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

Bug: v8:9909
Change-Id: I066621f12ce594280f6f43bc526988e55729eefa
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2162366Reviewed-by: Zhi An Ng <zhin@chromium.org>
Commit-Queue: Zhiguo Zhou <zhiguo.zhou@intel.com>
Cr-Commit-Position: refs/heads/master@{#67411}

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

@@ -2432,6 +2432,16 @@ void LiftoffAssembler::emit_i8x16_ne(LiftoffRegister dst, LiftoffRegister lhs,
   vmvn(liftoff::GetSimd128Register(dst), liftoff::GetSimd128Register(dst));
 }
 
+void LiftoffAssembler::emit_i8x16_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16ge_s");
+}
+
+void LiftoffAssembler::emit_i8x16_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16ge_u");
+}
+
 void LiftoffAssembler::emit_i16x8_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   vceq(Neon16, liftoff::GetSimd128Register(dst),
@@ -2445,6 +2455,16 @@ void LiftoffAssembler::emit_i16x8_ne(LiftoffRegister dst, LiftoffRegister lhs,
   vmvn(liftoff::GetSimd128Register(dst), liftoff::GetSimd128Register(dst));
 }
 
+void LiftoffAssembler::emit_i16x8_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8ge_s");
+}
+
+void LiftoffAssembler::emit_i16x8_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8ge_u");
+}
+
 void LiftoffAssembler::emit_i32x4_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   vceq(Neon32, liftoff::GetSimd128Register(dst),
@@ -2458,6 +2478,16 @@ void LiftoffAssembler::emit_i32x4_ne(LiftoffRegister dst, LiftoffRegister lhs,
   vmvn(liftoff::GetSimd128Register(dst), liftoff::GetSimd128Register(dst));
 }
 
+void LiftoffAssembler::emit_i32x4_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i32x4ge_s");
+}
+
+void LiftoffAssembler::emit_i32x4_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i32x4ge_u");
+}
+
 void LiftoffAssembler::emit_f32x4_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   vceq(liftoff::GetSimd128Register(dst), liftoff::GetSimd128Register(lhs),

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

@@ -1540,6 +1540,16 @@ void LiftoffAssembler::emit_i8x16_ne(LiftoffRegister dst, LiftoffRegister lhs,
   Mvn(dst.fp().V16B(), dst.fp().V16B());
 }
 
+void LiftoffAssembler::emit_i8x16_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16ge_s");
+}
+
+void LiftoffAssembler::emit_i8x16_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i8x16ge_u");
+}
+
 void LiftoffAssembler::emit_i16x8_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   Cmeq(dst.fp().V8H(), lhs.fp().V8H(), rhs.fp().V8H());
@@ -1551,6 +1561,16 @@ void LiftoffAssembler::emit_i16x8_ne(LiftoffRegister dst, LiftoffRegister lhs,
   Mvn(dst.fp().V8H(), dst.fp().V8H());
 }
 
+void LiftoffAssembler::emit_i16x8_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8ge_s");
+}
+
+void LiftoffAssembler::emit_i16x8_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i16x8ge_u");
+}
+
 void LiftoffAssembler::emit_i32x4_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   Cmeq(dst.fp().V4S(), lhs.fp().V4S(), rhs.fp().V4S());
@@ -1562,6 +1582,16 @@ void LiftoffAssembler::emit_i32x4_ne(LiftoffRegister dst, LiftoffRegister lhs,
   Mvn(dst.fp().V4S(), dst.fp().V4S());
 }
 
+void LiftoffAssembler::emit_i32x4_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i32x4ge_s");
+}
+
+void LiftoffAssembler::emit_i32x4_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  bailout(kSimd, "i32x4ge_u");
+}
+
 void LiftoffAssembler::emit_f32x4_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   Fcmeq(dst.fp().V4S(), lhs.fp().V4S(), rhs.fp().V4S());

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

@@ -2033,6 +2033,30 @@ void LiftoffAssembler::emit_i8x16_ne(LiftoffRegister dst, LiftoffRegister lhs,
   Pxor(dst.fp(), liftoff::kScratchDoubleReg);
 }
 
