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: [patch] gcc.dg/tree-ssa/*.c: Use -O2 in some testcases.

On Mon, May 09, 2005 at 07:15:30PM -0400, Kazu Hirata wrote:

> The patch partially fixes this problem by raising -O1 to -O2.  In
> 20040721-1.c, I was able to disable DOM and still get expected
> optimizations.  In other testcases, we seem to at least partially
> depend on DOM to get expected optimizations.
After I commit the pass reorg patch I'm working on, DOM will run
after the initial scalar cleanups, so this should not be
necessary.  But this is fine in the meantime.

> A complete fix would involve checking messages from the propagator so
> that expected optimizations will happen in nothing but store_ccp, but
> that's not as easy as I thought.  For example, we can't just turn the
> FRE portion of DOM and disable the rest, at least in the current pass
> ordering.
Well, I'm not so interested in *what* does the optimization so
much as in getting it done.  The problem with testing
pass-specific optimizations is that reshuffling the optimization
pipeline becomes an annoying process.  I would much prefer doing
either link-time tests or .optimized dump scans.

I know it's not perfect, because we may get the program optimized
by blind-luck otherwise, but we could for instance enable just
the pass that is supposed to do the work, and nothing else.

> 	PR testsuite/21443
> 	* gcc.dg/tree-ssa/20030731-2.c, gcc.dg/tree-ssa/20030917-1.c,
> 	gcc.dg/tree-ssa/20030917-3.c, gcc.dg/tree-ssa/ssa-ccp-1.c,
> 	gcc.dg/tree-ssa/ssa-ccp-2.c, gcc.dg/tree-ssa/ssa-ccp-3.c,
> 	gcc.dg/tree-ssa/ssa-ccp-7.c, gcc.dg/tree-ssa/ssa-ccp-9.c:
> 	Change -O1 to -O2.
> 	* gcc.dg/tree-ssa/20040721-1.c: Change -O1 to -O2.  Add
> 	-fno-dominator-opts.


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