[llvm-commits] [llvm][PATCH][Review Requested] Maintain live-ins for anti-dependency checking

[Gurd, Preston]

Hello Anton,

Thank you for your comments.

We will reformat the code as you suggested before resubmitting the patch.

As it stands right now, a few of our test cases break when run on the Intel Atom processor, without the proposed patch, because the code generator for the Atom uses the Post-RA scheduler.

To eliminate the possibility that regular X86 code generation might be affected by this change, the proposed patch will be amended to make sure that the register scavenger will only run when doing Post-RA scheduling for Intel Atom and will not run for any other X86 subtarget.

If it is still desired that this patch should not go into 3.1, then the short term fix will be to change the Post-RA scheduler to turn off anti-dependency checking by using ANTIDEP_NONE instead of ANTIDEP_CRITICAL and by temporarily XFAILing one or two tests.

Preston


-----Original Message-----
From: Anton Korobeynikov [mailto:anton at korobeynikov.info] 
Sent: Monday, April 16, 2012 2:21 PM
To: Gurd, Preston
Cc: llvm-commits at cs.uiuc.edu
Subject: Re: [llvm-commits] [llvm][PATCH][Review Requested] Maintain live-ins for anti-dependency checking

Hello, Preston

First, one minor nit-pick: please follow LLVM's code style (look into surrounding code).
Next, register scavenger dependes heavily on correct register liveness information. Given that X86 never had regscavenger enabled before, I'd turn it for 3.2, not 3.1.
This will ensure that such really important change for important target will be carefully tested and bugs (if any - though, for example, 2.9 was not machine verifier clean on x86) will be fixed.
This is definitely not the last minute change.

--
With best regards, Anton Korobeynikov
Faculty of Mathematics and Mechanics, Saint Petersburg State University