Bug 1110164: Extend MSimdSelect to also handle element-wise select; r=sunfish

--HG--
extra : rebase_source : 4502a8718d467da247d481476013e8b718f2f38a
extra : histedit_source : 9dd605bf892aa843cc892ab9f0fa07925225df85%2C765790385369b77fa4f13a6921d79109666a9017

js/src/asmjs/AsmJSValidate.cpp

@@ -2656,7 +2656,7 @@ class FunctionCompiler
         MOZ_ASSERT(mask->type() == MIRType_Int32x4);
         MOZ_ASSERT(IsSimdType(lhs->type()) && rhs->type() == lhs->type());
         MOZ_ASSERT(lhs->type() == type);
-        MSimdSelect *ins = MSimdSelect::NewAsmJS(alloc(), mask, lhs, rhs, type);
+        MSimdSelect *ins = MSimdSelect::NewAsmJS(alloc(), mask, lhs, rhs, type, /* isElementWise */ true);
         curBlock_->add(ins);
         return ins;
     }

js/src/jit-test/tests/asm.js/testSIMD.js

@@ -836,13 +836,24 @@ assertAsmTypeFail('glob', USE_ASM + I32 + F32 + I32SEL + "function f() {var m=i4
 assertAsmTypeFail('glob', USE_ASM + I32 + F32 + I32SEL + "function f() {var m=i4(1,2,3,4); var x=i4(1,2,3,4); var y=f4(5,6,7,8); return i4(i4sel(m,x,y));} return f");
 assertAsmTypeFail('glob', USE_ASM + I32 + F32 + I32SEL + "function f() {var m=i4(1,2,3,4); var x=f4(1,2,3,4); var y=f4(5,6,7,8); return i4(i4sel(m,x,y));} return f");
 
-
 assertAsmTypeFail('glob', USE_ASM + F32 + F32SEL + "function f() {var m=f4(1,2,3,4); return f4(f4sel(x,x,x));} return f");
 assertAsmTypeFail('glob', USE_ASM + I32 + F32 + F32SEL + "function f() {var m=f4(1,2,3,4); var x=i4(1,2,3,4); return f4(f4sel(m,x,x));} return f");
 assertAsmTypeFail('glob', USE_ASM + I32 + F32 + F32SEL + "function f() {var m=f4(1,2,3,4); var x=f4(1,2,3,4); return f4(f4sel(m,x,x));} return f");
 assertAsmTypeFail('glob', USE_ASM + I32 + F32 + F32SEL + "function f() {var m=i4(1,2,3,4); var x=f4(1,2,3,4); var y=i4(5,6,7,8); return f4(f4sel(m,x,y));} return f");
 assertAsmTypeFail('glob', USE_ASM + I32 + F32 + F32SEL + "function f() {var m=i4(1,2,3,4); var x=i4(1,2,3,4); var y=f4(5,6,7,8); return f4(f4sel(m,x,y));} return f");
 
+// These pass with select but not bitselect
+assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + I32SEL + "function f() {var m=i4(0,0,0,0); var x=i4(1,2,3,4); var y=i4(5,6,7,8); return i4(i4sel(m,x,y)); } return f"), this)(), [5, 6, 7, 8]);
+assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + I32SEL + "function f() {var m=i4(-1,-2,-3,-42); var x=i4(1,2,3,4); var y=i4(5,6,7,8); return i4(i4sel(m,x,y)); } return f"), this)(), [1, 2, 3, 4]);
+assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + I32SEL + "function f() {var m=i4(1,-1,2,-2); var x=i4(1,2,3,4); var y=i4(5,6,7,8); return i4(i4sel(m,x,y)); } return f"), this)(), [5, 2, 7, 4]);
+assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + I32SEL + "function f() {var m=i4(42,45,-42,-47); var x=i4(1,2,3,4); var y=i4(5,6,7,8); return i4(i4sel(m,x,y)); } return f"), this)(), [5, 6, 3, 4]);
+
+assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + F32 + F32SEL + "function f() {var m=i4(0,0,0,0); var x=f4(1,2,3,4); var y=f4(5,6,7,8); return f4(f4sel(m,x,y)); } return f"), this)(), [5, 6, 7, 8]);
+assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + F32 + F32SEL + "function f() {var m=i4(-1,-2,-3,-42); var x=f4(1,2,3,4); var y=f4(5,6,7,8); return f4(f4sel(m,x,y)); } return f"), this)(), [1, 2, 3, 4]);
+assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + F32 + F32SEL + "function f() {var m=i4(1,-1,2,-2); var x=f4(1,2,3,4); var y=f4(5,6,7,8); return f4(f4sel(m,x,y)); } return f"), this)(), [5, 2, 7, 4]);
+assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + F32 + F32SEL + "function f() {var m=i4(42,45,-42,-47); var x=f4(1,2,3,4); var y=f4(5,6,7,8); return f4(f4sel(m,x,y)); } return f"), this)(), [5, 6, 3, 4]);
+
+// These pass for both select and bitselect
 assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + I32SEL + "function f() {var m=i4(0,0,0,0); var x=i4(1,2,3,4); var y=i4(5,6,7,8); return i4(i4sel(m,x,y)); } return f"), this)(), [5, 6, 7, 8]);
 assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + I32SEL + "function f() {var m=i4(0xffffffff,0xffffffff,0xffffffff,0xffffffff); var x=i4(1,2,3,4); var y=i4(5,6,7,8); return i4(i4sel(m,x,y)); } return f"), this)(), [1, 2, 3, 4]);
 assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + I32SEL + "function f() {var m=i4(0,0xffffffff,0,0xffffffff); var x=i4(1,2,3,4); var y=i4(5,6,7,8); return i4(i4sel(m,x,y)); } return f"), this)(), [5, 2, 7, 4]);

