[wasm-simd][liftoff] Implement i32x4 shr_s shr_u

Implemented for x64, ia32, arm64, arm.

Bug: v8:9909
Change-Id: Ib1479525d44c29b2ba02de31d655da2cde21d9fc
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2202356
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Reviewed-by: Clemens Backes <clemensb@chromium.org>
Cr-Commit-Position: refs/heads/master@{#67835}

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

@@ -362,15 +362,19 @@ inline void EmitSimdShift(LiftoffAssembler* assm, LiftoffRegister dst,
              liftoff::GetSimd128Register(lhs), tmp);
 }
 
-template <NeonDataType dt>
-inline void EmitSimdShiftRightImmediate(LiftoffAssembler* assm,
-                                        LiftoffRegister dst,
-                                        LiftoffRegister lhs, int32_t rhs) {
+template <ShiftDirection dir, NeonDataType dt>
+inline void EmitSimdShiftImmediate(LiftoffAssembler* assm, LiftoffRegister dst,
+                                   LiftoffRegister lhs, int32_t rhs) {
   // vshr by 0 is not allowed, so check for it, and only move if dst != lhs.
   int32_t shift = rhs & MaskFromNeonDataType(dt);
   if (shift) {
-    assm->vshr(dt, liftoff::GetSimd128Register(dst),
-               liftoff::GetSimd128Register(lhs), shift);
+    if (dir == kLeft) {
+      assm->vshl(dt, liftoff::GetSimd128Register(dst),
+                 liftoff::GetSimd128Register(lhs), shift);
+    } else {
+      assm->vshr(dt, liftoff::GetSimd128Register(dst),
+                 liftoff::GetSimd128Register(lhs), shift);
+    }
   } else if (dst != lhs) {
     assm->vmov(liftoff::GetSimd128Register(dst),
                liftoff::GetSimd128Register(lhs));
@@ -2338,7 +2342,8 @@ void LiftoffAssembler::emit_i64x2_shr_s(LiftoffRegister dst,
 
 void LiftoffAssembler::emit_i64x2_shri_s(LiftoffRegister dst,
                                          LiftoffRegister lhs, int32_t rhs) {
-  liftoff::EmitSimdShiftRightImmediate<NeonS64>(this, dst, lhs, rhs);
+  liftoff::EmitSimdShiftImmediate<liftoff::kRight, NeonS64>(this, dst, lhs,
+                                                            rhs);
 }
 
 void LiftoffAssembler::emit_i64x2_shr_u(LiftoffRegister dst,
@@ -2349,7 +2354,8 @@ void LiftoffAssembler::emit_i64x2_shr_u(LiftoffRegister dst,
 
 void LiftoffAssembler::emit_i64x2_shri_u(LiftoffRegister dst,
                                          LiftoffRegister lhs, int32_t rhs) {
-  liftoff::EmitSimdShiftRightImmediate<NeonU64>(this, dst, lhs, rhs);
+  liftoff::EmitSimdShiftImmediate<liftoff::kRight, NeonU64>(this, dst, lhs,
+                                                            rhs);
 }
 
 void LiftoffAssembler::emit_i64x2_add(LiftoffRegister dst, LiftoffRegister lhs,
@@ -2445,6 +2451,30 @@ void LiftoffAssembler::emit_i32x4_shli(LiftoffRegister dst, LiftoffRegister lhs,
        liftoff::GetSimd128Register(lhs), rhs & 31);
 }
 
+void LiftoffAssembler::emit_i32x4_shr_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShift<liftoff::kRight, NeonS32, Neon32>(this, dst, lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i32x4_shri_s(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftImmediate<liftoff::kRight, NeonS32>(this, dst, lhs,
+                                                            rhs);
+}
+
+void LiftoffAssembler::emit_i32x4_shr_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShift<liftoff::kRight, NeonU32, Neon32>(this, dst, lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i32x4_shri_u(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftImmediate<liftoff::kRight, NeonU32>(this, dst, lhs,
+                                                            rhs);
+}
+
 void LiftoffAssembler::emit_i32x4_add(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   vadd(Neon32, liftoff::GetSimd128Register(dst),

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

@@ -1412,6 +1412,35 @@ void LiftoffAssembler::emit_i32x4_shli(LiftoffRegister dst, LiftoffRegister lhs,
   Shl(dst.fp().V4S(), lhs.fp().V4S(), rhs & 31);
 }
 
