869525 - Use xor+setcc instead of setcc+movzbl in BaselineIC-x86-shared.cpp

Status: RESOLVED FIXED

(Core :: JavaScript Engine, enhancement)

Description

[Dan Gohman [:sunfish]]

BaselineIC-x86-shared.cpp has code that does this:

  ucomisd %xmm0, %xmm1
  seta    %cl
  movzbl  %cl, %ecx

Since it happens to be convenient to do so, this would be better:

  xorl    %ecx, %ecx
  ucomisd %xmm0, %xmm1
  seta    %cl

because it has smaller code size and on common processors today, and the xor of a register with itself is handled specially.

Comment 1

[Dan Gohman [:sunfish]]

Attachment: a patch containing a proposed fix

Comment 2

[Dan Gohman [:sunfish]]

This patch depends on the patch in bug 869532.

Comment 3

[Dan Gohman [:sunfish]]

Attachment: refresh patch to apply to trunk

Comment 4

[Nicolas B. Pierron [:nbp]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/0f05638c8f26

Comment 5

[Jeff Muizelaar [:jrmuizel]]

Are you sure that this is better? Both gcc and clang use the mozbl variant.

Comment 6

[Dan Gohman [:sunfish]]

GCC (4.6 and 4.8) on my system uses the xor variant, for this C code with -O2 for example:

long foo(double x) { return x > 0; }

The xor variant is part of the "peephole2" pass in GCC, which specifically rewrites setcc+movzbl to xor+setcc when it can.

Also, the xor variant is recommended first by Agner's "Optimizing subroutines in assembly 
language", section "Replacing conditional jumps with conditional set instructions" [0].

Also, Intel uses the xor variant in their x86 optimizing manual [1]. It isn't a specific recommendation, but they do use it in an example (Example 3-2. "Code Optimization to Eliminate Branches").

Also, the xor trick has a smaller overall encoding.

[0] http://www.agner.org/optimize/optimizing_assembly.pdf
[1] http://www.intel.com/content/dam/doc/manual/64-ia-32-architectures-optimization-manual.pdf

Comment 7

[Jeff Muizelaar [:jrmuizel]]

(In reply to Dan Gohman from comment #6)
> GCC (4.6 and 4.8) on my system uses the xor variant, for this C code with
> -O2 for example:


Very true. I forgot my gcc was very old. I've filed a bug against clang to add the optimization there:
http://llvm.org/bugs/show_bug.cgi?id=15946

Comment 8

[Ed Morley [:emorley]]

https://hg.mozilla.org/mozilla-central/rev/0f05638c8f26