js/src/jit/LIR-Common.h

@@ -498,6 +498,9 @@ class LSimdSelect : public LInstructionHelper<1, 3, 1>
     const LDefinition *temp() {
         return getTemp(0);
     }
+    MSimdSelect *mir() const {
+        return mir_->toSimdSelect();
+    }
 };
 
 // Constant 32-bit integer.

js/src/jit/MIR.h

@@ -2023,9 +2023,11 @@ class MSimdShift : public MBinaryInstruction
 
 class MSimdSelect : public MTernaryInstruction
 {
-  private:
-    MSimdSelect(MDefinition *mask, MDefinition *lhs, MDefinition *rhs, MIRType type)
-      : MTernaryInstruction(mask, lhs, rhs)
+    bool isElementWise_;
+
+    MSimdSelect(MDefinition *mask, MDefinition *lhs, MDefinition *rhs, MIRType type,
+                bool isElementWise)
+      : MTernaryInstruction(mask, lhs, rhs), isElementWise_(isElementWise)
     {
         MOZ_ASSERT(IsSimdType(type));
         MOZ_ASSERT(mask->type() == MIRType_Int32x4);
@@ -2038,15 +2040,29 @@ class MSimdSelect : public MTernaryInstruction
   public:
     INSTRUCTION_HEADER(SimdSelect);
     static MSimdSelect *NewAsmJS(TempAllocator &alloc, MDefinition *mask, MDefinition *lhs,
-                                 MDefinition *rhs, MIRType t)
+                                 MDefinition *rhs, MIRType t, bool isElementWise)
     {
-        return new(alloc) MSimdSelect(mask, lhs, rhs, t);
+        return new(alloc) MSimdSelect(mask, lhs, rhs, t, isElementWise);
+    }
+
+    MDefinition *mask() const {
+        return getOperand(0);
     }
 
     AliasSet getAliasSet() const {
         return AliasSet::None();
     }
 
+    bool isElementWise() const {
+        return isElementWise_;
+    }
+
+    bool congruentTo(const MDefinition *ins) const {
+        if (!congruentIfOperandsEqual(ins))
+            return false;
+        return isElementWise_ == ins->toSimdSelect()->isElementWise();
+    }
+
     ALLOW_CLONE(MSimdSelect)
 };
 

js/src/jit/shared/CodeGenerator-x86-shared.cpp

@@ -2964,7 +2964,22 @@ CodeGeneratorX86Shared::visitSimdSelect(LSimdSelect *ins)
     if (mask != temp)
         masm.vmovaps(mask, temp);
 
-    masm.bitwiseAndX4(Operand(mask), output);
+    MSimdSelect *mir = ins->mir();
+    if (mir->isElementWise()) {
+        if (AssemblerX86Shared::HasAVX()) {
+            masm.vblendvps(mask, onTrue, onFalse, output);
+            return;
+        }
+
+        // SSE4.1 has plain blendvps which can do this, but it is awkward
+        // to use because it requires the mask to be in xmm0.
+
+        // Propagate sign to all bits of mask vector, if necessary.
+        if (!mir->mask()->isSimdBinaryComp())
+            masm.packedRightShiftByScalar(Imm32(31), temp);
+    }
+
+    masm.bitwiseAndX4(Operand(temp), output);
     masm.bitwiseAndNotX4(Operand(onFalse), temp);
     masm.bitwiseOrX4(Operand(temp), output);
 }