[PATCH, i386]: Fix PR 46716, [4.4/4.5/4.6/4.7/4.8 Regression] wrong code generated with -mno-sse2 -m64

[Uros Bizjak <ubizjak at gmail dot com>]

Hello!

i386 maintains stable ABI by passing "unsupported" SSE modes in hard
SSE registers, even if original mode does not fit there. This works
quite well for 32bit targets (-msse2 vs -msse), but 64bit targets
assume SSE2 by default, so various paths aren't well tested for
-mno-sse2.

Attached patch fixes one particular problem with -mno-sse2 on 64bit
targets. If the "natural" mode doesn't agree with the mode, choosen by
compiler type system, then compiler decomposes "natural" mode in some
supported mode, breaking various assumptions on how arguments in
"natural" mode are passed around. The patch adopts the same approach
as 32bit targets - it wraps argument in a parallel RTX in a couple of
missed places, tricking the compiler to pass the argument in "natural"
mode. (Please see the comment before gen_reg_or_parallel for further
details).

2012-03-02  Uros Bizjak  <ubizjak@gmail.com>

	PR target/46716
	* config/i386/i386.c (construct_container): Use gen_reg_or_parallel
	to pass the argument in the register of "natural" mode.

testsuite/ChangeLog:

2012-03-02  Uros Bizjak  <ubizjak@gmail.com>

	PR target/46716
	* gcc.target/i386/pr46176.c: New test.

Patch was tested on x86_64-pc-linux-gnu {,-m32}.  Patch was committed
to mainline, is it OK to commit to 4.7 after it opens?

I will backport the patch to other release branches in a couple of days.

Uros.

Attachment: p.diff.txt
Description: Text document