[wasm-simd][liftoff] Implement min of i8x16, i16x8, i32x4 on x64 and ia32

Bug: v8:9909
Change-Id: I11fae10593d135e007866411c61076300ac4ae7d
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2129176Reviewed-by: Clemens Backes <clemensb@chromium.org>
Reviewed-by: Zhi An Ng <zhin@chromium.org>
Commit-Queue: Zhiguo Zhou <zhiguo.zhou@intel.com>
Cr-Commit-Position: refs/heads/master@{#66972}

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

@@ -1763,6 +1763,18 @@ void LiftoffAssembler::emit_i32x4_mul(LiftoffRegister dst, LiftoffRegister lhs,
        liftoff::GetSimd128Register(rhs.low_fp()));
 }
 
+void LiftoffAssembler::emit_i32x4_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i32x4_min_s");
+}
+
+void LiftoffAssembler::emit_i32x4_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i32x4_min_u");
+}
+
 void LiftoffAssembler::emit_i16x8_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   vdup(Neon16, liftoff::GetSimd128Register(dst.low_fp()), src.gp());
@@ -1801,6 +1813,18 @@ void LiftoffAssembler::emit_i16x8_add_saturate_u(LiftoffRegister dst,
   bailout(kSimd, "i16x8addsaturate_u");
 }
 
+void LiftoffAssembler::emit_i16x8_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8_min_s");
+}
+
+void LiftoffAssembler::emit_i16x8_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8_min_u");
+}
+
 void LiftoffAssembler::emit_i16x8_extract_lane_u(LiftoffRegister dst,
                                                  LiftoffRegister lhs,
                                                  uint8_t imm_lane_idx) {
@@ -1885,6 +1909,18 @@ void LiftoffAssembler::emit_i8x16_add_saturate_u(LiftoffRegister dst,
   bailout(kSimd, "i8x16addsaturate_u");
 }
 
+void LiftoffAssembler::emit_i8x16_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16_min_s");
+}
+
+void LiftoffAssembler::emit_i8x16_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16_min_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

@@ -1244,6 +1244,18 @@ void LiftoffAssembler::emit_i32x4_mul(LiftoffRegister dst, LiftoffRegister lhs,
   Mul(dst.fp().V4S(), lhs.fp().V4S(), rhs.fp().V4S());
 }
 
+void LiftoffAssembler::emit_i32x4_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i32x4_min_s");
+}
+
+void LiftoffAssembler::emit_i32x4_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i32x4_min_u");
+}
+
 void LiftoffAssembler::emit_i16x8_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Dup(dst.fp().V8H(), src.gp().W());
@@ -1298,6 +1310,18 @@ void LiftoffAssembler::emit_i16x8_add_saturate_u(LiftoffRegister dst,
   bailout(kSimd, "i16x8addsaturate_u");
 }
 
+void LiftoffAssembler::emit_i16x8_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8_min_s");
+}
+
+void LiftoffAssembler::emit_i16x8_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8_min_u");
+}
+
 void LiftoffAssembler::emit_i8x16_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Dup(dst.fp().V16B(), src.gp().W());
@@ -1352,6 +1376,18 @@ void LiftoffAssembler::emit_i8x16_add_saturate_u(LiftoffRegister dst,
   bailout(kSimd, "i8x16addsaturate_u");
 }
 
+void LiftoffAssembler::emit_i8x16_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16_min_s");
+}
+
+void LiftoffAssembler::emit_i8x16_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16_min_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

@@ -2200,6 +2200,20 @@ void LiftoffAssembler::emit_i32x4_mul(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs, base::Optional<CpuFeature>(SSE4_1));
 }
 
+void LiftoffAssembler::emit_i32x4_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsd, &Assembler::pminsd>(
+      this, dst, lhs, rhs, base::Optional<CpuFeature>(SSE4_1));
+}
+
+void LiftoffAssembler::emit_i32x4_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminud, &Assembler::pminud>(
+      this, dst, lhs, rhs, base::Optional<CpuFeature>(SSE4_1));
+}
+
 void LiftoffAssembler::emit_i16x8_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Movd(dst.fp(), src.gp());
@@ -2265,6 +2279,20 @@ void LiftoffAssembler::emit_i16x8_add_saturate_u(LiftoffRegister dst,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i16x8_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsw, &Assembler::pminsw>(
+      this, dst, lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminuw, &Assembler::pminuw>(
+      this, dst, lhs, rhs, base::Optional<CpuFeature>(SSE4_1));
+}
+
 void LiftoffAssembler::emit_i8x16_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Movd(dst.fp(), src.gp());
@@ -2387,6 +2415,20 @@ void LiftoffAssembler::emit_i8x16_add_saturate_u(LiftoffRegister dst,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i8x16_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsb, &Assembler::pminsb>(
+      this, dst, lhs, rhs, base::Optional<CpuFeature>(SSE4_1));
+}
+
+void LiftoffAssembler::emit_i8x16_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminub, &Assembler::pminub>(
+      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

