557705 - nanojit: clean up codegen of i386 jumps

Status: RESOLVED FIXED

(Core Graveyard :: Nanojit, defect)

Description

[Nicholas Nethercote [inactive]]

Attachment: patch

This patch:

- Simplifies printing of jmp targets -- it prints 't' directly rather than
  computing (in a roundabout fashion) 't - _nIns + _nIns'.

- In JMP(), handles correctly a target of 0 (which requires later patching) by
  design, not just by luck.  In JCC() and JMP_long() it asserts that the
  target is not zero.

- Inlines JMP_long_nochk_offset() into its two callers and then removes it.
  Having that code factored out actually made things more complicated.

- Moves three jmp constants to where they're more visible.

These changes will help bug 506693.  I confirmed that codegen is unchanged
for TM on SunSpider.

Comment 1

[Nicholas Nethercote [inactive]]

http://hg.mozilla.org/projects/nanojit-central/rev/ec93bc283385

http://hg.mozilla.org/tracemonkey/rev/315c120ca1e7

Comment 2

[Edwin Smith]

Attachment: Loosen code in JCC() to allow for patchable backwards branches

TR generates conditional branches for loop edges (back-branches), which requires us to call asm_branch with target == 0, then patch later.

This patch loses the assert that t (target) != 0, but explicitly handles the patch case (target == 0) like the way its done in the X64 backend, instead of working by accident.

Comment 3

[Edwin Smith]

NJ fix: http://hg.mozilla.org/projects/nanojit-central/rev/5f12e1da6b7f

TR original: http://hg.mozilla.org/tamarin-redux/rev/7fc928af3c78
TR fix: http://hg.mozilla.org/tamarin-redux/rev/0dc2670f9505

Comment 4

[Nicholas Nethercote [inactive]]

TM fix: http://hg.mozilla.org/tracemonkey/rev/53895deb2fed

Comment 5

[Robert Sayre]

http://hg.mozilla.org/mozilla-central/rev/315c120ca1e7