A serious -fpic and -fomit-frame-pointer bug in egcs 1.0.3/1.1

[H.J. Lu]

>   > That is fine to put the PIC base register in ebx. But eliminate_regs ()
>   > doesn't know anything about PIC and ebx. When eliminate_regs replaces
>   > ebp with esp, it assumes the top of the stack is for the temporary
>   > variable. But with PIC, the top of the stack is changed to ebx. When
>   > the function epilogue pops ebx, some garbage is put into ebx. You
>   > can see the asm output for yourself.
> I don't see how eliminate_regs plays into this at all.

eliminate_regs called by reload in egcs 1.0.3 around line 979

      for (i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)              
        if (reg_renumber[i] < 0 && reg_equiv_memory_loc[i])
          {                                                      
            rtx x = eliminate_regs (reg_equiv_memory_loc[i], 0, NULL_RTX, 0);

In my test case, reg_equiv_memory_loc[i] is 16(%ebp) and (%esp) is
returned since eliminate_regs knows 16(%ebp) == (%esp). But it is
in the global register allocation. It knows nothing about PIC and
ebx.

Now the fucntion prologue/epilogue do

exp__H1Zd_RCt7complex1ZX01_t7complex1ZX01:
	subl $16,%esp
	pushl %ebx
        call .L22
.L22:
        popl %ebx
        addl $_GLOBAL_OFFSET_TABLE_+[.-.L22],%ebx
	......
	fstpl (%esp)
	......
        popl %ebx
        addl $16,%esp
        ret $4

Please tell me what happens to ebx. It is because the fucntion
prologue/epilogue do something behind the back of the compiler.
eliminate_regs doesn't know it. If we disable -fomit-frame-pointer
for PIC or change the prologue/epilogue to

exp__H1Zd_RCt7complex1ZX01_t7complex1ZX01:
	pushl %ebx
        call .L22
.L22:
        popl %ebx
        addl $_GLOBAL_OFFSET_TABLE_+[.-.L22],%ebx
	subl $16,%esp
	......
	fstpl (%esp)
	......
        addl $16,%esp
        popl %ebx
        ret $4

the bug will disappear. BTW, can we change the prologue/epilogue
like that? Will it break gdb?


H.J.