+void LiftoffAssembler::emit_i8x16_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(liftoff::kScratchDoubleReg, rhs.fp());
+    ref = liftoff::kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsb, &Assembler::pminsb>(
+      this, dst, lhs, rhs, SSE4_1);
+  Pcmpeqb(dst.fp(), ref);
+}
+
+void LiftoffAssembler::emit_i8x16_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(liftoff::kScratchDoubleReg, rhs.fp());
+    ref = liftoff::kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminub, &Assembler::pminub>(
+      this, dst, lhs, rhs);
+  Pcmpeqb(dst.fp(), ref);
+}
+
 void LiftoffAssembler::emit_i16x8_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vpcmpeqw, &Assembler::pcmpeqw>(
@@ -2047,6 +2071,30 @@ void LiftoffAssembler::emit_i16x8_ne(LiftoffRegister dst, LiftoffRegister lhs,
   Pxor(dst.fp(), liftoff::kScratchDoubleReg);
 }
 
+void LiftoffAssembler::emit_i16x8_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(liftoff::kScratchDoubleReg, rhs.fp());
+    ref = liftoff::kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsw, &Assembler::pminsw>(
+      this, dst, lhs, rhs);
+  Pcmpeqw(dst.fp(), ref);
+}
+
+void LiftoffAssembler::emit_i16x8_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(liftoff::kScratchDoubleReg, rhs.fp());
+    ref = liftoff::kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminuw, &Assembler::pminuw>(
+      this, dst, lhs, rhs, SSE4_1);
+  Pcmpeqw(dst.fp(), ref);
+}
+
 void LiftoffAssembler::emit_i32x4_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vpcmpeqd, &Assembler::pcmpeqd>(
@@ -2061,6 +2109,30 @@ void LiftoffAssembler::emit_i32x4_ne(LiftoffRegister dst, LiftoffRegister lhs,
   Pxor(dst.fp(), liftoff::kScratchDoubleReg);
 }
 
+void LiftoffAssembler::emit_i32x4_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(liftoff::kScratchDoubleReg, rhs.fp());
+    ref = liftoff::kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsd, &Assembler::pminsd>(
+      this, dst, lhs, rhs, SSE4_1);
+  Pcmpeqd(dst.fp(), ref);
+}
+
+void LiftoffAssembler::emit_i32x4_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(liftoff::kScratchDoubleReg, rhs.fp());
+    ref = liftoff::kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminud, &Assembler::pminud>(
+      this, dst, lhs, rhs, SSE4_1);
+  Pcmpeqd(dst.fp(), ref);
+}
+
 void LiftoffAssembler::emit_f32x4_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vcmpeqps, &Assembler::cmpeqps>(

src/wasm/baseline/liftoff-assembler.h

@@ -733,14 +733,26 @@ class LiftoffAssembler : public TurboAssembler {
                             LiftoffRegister rhs);
   inline void emit_i8x16_ne(LiftoffRegister dst, LiftoffRegister lhs,
                             LiftoffRegister rhs);
+  inline void emit_i8x16_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                              LiftoffRegister rhs);
+  inline void emit_i8x16_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                              LiftoffRegister rhs);
   inline void emit_i16x8_eq(LiftoffRegister dst, LiftoffRegister lhs,
                             LiftoffRegister rhs);
   inline void emit_i16x8_ne(LiftoffRegister dst, LiftoffRegister lhs,
                             LiftoffRegister rhs);
+  inline void emit_i16x8_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                              LiftoffRegister rhs);
+  inline void emit_i16x8_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                              LiftoffRegister rhs);
   inline void emit_i32x4_eq(LiftoffRegister dst, LiftoffRegister lhs,
                             LiftoffRegister rhs);
   inline void emit_i32x4_ne(LiftoffRegister dst, LiftoffRegister lhs,
                             LiftoffRegister rhs);
+  inline void emit_i32x4_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                              LiftoffRegister rhs);
+  inline void emit_i32x4_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                              LiftoffRegister rhs);
   inline void emit_f32x4_eq(LiftoffRegister dst, LiftoffRegister lhs,
                             LiftoffRegister rhs);
   inline void emit_f32x4_ne(LiftoffRegister dst, LiftoffRegister lhs,

