egcs-1.1.1 bootstrap-failure on i386-next-nextstep3

[Eric Norum]

I've discovered what's going wrong with egcs-1.1.1 on OPENSTEP/Intel.

Symptoms
========
The gencheck program compiled with stage1/xgcc does not run  
(Segmentation fault).


Investigation
=============
A little poking with gdb shows that the program is dying on line 53:
53        for (i = 0; tree_codes[i]; i++)

Disassembling the instructions shows:
0x291e <main+34>:       addl    $0x04,%esp
0x2921 <main+37>:       cmpl    $0x00,tree_codes
0x2928 <main+44>:       je      0x00002957
0x292a <main+46>:       movl    $tree_codes,%ebx
0x292f <main+51>:       addb    %cl,0x68505003(%ebx)  <<=== ???
0x2935 <main+57>:       pushal
0x2936 <main+58>:       aas

Total garbage!


Identification
==============
Looking at the assembler code generated by the stage 1 compiler reveals:
        .stabd 68,0,53
        addl $4,%esp
        cmpl $0,_tree_codes
        je L95
        movl $_tree_codes,%ebx
        .align 2                   <<<<<======= !!!!!!!!!!!
L97:
        .stabd 68,0,55
        movl (%ebx),%eax
        pushl %eax
        pushl %eax

What's an `align' doing in the middle of an instruction  
sequence??!!!  It's forcing an extra byte between the end of the  
`movl $_tree_codes,%ebx' instruction and the beginning of the `movl  
(%ebx),%eax' instruction.  This, of course, completely messes up the  
instructions seen by the processor.


The solution
============
This I don't have :-(
Perhaps an egcs expert out there can figure out what needs to be done.

---
Eric Norum                                 eric@skatter.usask.ca
Saskatchewan Accelerator Laboratory        Phone: (306) 966-6308
University of Saskatchewan                 FAX:   (306) 966-6058
Saskatoon, Canada.