@@ -759,6 +759,10 @@ class LiftoffAssembler : public TurboAssembler {
                              LiftoffRegister rhs);
   inline void emit_i32x4_mul(LiftoffRegister dst, LiftoffRegister lhs,
                              LiftoffRegister rhs);
+  inline void emit_i32x4_min_s(LiftoffRegister dst, LiftoffRegister lhs,
+                               LiftoffRegister rhs);
+  inline void emit_i32x4_min_u(LiftoffRegister dst, LiftoffRegister lhs,
+                               LiftoffRegister rhs);
   inline void emit_i16x8_splat(LiftoffRegister dst, LiftoffRegister src);
   inline void emit_i16x8_extract_lane_u(LiftoffRegister dst,
                                         LiftoffRegister lhs,
@@ -781,6 +785,10 @@ class LiftoffAssembler : public TurboAssembler {
   inline void emit_i16x8_add_saturate_u(LiftoffRegister dst,
                                         LiftoffRegister lhs,
                                         LiftoffRegister rhs);
+  inline void emit_i16x8_min_s(LiftoffRegister dst, LiftoffRegister lhs,
+                               LiftoffRegister rhs);
+  inline void emit_i16x8_min_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,
@@ -803,6 +811,10 @@ class LiftoffAssembler : public TurboAssembler {
   inline void emit_i8x16_add_saturate_u(LiftoffRegister dst,
                                         LiftoffRegister lhs,
                                         LiftoffRegister rhs);
+  inline void emit_i8x16_min_s(LiftoffRegister dst, LiftoffRegister lhs,
+                               LiftoffRegister rhs);
+  inline void emit_i8x16_min_u(LiftoffRegister dst, LiftoffRegister lhs,
+                               LiftoffRegister rhs);
 
   inline void StackCheck(Label* ool_code, Register limit_address);
 

src/wasm/baseline/liftoff-compiler.cc

@@ -2316,6 +2316,10 @@ class LiftoffCompiler {
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_sub);
       case wasm::kExprI32x4Mul:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_mul);
+      case wasm::kExprI32x4MinS:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_min_s);
+      case wasm::kExprI32x4MinU:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_min_u);
       case wasm::kExprI16x8Splat:
         return EmitUnOp<kI32, kS128>(&LiftoffAssembler::emit_i16x8_splat);
       case wasm::kExprI16x8Add:
@@ -2330,6 +2334,10 @@ class LiftoffCompiler {
       case wasm::kExprI16x8AddSaturateU:
         return EmitBinOp<kS128, kS128>(
             &LiftoffAssembler::emit_i16x8_add_saturate_u);
+      case wasm::kExprI16x8MinS:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_min_s);
+      case wasm::kExprI16x8MinU:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_min_u);
       case wasm::kExprI8x16Splat:
         return EmitUnOp<kI32, kS128>(&LiftoffAssembler::emit_i8x16_splat);
       case wasm::kExprI8x16Add:
@@ -2344,6 +2352,10 @@ class LiftoffCompiler {
       case wasm::kExprI8x16AddSaturateU:
         return EmitBinOp<kS128, kS128>(
             &LiftoffAssembler::emit_i8x16_add_saturate_u);
+      case wasm::kExprI8x16MinS:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_min_s);
+      case wasm::kExprI8x16MinU:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_min_u);
       default:
         unsupported(decoder, kSimd, "simd");
     }

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

@@ -2131,6 +2131,20 @@ void LiftoffAssembler::emit_i32x4_mul(LiftoffRegister dst, LiftoffRegister lhs,
       this, dst, lhs, rhs, base::Optional<CpuFeature>(SSE4_1));
 }
 
+void LiftoffAssembler::emit_i32x4_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsd, &Assembler::pminsd>(
+      this, dst, lhs, rhs, base::Optional<CpuFeature>(SSE4_1));
+}
+
+void LiftoffAssembler::emit_i32x4_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminud, &Assembler::pminud>(
+      this, dst, lhs, rhs, base::Optional<CpuFeature>(SSE4_1));
+}
+
 void LiftoffAssembler::emit_i16x8_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Movd(dst.fp(), src.gp());
@@ -2196,6 +2210,20 @@ void LiftoffAssembler::emit_i16x8_add_saturate_u(LiftoffRegister dst,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i16x8_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsw, &Assembler::pminsw>(
+      this, dst, lhs, rhs);
+}
+
+void LiftoffAssembler::emit_i16x8_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminuw, &Assembler::pminuw>(
+      this, dst, lhs, rhs, base::Optional<CpuFeature>(SSE4_1));
+}
+
 void LiftoffAssembler::emit_i8x16_splat(LiftoffRegister dst,
                                         LiftoffRegister src) {
   Movd(dst.fp(), src.gp());
@@ -2318,6 +2346,20 @@ void LiftoffAssembler::emit_i8x16_add_saturate_u(LiftoffRegister dst,
       this, dst, lhs, rhs);
 }
 
+void LiftoffAssembler::emit_i8x16_min_s(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsb, &Assembler::pminsb>(
+      this, dst, lhs, rhs, base::Optional<CpuFeature>(SSE4_1));
+}
+
+void LiftoffAssembler::emit_i8x16_min_u(LiftoffRegister dst,
+                                        LiftoffRegister lhs,
+                                        LiftoffRegister rhs) {
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminub, &Assembler::pminub>(
+      this, dst, lhs, rhs);
+}
+
 void LiftoffAssembler::StackCheck(Label* ool_code, Register limit_address) {
   cmpq(rsp, Operand(limit_address, 0));
   j(below_equal, ool_code);