Re: grokfield miscompiled with -O9

[Alexandre Oliva <oliva at lsd dot ic dot unicamp dot br>]

On Jan 10, 2000, Alexandre Oliva <oliva@lsd.ic.unicamp.br> wrote:

> On Jan  8, 2000, Richard Henderson <rth@cygnus.com> wrote:
>> But I don't understand the intent of the code I've killed, particularly
>> in context of the other places this function is called.  It may be that
>> this code is correct, but that set_nonzero_bits_and_sign_copies shouldn't
>> be using expand_field_assignment.

> How about skipping this simplification just while computing
> nonzero_bits?

> 	* combine.c (expand_field_assignment): Do not discard SUBREGs
> 	while computing nonzero_bits.
	
BTW, the code fragment in i386.c:ix86_expand_setcc got me wondering
why it would be of any help to generate a scratch QI register to store
CC, and whether it had not been coded the way it was because of the
bug fixed (?) in the patch above.  So I experimented removing the
scratch register generation, and a bit of benchmarking seems to imply
it doesn't affect performance at all, or, if it does, it improves it a
little bit.

My ``benchmarks'' were to measure the time the stage2 compiler (built
from the patched sources) took to build stage3, and the time taken to
run the bzip2 testsuite:

gcc stage3 build
before:
real    9m20.331s
user    8m49.780s
sys     0m21.940s

after:
real    9m15.239s
user    8m48.690s
sys     0m20.990s


bzip2 test
before:
real    0m2.219s
user    0m2.090s
sys     0m0.100s

after:
real    0m2.207s
user    0m2.040s
sys     0m0.140s


Is the following patch ok to install? (assuming the previous combine.c
patch is ok too)

Index: gcc/ChangeLog
from  Alexandre Oliva  <oliva@lsd.ic.unicamp.br>
	
	* config/i386/i386.c (ix86_expand_setcc): Do not create a QImode
	scratch register.
	
Index: gcc/config/i386/i386.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/config/i386/i386.c,v
retrieving revision 1.118
diff -u -p -r1.118 i386.c
--- gcc/config/i386/i386.c	2000/01/04 14:44:03	1.118
+++ gcc/config/i386/i386.c	2000/01/10 04:43:40
@@ -4362,12 +4362,7 @@ ix86_expand_setcc (code, unordered, dest
       tmp = gen_rtx_STRICT_LOW_PART (VOIDmode, tmp);
     }
   else if (type == 1)
-    {
-      if (!cse_not_expected)
-	tmp = gen_reg_rtx (QImode);
-      else
-        tmp = gen_lowpart (QImode, dest);
-    }
+    tmp = gen_lowpart (QImode, dest);
 
   emit_insn (gen_rtx_SET (VOIDmode, tmp, ret));
 

-- 
Alexandre Oliva http://www.ic.unicamp.br/~oliva IC-Unicamp, Bra[sz]il
oliva@{lsd.ic.unicamp.br,guarana.{org,com}} aoliva@{acm,computer}.org
oliva@{gnu.org,kaffe.org,{egcs,sourceware}.cygnus.com,samba.org}
** I may forward mail about projects to mailing lists; please use them