[wasm-simd][liftoff] Implement s8x16.swizzle

Implemented for ia32, x64, arm, arm64, all in one patch (phew). The
code is simple enough (short paragraph) that are the same as TurboFan
codegen.

Bug: v8:9909
Change-Id: Idbc1cbd58c16e455b1656c2367c8d9db10308e35
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2208610
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Reviewed-by: Clemens Backes <clemensb@chromium.org>
Cr-Commit-Position: refs/heads/master@{#68022}

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

@@ -2180,6 +2180,23 @@ void LiftoffAssembler::LoadTransform(LiftoffRegister dst, Register src_addr,
   }
 }
 
+void LiftoffAssembler::emit_s8x16_swizzle(LiftoffRegister dst,
+                                          LiftoffRegister lhs,
+                                          LiftoffRegister rhs) {
+  UseScratchRegisterScope temps(this);
+
+  NeonListOperand table(liftoff::GetSimd128Register(lhs));
+  if (dst == lhs) {
+    // dst will be overwritten, so keep the table somewhere else.
+    QwNeonRegister tbl = temps.AcquireQ();
+    TurboAssembler::Move(tbl, liftoff::GetSimd128Register(lhs));
+    table = NeonListOperand(tbl);
+  }
+
+  vtbl(dst.low_fp(), table, rhs.low_fp());
+  vtbl(dst.high_fp(), table, rhs.high_fp());
+}
+
 void LiftoffAssembler::emit_f64x2_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   TurboAssembler::Move(dst.low_fp(), src.fp());

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

@@ -1200,6 +1200,12 @@ void LiftoffAssembler::LoadTransform(LiftoffRegister dst, Register src_addr,
   }
 }
 
+void LiftoffAssembler::emit_s8x16_swizzle(LiftoffRegister dst,
+                                          LiftoffRegister lhs,
+                                          LiftoffRegister rhs) {
+  Tbl(dst.fp().V16B(), lhs.fp().V16B(), rhs.fp().V16B());
+}
+
 void LiftoffAssembler::emit_f64x2_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Dup(dst.fp().V2D(), src.fp().D(), 0);

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

@@ -2170,6 +2170,22 @@ void LiftoffAssembler::LoadTransform(LiftoffRegister dst, Register src_addr,
   }
 }
 
+void LiftoffAssembler::emit_s8x16_swizzle(LiftoffRegister dst,
+                                          LiftoffRegister lhs,
+                                          LiftoffRegister rhs) {
+  XMMRegister mask =
+      GetUnusedRegister(kFpReg, LiftoffRegList::ForRegs(lhs, rhs)).fp();
+  // Out-of-range indices should return 0, add 112 (0x70) so that any value > 15
+  // saturates to 128 (top bit set), so pshufb will zero that lane.
+  TurboAssembler::Move(mask, uint32_t{0x70707070});
+  Pshufd(mask, mask, uint8_t{0x0});
+  Paddusb(mask, rhs.fp());
+  if (lhs != dst) {
+    Movaps(dst.fp(), lhs.fp());
+  }
+  Pshufb(dst.fp(), mask);
+}
+
 void LiftoffAssembler::emit_i8x16_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Movd(dst.fp(), src.gp());

src/wasm/baseline/liftoff-assembler.h

@@ -746,7 +746,8 @@ class LiftoffAssembler : public TurboAssembler {
                             Register offset_reg, uint32_t offset_imm,
                             LoadType type, LoadTransformationKind transform,
                             uint32_t* protected_load_pc);
-
+  inline void emit_s8x16_swizzle(LiftoffRegister dst, LiftoffRegister lhs,
+                                 LiftoffRegister rhs);
   inline void emit_i8x16_splat(LiftoffRegister dst, LiftoffRegister src);
   inline void emit_i16x8_splat(LiftoffRegister dst, LiftoffRegister src);
   inline void emit_i32x4_splat(LiftoffRegister dst, LiftoffRegister src);

src/wasm/baseline/liftoff-compiler.cc

@@ -2453,6 +2453,8 @@ class LiftoffCompiler {
       return unsupported(decoder, kSimd, "simd");
     }
     switch (opcode) {
+      case wasm::kExprS8x16Swizzle:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_s8x16_swizzle);
       case wasm::kExprI8x16Splat:
         return EmitUnOp<kI32, kS128>(&LiftoffAssembler::emit_i8x16_splat);
       case wasm::kExprI16x8Splat:

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

@@ -2222,6 +2222,22 @@ void LiftoffAssembler::LoadTransform(LiftoffRegister dst, Register src_addr,
   }
 }
 
+void LiftoffAssembler::emit_s8x16_swizzle(LiftoffRegister dst,
+                                          LiftoffRegister lhs,
+                                          LiftoffRegister rhs) {
+  XMMRegister mask =
+      GetUnusedRegister(kFpReg, LiftoffRegList::ForRegs(lhs, rhs)).fp();
+  // Out-of-range indices should return 0, add 112 (0x70) so that any value > 15
+  // saturates to 128 (top bit set), so pshufb will zero that lane.
+  TurboAssembler::Move(mask, uint32_t{0x70707070});
+  Pshufd(mask, mask, uint8_t{0x0});
+  Paddusb(mask, rhs.fp());
+  if (lhs != dst) {
+    Movaps(dst.fp(), lhs.fp());
+  }
+  Pshufb(dst.fp(), mask);
+}
+
 void LiftoffAssembler::emit_i8x16_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Movd(dst.fp(), src.gp());