[Bug bootstrap/45445] [4.6 regression] ARM bootstrap failure: comparison failures after stage 3

[mikpe at it dot uu dot se]

------- Comment #7 from mikpe at it dot uu dot se  2010-09-09 10:21 -------
It's not a stage2/stage3 debug difference as far as I can tell.  I've
recompiled every differing .o file with the stage 1/2/3 xgccs -fcompare-debug
without complaints.

The test case showing the different code generation is trivial (this is in the
objdir for a native bootstrap of r163951):

> cat sreal.c
void normalize(unsigned long long *sreal_sig)
{
    *sreal_sig <<= 1;
}
> stage1-gcc/xgcc -Bstage1-gcc -O2 -c sreal.c; objdump -d sreal.o

sreal.o:     file format elf32-littlearm


Disassembly of section .text:

00000000 <normalize>:
   0:   e1c020d0        ldrd    r2, [r0]
   4:   e0922002        adds    r2, r2, r2
   8:   e0a33003        adc     r3, r3, r3
   c:   e1c020f0        strd    r2, [r0]
  10:   e12fff1e        bx      lr
> stage2-gcc/xgcc -Bstage2-gcc -O2 -c sreal.c ; objdump -d sreal.o

sreal.o:     file format elf32-littlearm


Disassembly of section .text:

00000000 <normalize>:
   0:   e1c020d0        ldrd    r2, [r0]
   4:   e1b02082        lsls    r2, r2, #1
   8:   e0a33003        adc     r3, r3, r3
   c:   e1c020f0        strd    r2, [r0]
  10:   e12fff1e        bx      lr
> stage3-gcc/xgcc -Bstage3-gcc -O2 -c sreal.c ; objdump -d sreal.o

sreal.o:     file format elf32-littlearm


Disassembly of section .text:

00000000 <normalize>:
   0:   e1c020d0        ldrd    r2, [r0]
   4:   e1b02082        lsls    r2, r2, #1
   8:   e0a33003        adc     r3, r3, r3
   c:   e1c020f0        strd    r2, [r0]
  10:   e12fff1e        bx      lr

So stage1 chooses adds but stage2 and stage3 choose lsls for << of the lower
half of a long long.  Since the behaviour of a stageN xgcc depends on both the
gcc source code and the compiler used to build it, I have to suspect a source
code ambiguity (e.g. evaluation order dependence) that the bootstrap compiler
(gcc-4.4.4 in my case) resolves differently from post-r162417 4.6.

I've so far not been able to reproduce this difference in a cross from i686;
those cross compilers always seem to choose the adds form regardless of the
version of gcc I'm building with.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45445