[gcc] Eliminate control integrity pass in favor of nacl instructions generation from the start.

[gcc] Eliminate control integrity pass in favor of nacl instructions generation from the start.

This is a refactoring/cleanup CL. The advatage of this new control integrity implementation is that it touches less files and is fully contained in i386 backend. This seems to be easier to push in upstream.

Specs pass, and performance seems unaffected. I have a concern about generating unspecs at rtl expand phase, as it can possibly kill several optimizations. From the other side, it seems unlikely as optimizations that touch indirect calls and jumps are hopefully done before rtl. In any case, I have an idea how to improve this if we discover that some optimization suffers from the change.

[eaeltsin, 2010-06-21 19:04:09]

This issue seems to be created correctly.
Please review.

[pasko-google - do not use, 2010-06-21 21:42:11]

http://codereview.chromium.org/2850019/diff/1/3
File gcc/gcc/config/i386/i386.c (right):

http://codereview.chromium.org/2850019/diff/1/3#newcode19119
gcc/gcc/config/i386/i386.c:19119: +    call = gen_rtx_UNSPEC (VOIDmode,
gen_rtvec (2, fnaddr, callarg1),
did you investigate the possibility to make a (call (unspec ...))

this way we would be able to say things like "this insn is a call" and some
rules can be applied to that (say, register save/restore, or PIC?). Though it's
the backend where this information is probably already have been used. Not sure.


Please, also run the GCC testsuite with this change.

[eaeltsin, 2010-06-22 04:04:55]

http://codereview.chromium.org/2850019/diff/1/3
File gcc/gcc/config/i386/i386.c (right):

http://codereview.chromium.org/2850019/diff/1/3#newcode19119
gcc/gcc/config/i386/i386.c:19119: +    call = gen_rtx_UNSPEC (VOIDmode,
gen_rtvec (2, fnaddr, callarg1),
On 2010/06/21 21:42:11, pasko wrote:
> did you investigate the possibility to make a (call (unspec ...))

I generate rtl that is matchable with naclcall define_insns. Their patterns are
unspec _instead_of_ call.

At some point it may be worth trying what you suggest, but this will require
changing all naclcall define_insns rtl patterns.

> 
> this way we would be able to say things like "this insn is a call" and some
> rules can be applied to that (say, register save/restore, or PIC?). Though
it's
> the backend where this information is probably already have been used. Not
sure.

This is still a CALL_INSN, so the stuff you mention applies.

However, making its rtl to have a CALL expression may help gcc to optimize. It
is worth investigating, but as the performance didn't regress with this change,
it is of low priority.

> 
> 
> Please, also run the GCC testsuite with this change.

[pasko-google - do not use, 2010-06-22 08:34:43]

LGTM