958799 - Gecko profiler inhibits some JIT optimizations?

Status: RESOLVED FIXED

(Core :: Gecko Profiler, defect)

Description

[Boris Zbarsky [:bzbarsky]]

Attachment: test.html

If I run the attached testcase with Gecko profiler active, it shows 5ns or so per iteration.  Without Gecko profiler active it's about 0.7ns.

Here's what JIT inspector shows without the profiler active:

  Block #3 -> #4 -> #5 :: 9998904 hits
  [Label]
  [InterruptCheckImplicit]
  [OsiPoint]
  [CompareAndBranch:lt]
  cmpl       %ecx, %edx
  jge        ((571))
  
  Block #4 -> #3 :: 9998903 hits
  [Label]
  [AddI]
  addl       $0x1, %edx
  [Goto]
  jmp        ((593))
  ##link     ((593)) jumps to ((593))

Here's the same thing with the profiler active:

  Block #3 -> #4 -> #7 :: 9998904 hits
  [Label]
  [InterruptCheckImplicit]
  [OsiPoint]
  [CompareAndBranch:lt]
  cmpl       %ecx, %edx
  jge        ((981))
  
  Block #4 -> #5 :: 9998903 hits
  [Label]
  [FunctionBoundary]
  movabsq    $0x100572040, %rbx
  movl       0x0(%rbx), %ebx
  addl       $0xffffffff, %ebx
  cmpl       $0x400, %ebx
  jge        ((1022))
  shlq       $5, %rbx
  movabsq    $0x10056a000, %r11
  addq       %r11, %rbx
  movl       $0x2c, 0x18(%rbx)
  ##link     ((1022)) jumps to ((1046))
  movabsq    $0x100572040, %rbx
  movl       0x0(%rbx), %ebx
  cmpl       $0x400, %ebx
  jge        ((1070))
  shlq       $5, %rbx
  movabsq    $0x10056a000, %r11
  addq       %r11, %rbx
  movabsq    $0x1abd1be50, %r11
  movq       %r11, 0x0(%rbx)
  movabsq    $0x1379cf480, %r11
  movq       %r11, 0x10(%rbx)
  movq       $0, 0x8(%rbx)
  movl       $0xffffffff, 0x18(%rbx)
  ##link     ((1070)) jumps to ((1129))
  movabsq    $0x100572040, %rbx
  addl       $1, 0x0(%rbx)
  [Goto]
  
  Block #5 -> #6 :: 9998903 hits
  [Label]
  [FunctionBoundary]
  movabsq    $0x100572040, %rbx
  addl       $-1, 0x0(%rbx)
  [Goto]
  
  Block #6 -> #3 :: 9998903 hits
  [Label]
  [FunctionBoundary]
  movabsq    $0x100572040, %rbx
  movl       0x0(%rbx), %ebx
  addl       $0xffffffff, %ebx
  cmpl       $0x400, %ebx
  jge        ((1210))
  shlq       $5, %rbx
  movabsq    $0x10056a000, %r11
  addq       %r11, %rbx
  movl       $0xffffffff, 0x18(%rbx)
  ##link     ((1210)) jumps to ((1234))
  [AddI]
  addl       $0x1, %edx
  [Goto]
  jmp        ((1242))
  ##link     ((1242)) jumps to ((1242))
  
For simple enough function bodies, the instrumentation seems to dominate the actual inlined function or something.  :(  Not sure what we can do here...

Comment 1

[Boris Zbarsky [:bzbarsky]]

Hmm.  So MFunctionBoundary doesn't alias anything, right?  But it's also not movable or CSEable...  So it looks quite possible for the contents of the function, after inlining to be moved/CSEd across MFunctionBoundary?

Comment 2

[Benoit Girard (:BenWa)]

I think this is a dupe.

Comment 3

[Jan de Mooij [:jandem]]

See also the discussion in bug 948229.

Comment 4

[Kannan Vijayan [:djvj]]

Boris, can you re-check this and let me know if the performance has improved?  I just landed 948229 so this should be fixed now.  Confirmation would be good.

Comment 5

[Boris Zbarsky [:bzbarsky]]

Yes, this seems to be fixed.