516093 - NJ: Bad register allocation in NativeX64.cpp!regalloc_binary

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: patch

regalloc_binary() shortcuts if the RHS already has a reservation with a valid register, but it doesn't make sure this register is in the allow mask. This meant the backend could try to generate code like:

addsd xmm0, r12

Which would come out as:

addsd xmm0, xmm12

This can easily happen if the previous use of the RHS is a store, and also the last use, because it will get a GPR by default.

This patch does three things:
 1. Fetches a new allowable register in the case described above.
 2. Asserts in the x64 regalloc functions that the outputs are valid.
 3. Adds a special-case hint to asm_store64, which should be a small register allocation improvement.

Comment 1

[Andreas Gal :gal]

Comment on attachment 400197 [details] [diff] [review]
patch

You should add an assert that this fixup only happens for the few cases we know this makes sense for (quad, ldq).

Comment 2

[David Anderson [:dvander] - inactive, e-mail if emergency]

http://hg.mozilla.org/tracemonkey/rev/8ac0950cbd76

Comment 3

[Blake Kaplan (:mrbkap) (inactive)]

http://hg.mozilla.org/mozilla-central/rev/8ac0950cbd76

Comment 4

[Rick Reitmaier]

Does it also make sense for the conversion functions; i2f u2f ?

Comment 5

[Robert Sayre]

http://hg.mozilla.org/releases/mozilla-1.9.2/rev/b8bc23c69e0e