+void LiftoffAssembler::emit_i32x4_shr_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShift<liftoff::ShiftDirection::kRight,
+                         liftoff::ShiftSign::kSigned>(
+      this, dst.fp().V4S(), lhs.fp().V4S(), rhs.gp(), kFormat4S);
+}
+
+void LiftoffAssembler::emit_i32x4_shri_s(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftRightImmediate<kFormat4S, liftoff::ShiftSign::kSigned>(
+      this, dst.fp().V4S(), lhs.fp().V4S(), rhs);
+}
+
+void LiftoffAssembler::emit_i32x4_shr_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShift<liftoff::ShiftDirection::kRight,
+                         liftoff::ShiftSign::kUnsigned>(
+      this, dst.fp().V4S(), lhs.fp().V4S(), rhs.gp(), kFormat4S);
+}
+
+void LiftoffAssembler::emit_i32x4_shri_u(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftRightImmediate<kFormat4S,
+                                       liftoff::ShiftSign::kUnsigned>(
+      this, dst.fp().V4S(), lhs.fp().V4S(), rhs);
+}
+
 void LiftoffAssembler::emit_i32x4_add(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   Add(dst.fp().V4S(), lhs.fp().V4S(), rhs.fp().V4S());

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

@@ -2654,6 +2654,32 @@ void LiftoffAssembler::emit_i32x4_shli(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i32x4_shr_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShiftOp<&Assembler::vpsrad, &Assembler::psrad, 5>(this, dst,
+                                                                     lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i32x4_shri_s(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftOpImm<&Assembler::vpsrad, &Assembler::psrad, 5>(
+      this, dst, lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i32x4_shr_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShiftOp<&Assembler::vpsrld, &Assembler::psrld, 5>(this, dst,
+                                                                     lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i32x4_shri_u(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftOpImm<&Assembler::vpsrld, &Assembler::psrld, 5>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i32x4_add(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddd, &Assembler::paddd>(

src/wasm/baseline/liftoff-assembler.h

@@ -876,6 +876,14 @@ class LiftoffAssembler : public TurboAssembler {
                              LiftoffRegister rhs);
   inline void emit_i32x4_shli(LiftoffRegister dst, LiftoffRegister lhs,
                               int32_t rhs);
+  inline void emit_i32x4_shr_s(LiftoffRegister dst, LiftoffRegister lhs,
+                               LiftoffRegister rhs);
+  inline void emit_i32x4_shri_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                int32_t rhs);
+  inline void emit_i32x4_shr_u(LiftoffRegister dst, LiftoffRegister lhs,
+                               LiftoffRegister rhs);
+  inline void emit_i32x4_shri_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                int32_t rhs);
   inline void emit_i32x4_add(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
   inline void emit_i32x4_sub(LiftoffRegister dst, LiftoffRegister lhs,

src/wasm/baseline/liftoff-compiler.cc

@@ -2566,6 +2566,12 @@ class LiftoffCompiler {
       case wasm::kExprI32x4Shl:
         return EmitSimdShiftOp(&LiftoffAssembler::emit_i32x4_shl,
                                &LiftoffAssembler::emit_i32x4_shli);
+      case wasm::kExprI32x4ShrS:
+        return EmitSimdShiftOp(&LiftoffAssembler::emit_i32x4_shr_s,
+                               &LiftoffAssembler::emit_i32x4_shri_s);
+      case wasm::kExprI32x4ShrU:
+        return EmitSimdShiftOp(&LiftoffAssembler::emit_i32x4_shr_u,
+                               &LiftoffAssembler::emit_i32x4_shri_u);
       case wasm::kExprI32x4Add:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_add);
       case wasm::kExprI32x4Sub:

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

@@ -2683,6 +2683,32 @@ void LiftoffAssembler::emit_i32x4_shli(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i32x4_shr_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShiftOp<&Assembler::vpsrad, &Assembler::psrad, 5>(this, dst,
+                                                                     lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i32x4_shri_s(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftOpImm<&Assembler::vpsrad, &Assembler::psrad, 5>(
+      this, dst, lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i32x4_shr_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdShiftOp<&Assembler::vpsrld, &Assembler::psrld, 5>(this, dst,
+                                                                     lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i32x4_shri_u(LiftoffRegister dst,
+                                         LiftoffRegister lhs, int32_t rhs) {
+  liftoff::EmitSimdShiftOpImm<&Assembler::vpsrld, &Assembler::psrld, 5>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::emit_i32x4_add(LiftoffRegister dst, LiftoffRegister lhs,
                                       LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vpaddd, &Assembler::paddd>(