Re: bootstrap failure, x86: stage1 miscompiles gengenrtl

[Alan Modra <alan at linuxcare dot com dot au>]

On Thu, 30 Mar 2000, Zack Weinberg wrote:

> stage1/xgcc -Bstage1/ -B/work/inst/i686-pc-linux-gnu/bin/ -c -DIN_GCC
>     -W -Wall -Wtraditional -O2 -g -W -Wall -Wtraditional -DHAVE_CONFIG_H
>     -I. -I/work/src/hashtab.gcc/gcc -I/work/src/hashtab.gcc/gcc/config
>     -I/work/src/hashtab.gcc/gcc/../include
>     /work/src/hashtab.gcc/gcc/gengenrtl.c
> stage1/xgcc -Bstage1/ -B/work/inst/i686-pc-linux-gnu/bin/ -DIN_GCC -W
>      -Wall -Wtraditional -O2 -g -W -Wall -Wtraditional -DHAVE_CONFIG_H 
>      -o gengenrtl gengenrtl.o obstack.o
> ./gengenrtl -h >tmp-genrtl.h
> Segmentation fault
> 
> $ catchsegv ./gengenrtl -h
> 
> ...
> extern rtx gen_rtx_fmt_uuuu     PARAMS ((RTX_CODE, enum machine_mode mode,
>                                        rtx arg117, rtx arg117, rtx arg117,
>                                        rtx arg117));
> 
> *** Segmentation fault
> Backtrace:
> libc.so.6(_IO_vfprintf+0x17df)
> libc.so.6(printf+0x2a)
> gengenrtl.c:264(genmacro)
> gengenrtl.c:344(genheader)
> gengenrtl.c:405(main)
> 
> Line 264 of gengenrtl.c is
> 
>   printf ("#define gen_rtx_%s%s(MODE",
>            special_rtx (idx) ? "raw_" : "", defs[idx].enumname);
> 
> corresponding to this hunk of assembly output:
> 
> 
>         movl    8(%ebp), %edi
>         .stabn 68,0,261,.LM74-genmacro
> .LM74:
>         leal    (%edi,%edi,2), %eax
>         leal    0(,%eax,4), %ebx
>         movl    defs(%ebx), %eax
>         pushl   %eax
>         subl    $12, %esp

The above two instructions seem to be in the wrong order.  That's likely
your segv problem.

>         pushl   %edi
>         call    special_rtx
>         testl   %eax, %eax
>         movl    $.LC125, %eax
>         je      .L362
>         movl    $.LC340, %eax
> .L362:
>         pushl   %eax
>         .stabn 68,0,264,.LM75-genmacro
> .LM75:
>         xorl    %esi, %esi
>         .stabn 68,0,261,.LM76-genmacro
> .LM76:
>         pushl   $.LC341
>         call    printf
> ...
> .LC125:
> 	.string ""
> .LC340:
> 	.string "raw_"
> .LC341:
> 	.string "#define gen_rtx_%s%s(MODE"
> 
> Given the recent sizing changes, I am inclined to suspect this chunk
> of code:
> 
>         movl    8(%ebp), %edi
>         leal    (%edi,%edi,2), %eax
>         leal    0(,%eax,4), %ebx
>         movl    defs(%ebx), %eax
>         pushl   %eax
> 
> That's supposed to calculate defs[idx].enumname.  The analogous code
> generated by gcc 2.95 reads
> 
> 	movl	8(%ebp), %edi
> 	addl	%edi, %edi
> 	movl	8(%ebp), %edx
> 	leal	(%edx,%edi), %eax
> 	leal	0(,%eax,4), %esi
> 	movl	$defs, %ebx
> 	pushl	(%esi, %ebx)
> 
> I'm not exactly an x86 assembly expert, but it appears to me that the
> code generated by 2.95 is multiplying idx by 3 before the 
> leal 0(,%eax,4) instruction, while the current tree is multiplying it
> by 4.  sizeof(*defs) is 12; both versions agree on that if explicitly
> asked.

Nup.  "leal (%edi,%edi,2), %eax" does eax = edi + edi * 2

--
Linuxcare.  Support for the Revolution.