This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [tree-ssa] SSA verifier

> Adds Jan's SSA and statement verifier together with fixes for the bugs
> it caught.
> In SRA, we weren't marking all the variables that needed to be renamed
> in various cases.  In the must-alias pass, we were promoting variables
> with hidden uses, and in fold_stmt() we weren't resetting VDEFs and
> VUSEs when replacing the whole statement.
> It also caught a PRE bug that Dan fixed earlier today.
> I have added a call to verify_ssa() after every pass.  This is only
> active with --enable-checking, and it's quite heavy handed.  It adds
> about 18 seconds to a 560 second compilation of a typical GCC bootstrap.
> It hasn't really been noticeable for me, but we may want to cut down on
> the number of calls to verify_ssa().  We could probably introduce a new
> --enable-checking argument for checking data flow.  Thoughts?

18 second slodwon is very fine.  I was having slowdown approximately
20-40% originally.  I think that recent work on making
vdefs/vuses/uses/defs more memory friendly is responsible for fixing
this slodown, so I no longer opose against using it after each pass.

Thank you!

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]