSign::kUnsigned>(
+      this, dst.fp().V8H(), lhs.fp().V8H(), rhs.gp(), kFormat8H);
+}
+
+void LiftoffAssembler::emit_i16x8_shri_u(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftRightImmediate<kFormat8H,
+                                       liftoff::ShiftSign::kUnsigned>(
+      this, dst.fp().V8H(), lhs.fp().V8H(), rhs);
+}
+
 void LiftoffAssembler::emit_i16x8_add(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   Add(dst.fp().V8H(), lhs.fp().V8H(), rhs.fp().V8H());

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

@@ -2556,6 +2556,32 @@ void LiftoffAssembler::emit_i16x8_shli(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i16x8_shr_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShiftOp<&Assembler::vpsraw, &Assembler::psraw, 4>(this, dst,
+                                                                     lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_shri_s(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftOpImm<&Assembler::vpsraw, &Assembler::psraw, 4>(
+      this, dst, lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_shr_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShiftOp<&Assembler::vpsrlw, &Assembler::psrlw, 4>(this, dst,
+                                                                     lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_shri_u(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftOpImm<&Assembler::vpsrlw, &Assembler::psrlw, 4>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i16x8_add(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddw, &Assembler::paddw>(

src/wasm/baseline/liftoff-assembler.h

@@ -845,6 +845,14 @@ class LiftoffAssembler : public TurboAssembler {
                              LiftoffRegister rhs);
   inline void emit_i16x8_shli(LiftoffRegister dst, LiftoffRegister lhs,
                               int32_t rhs);
+  inline void emit_i16x8_shr_s(LiftoffRegister dst, LiftoffRegister lhs,
+                               LiftoffRegister rhs);
+  inline void emit_i16x8_shri_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                int32_t rhs);
+  inline void emit_i16x8_shr_u(LiftoffRegister dst, LiftoffRegister lhs,
+                               LiftoffRegister rhs);
+  inline void emit_i16x8_shri_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                int32_t rhs);
   inline void emit_i16x8_add(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
   inline void emit_i16x8_add_saturate_s(LiftoffRegister dst,

src/wasm/baseline/liftoff-compiler.cc

@@ -2535,6 +2535,12 @@ class LiftoffCompiler {
       case wasm::kExprI16x8Shl:
         return EmitSimdShiftOp(&LiftoffAssembler::emit_i16x8_shl,
                                &LiftoffAssembler::emit_i16x8_shli);
+      case wasm::kExprI16x8ShrS:
+        return EmitSimdShiftOp(&LiftoffAssembler::emit_i16x8_shr_s,
+                               &LiftoffAssembler::emit_i16x8_shri_s);
+      case wasm::kExprI16x8ShrU:
+        return EmitSimdShiftOp(&LiftoffAssembler::emit_i16x8_shr_u,
+                               &LiftoffAssembler::emit_i16x8_shri_u);
       case wasm::kExprI16x8Add:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_add);
       case wasm::kExprI16x8AddSaturateS:

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

@@ -2585,6 +2585,32 @@ void LiftoffAssembler::emit_i16x8_shli(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i16x8_shr_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShiftOp<&Assembler::vpsraw, &Assembler::psraw, 4>(this, dst,
+                                                                     lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_shri_s(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftOpImm<&Assembler::vpsraw, &Assembler::psraw, 4>(
+      this, dst, lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_shr_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShiftOp<&Assembler::vpsrlw, &Assembler::psrlw, 4>(this, dst,
+                                                                     lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_shri_u(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftOpImm<&Assembler::vpsrlw, &Assembler::psrlw, 4>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i16x8_add(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddw, &Assembler::paddw>(