Re: Fix PR399955: rs6000 needs to handle SSA_NAMEs

[Michael Meissner <meissner at linux dot vnet dot ibm dot com>]

On Wed, Apr 29, 2009 at 07:48:07PM +0200, Michael Matz wrote:
> Hi,
> 
> currently we get ICEs on powerpc when SDmode variables are involved, 
> because the backend goes over the gimple instructions to detect the usage 
> of that mode and allocate a stack slot.  With SSA expand this can contain 
> SSA_NAMEs which we need to look at too, not just VAR_DECLs.
> 
> This patch was regstrapped on powerpc64-linux with the same configure 
> options that Janis is using (so I can compare testresults posting with my 
> results).
> 
> The FAILs I now see are either preexisting (from 
> http://gcc.gnu.org/ml/gcc-testresults/2009-04/msg02785.html), are Richards 
> ICEs with the type verifier where the vectorizer introduces invalid 
> conversions, or are these ones:
> 
> FAIL: gcc.dg/sms-8.c scan-rtl-dump-times sms "SMS succeeded" 1
> FAIL: gcc.dg/vect/slp-perm-8.c scan-tree-dump-times vect "vectorized 1 loops" 2
> FAIL: gcc.target/powerpc/20050603-3.c scan-assembler-not inm
> FAIL: gcc.target/powerpc/405-dlmzb-strlen-1.c scan-assembler dlmzb\\\\.
> FAIL: gcc.target/powerpc/440-dlmzb-strlen-1.c scan-assembler dlmzb\\\\.
> FAIL: events output
> FAIL: 22_locale/codecvt/unshift/char/1.cc execution test
> 
> sms-8 uses long long arithmetic on 32 bit, which involves setting 
> SImode subregs of DImode regs, which SMS can't handle.  No idea why it 
> should have worked before.  The other scan-asm and vector test are most 
> probably changes exposed by the limited TER, but I haven't investigated in 
> detail yet.  1.cc seems to fail also for others sometimes.  no idea about 
> the "events output" one.
> 
> Okay?

While the whole SDmode stack slot calls out for rewriting, this patch looks
like the minimum to get it going again.  I have installed this patch into the
tree.

-- 
Michael Meissner, IBM
4 Technology Place Drive, MS 2203A, Westford, MA, 01886, USA
meissner@linux.vnet.ibm.com