src/wasm/baseline/liftoff-compiler.cc

@@ -2349,14 +2349,26 @@ class LiftoffCompiler {
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_eq);
       case wasm::kExprI8x16Ne:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_ne);
+      case wasm::kExprI8x16GeS:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_ge_s);
+      case wasm::kExprI8x16GeU:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i8x16_ge_u);
       case wasm::kExprI16x8Eq:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_eq);
       case wasm::kExprI16x8Ne:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_ne);
+      case wasm::kExprI16x8GeS:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_ge_s);
+      case wasm::kExprI16x8GeU:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i16x8_ge_u);
       case wasm::kExprI32x4Eq:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_eq);
       case wasm::kExprI32x4Ne:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_ne);
+      case wasm::kExprI32x4GeS:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_ge_s);
+      case wasm::kExprI32x4GeU:
+        return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_i32x4_ge_u);
       case wasm::kExprF32x4Eq:
         return EmitBinOp<kS128, kS128>(&LiftoffAssembler::emit_f32x4_eq);
       case wasm::kExprF32x4Ne:

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

@@ -1996,6 +1996,30 @@ void LiftoffAssembler::emit_i8x16_ne(LiftoffRegister dst, LiftoffRegister lhs,
   Pxor(dst.fp(), kScratchDoubleReg);
 }
 
+void LiftoffAssembler::emit_i8x16_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(kScratchDoubleReg, rhs.fp());
+    ref = kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsb, &Assembler::pminsb>(
+      this, dst, lhs, rhs, SSE4_1);
+  Pcmpeqb(dst.fp(), ref);
+}
+
+void LiftoffAssembler::emit_i8x16_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(kScratchDoubleReg, rhs.fp());
+    ref = kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminub, &Assembler::pminub>(
+      this, dst, lhs, rhs);
+  Pcmpeqb(dst.fp(), ref);
+}
+
 void LiftoffAssembler::emit_i16x8_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vpcmpeqw, &Assembler::pcmpeqw>(
@@ -2010,6 +2034,30 @@ void LiftoffAssembler::emit_i16x8_ne(LiftoffRegister dst, LiftoffRegister lhs,
   Pxor(dst.fp(), kScratchDoubleReg);
 }
 
+void LiftoffAssembler::emit_i16x8_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(kScratchDoubleReg, rhs.fp());
+    ref = kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsw, &Assembler::pminsw>(
+      this, dst, lhs, rhs);
+  Pcmpeqw(dst.fp(), ref);
+}
+
+void LiftoffAssembler::emit_i16x8_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(kScratchDoubleReg, rhs.fp());
+    ref = kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminuw, &Assembler::pminuw>(
+      this, dst, lhs, rhs, SSE4_1);
+  Pcmpeqw(dst.fp(), ref);
+}
+
 void LiftoffAssembler::emit_i32x4_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vpcmpeqd, &Assembler::pcmpeqd>(
@@ -2024,6 +2072,30 @@ void LiftoffAssembler::emit_i32x4_ne(LiftoffRegister dst, LiftoffRegister lhs,
   Pxor(dst.fp(), kScratchDoubleReg);
 }
 
+void LiftoffAssembler::emit_i32x4_ge_s(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(kScratchDoubleReg, rhs.fp());
+    ref = kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminsd, &Assembler::pminsd>(
+      this, dst, lhs, rhs, SSE4_1);
+  Pcmpeqd(dst.fp(), ref);
+}
+
+void LiftoffAssembler::emit_i32x4_ge_u(LiftoffRegister dst, LiftoffRegister lhs,
+                                       LiftoffRegister rhs) {
+  DoubleRegister ref = rhs.fp();
+  if (dst == rhs) {
+    Movaps(kScratchDoubleReg, rhs.fp());
+    ref = kScratchDoubleReg;
+  }
+  liftoff::EmitSimdCommutativeBinOp<&Assembler::vpminud, &Assembler::pminud>(
+      this, dst, lhs, rhs, SSE4_1);
+  Pcmpeqd(dst.fp(), ref);
+}
+
 void LiftoffAssembler::emit_f32x4_eq(LiftoffRegister dst, LiftoffRegister lhs,
                                      LiftoffRegister rhs) {
   liftoff::EmitSimdCommutativeBinOp<&Assembler::vcmpeqps, &Assembler::cmpeqps>(