525413 - lirasm: avoid operandCount

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Nicholas Nethercote [inactive]]

The main switch in assembleFragment() currently handles some LIR opcodes
specifically, and then falls back to generic code based on operandCounts for
other cases.  This is bad, see bug 505662 for more details.

This patch gets rid of this behaviour in lirasm by removing
assemble_general() and handling all LIR opcodes explicitly in the switch.
Some other benefits of this:

- LIR_start/LIR_label/LIR_callh now give useful error messages if they are
  used, rather than random crashes/asserts/wrong behaviour.

- LIR_cmov/LIR_qcmov are now implemented, as is qcall.

- LIR_float is now handled.  LIR_quad can now only take integral arguments.
  The tests have been updated accordingly.

- LIR_qcall is now handled.

- mOpMap is simpler -- the values are now just LOpcodes instead of a pair.

Other things in the patch:

- error messages now say "line 123: ..." instead of "instruction 123: ..."
  because it's really the line number.

- assemble_guard() and assemble_jump() have been tweaked.

Comment 1

[Nicholas Nethercote [inactive]]

Attachment: patch

Comment 2

[Graydon Hoare :graydon]

http://hg.mozilla.org/projects/nanojit-central/rev/f35c3aab8518

Comment 3

[Robert Sayre]

http://hg.mozilla.org/mozilla-central/rev/2cafb905b5a0