[gcc] Add clobber FLAGS_REG to NaCl indirect jumps

[gcc] Add clobber FLAGS_REG to NaCl indirect jumps

Compiler must know that NaCl indirect jump clobbers FLAGS register.

For call and return this is probably not important, as FLAGS is always clobbered by call.

For computed goto and switch, this IS important, so added explicit clobber.

TEST=very hard to reproduce, we only suspect this bug was the reason for LLVM miscompilation.

BUG=

[khim, 2010-10-20 14:01:59]

LGTM

[pasko-google - do not use, 2010-10-20 14:25:15]

do you have a test to reproduce the failure? can we commit it somewhere?

http://codereview.chromium.org/3892003/diff/1/2
File gcc/gcc/config/i386/i386.md (right):

http://codereview.chromium.org/3892003/diff/1/2#newcode15274
gcc/gcc/config/i386/i386.md:15274: if (TARGET_NACL)
I see no reason in creating an expand that does nothing with !TARGET_NACL.
"TARGET_NACL" two lines above?

[eaeltsin, 2010-10-20 14:30:43]

Unfortunately we don't have a reproducer, there is only a report from David
Meyer that gcc miscompiles LLVM with this bug being the likely reason. David was
unable to make the test even a bit smaller.

http://codereview.chromium.org/3892003/diff/1/2
File gcc/gcc/config/i386/i386.md (right):

http://codereview.chromium.org/3892003/diff/1/2#newcode15274
gcc/gcc/config/i386/i386.md:15274: if (TARGET_NACL)
On 2010/10/20 14:25:15, pasko wrote:
> I see no reason in creating an expand that does nothing with !TARGET_NACL.
> "TARGET_NACL" two lines above?

You simply forgot that if define_expand preparation statements do not end with
DONE or FAIL, it outputs the original rtl pattern.

[pasko-google - do not use, 2010-10-20 14:48:44]

LGTM

http://codereview.chromium.org/3892003/diff/1/2
File gcc/gcc/config/i386/i386.md (right):

http://codereview.chromium.org/3892003/diff/1/2#newcode15274
gcc/gcc/config/i386/i386.md:15274: if (TARGET_NACL)
On 2010/10/20 14:30:43, Evgeny Eltsin wrote:
> On 2010/10/20 14:25:15, pasko wrote:
> > I see no reason in creating an expand that does nothing with !TARGET_NACL.
> > "TARGET_NACL" two lines above?
> 
> You simply forgot that if define_expand preparation statements do not end with
> DONE or FAIL, it outputs the original rtl pattern.

OK, never knew that