818138 - enterFakeExitFrame can do a better job with the assembly it generates

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Boris Zbarsky [:bzbarsky]]

Right now, for something like enterFakeExitFrame(ION_FRAME_DOMGETTER) we get codegen like so on x86-84:

> movabsq    $0x1, %r11                   #bz Stick 1 in %r11
> push       %r11                         #bz Push 1
> movabsq    $0x0, %r11                   #bz Stick 0 in %r11
> push       %r11                         #bz Push 0

The problem is that it's not clear to me that the codeVal argument is always something that can be pushed as an immediate...  But null sure is.

Comment 1

[Boris Zbarsky [:bzbarsky]]

So what buildFakeExitFrame (which ends up with a pushl) does is this:

364         Push(Imm32(descriptor));

while enterFakeExitFrame does this:

        Push(ImmWord(uintptr_t(codeVal)));
        Push(ImmWord(uintptr_t(NULL)));

The problem is that presumably fake and real exit frames need to match size, and real exit frames have:

        exitCodePatch_ = PushWithPatch(ImmWord(-1));
        // Push VMFunction pointer, to mark arguments.
        Push(ImmWord(f));

which is in fact pushing two pointer-sized values.

Comment 2

[Nathan Froyd [:froydnj]]

Pushing pointer-sized values is fine; Push (or similar) just needs to be taught about the sign-extension semantics when pushing immediates.

Wonder what the codegen looks like on x86.

Comment 3

[Boris Zbarsky [:bzbarsky]]

Looks like pushl on x86-64 will in fact sign-extend 32-bit immediates to a 64-bit value.

I'm building a 32-bit build at the moment to answer that last question.  ;)

Comment 4

[Tom S. (please needinfo tschuster)]

Attachment: optimize push(immword) to push(imm32) sometimes

Comment 5

[Boris Zbarsky [:bzbarsky]]

In a 32-bit build, this bit of asm looks like so:

pushl      $0x1
pushl      $0x0

Comment 6

[Jan de Mooij [:jandem]]

Comment on attachment 688363 [details] [diff] [review]
optimize push(immword) to push(imm32) sometimes

Review of attachment 688363 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/ion/IonMacroAssembler.h
@@ +491,5 @@
>          exitCodePatch_ = PushWithPatch(ImmWord(-1));
>          // Push VMFunction pointer, to mark arguments.
>          Push(ImmWord(f));
>      }
> +    void enterFakeExitFrame(IonCode *codeVal = NULL) {

Pre-existing, but can you add a comment to this method like this:

// Note: use ImmWord, not ImmGCPtr, since codeVal is not a valid IonCode
// pointer but a magic value like ION_FRAME_DOMGETTER.

Comment 7

[Nathan Froyd [:froydnj]]

Comment on attachment 688363 [details] [diff] [review]
optimize push(immword) to push(imm32) sometimes

Review of attachment 688363 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/ion/x64/Assembler-x64.h
@@ +290,5 @@
>      }
>      void push(const ImmWord ptr) {
> +        // We often end up with ImmWords that actually fit into int32.
> +        // Be aware of the sign extension behavior.
> +        if (ptr.value <= INT32_MAX) {

This needs to check for ptr.value >= INT32_MIN as well, otherwise pointers whose numeric values are < INT32_MIN will be corrupted.  Or is the pointer guaranteed to fit into a 32-bit address space?  (Probably not a bad idea to check anyway.)

Comment 8

[Tom S. (please needinfo tschuster)]

ptr.value is an uintptr_t, so it can never be negative. So this should result in an unsigned compare.

Comment 9

[Tom S. (please needinfo tschuster)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/d148382e074c

Comment 10

[Ed Morley [:emorley]]

https://hg.mozilla.org/mozilla-central/rev/d148382e074c