685707 - IonMonkey: Implement JSOP_DIV

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[David Anderson [:dvander] - inactive, e-mail if emergency]

Need this to run a Mandelbrot benchmark.

Comment 1

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: fix

Comment 2

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: v2

The original patch broke constant folding.

Comment 3

[Sean Stangl [:sstangl]]

Comment on attachment 559353 [details] [diff] [review]
v2

Review of attachment 559353 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/ion/MIR.cpp
@@ +565,5 @@
> +    if (specialization_ == MIRType_None)
> +        return this;
> +
> +    MDefinition *lhs = getOperand(0);
> +    MDefinition *rhs = getOperand(1);

From Bug 685695,
MDefinition *lhs = lhs();
MDefinition *rhs = rhs();

@@ +571,5 @@
> +        return folded;
> +
> +    // 0 / x -> 0
> +    // x / 1 -> x
> +    if (IsConstant(lhs, 0) || IsConstant(rhs, 1))

Also "-0 / x -> -0" and "NaN / x -> NaN", but perhaps these doesn't pass the threshold for usefulness.

It might be more reasonable to handle the NaN case (for all arith) in EvaluateConstantOperands() -- if one of the operands is constant and NaN, the other does not need to be constant for the result to be known.

::: js/src/ion/shared/CodeGenerator-x86-shared.cpp
@@ +453,5 @@
> +{
> +    Register remainder = ToRegister(ins->remainder());
> +    Register lhs = ToRegister(ins->lhs());
> +    Register rhs = ToRegister(ins->rhs());
> +

If we alias %eax to %rax and %edx to %rdx for x64, as in JSC's Assembler, then we can assert that the Registers are as required.

@@ +472,5 @@
> +    // Prevent negative 0.
> +    Label nonzero;
> +    masm.testl(lhs, lhs);
> +    masm.j(Assembler::NonZero, &nonzero);
> +    masm.cmpl(rhs, Imm32(0));

I don't know whether a test on the sign bit should be preferred over cmp in this case, but I suspect cmp is fine.

@@ +483,5 @@
> +    masm.idiv(rhs);
> +
> +    // If the remainder is > 0, bailout since this must be a double.
> +    masm.testl(remainder, remainder);
> +    if (!bailoutIf(Assembler::NonZero, ins->snapshot()))

Having two integer operands and a double result seems like a very common case, but this is fine for now.

::: js/src/ion/shared/LIR-x86-shared.h
@@ +44,5 @@
> +
> +namespace js {
> +namespace ion {
> +
> +class LDivI : public LBinaryMath<1>

LBinaryMath<Temps(1)>, ideally.

::: js/src/ion/x64/Lowering-x64.cpp
@@ +215,5 @@
>  
> +bool
> +LIRGeneratorX64::lowerDivI(MDiv *div)
> +{
> +    LDivI *lir = new LDivI(useFixed(div->lhs(), rax), useRegister(div->lhs()), tempFixed(rdx));

Technically %eax and %edx in the JSC assembler are aliases for %rax and %rdx, respectively, so we could eliminate the duplication via some Lowering-x86-shared.{cpp,h}. The duplication is fine here, though.

::: js/src/jit-test/tests/sunspider/check-math-spectral-norm.js
@@ +4,5 @@
>  // contributed by Ian Osgood
>  
>  function A(i,j) {
> +  //var t = (i+j)*(i+j+1);
> +  //var k = 2+i+1;

Probably not intended for checkin.

Comment 4

[David Anderson [:dvander] - inactive, e-mail if emergency]

http://hg.mozilla.org/projects/ionmonkey/rev/e6f9d57d8836

I added the register aliases, but am going to hold off on Lowering-x86-shared until we get more duplication.

> LBinaryMath<Temps(1)>, ideally.

Where Temps is what?