Re: Is it a gcc bug? Optimization doesn't work

[lapo <lapo dot pasqui at tele2 dot it>]

Alle 14:10, mercoledì 15 settembre 2004, lapo.pasqui@tele2.it ha scritto:
Does anybody like to answer this question?
I'd rather prefer not to open an insane bug against the gcc.
Thanks
  Lapo
> Hi all,
> I've upgraded my cross gcc compiler from version 3.3.1 to 3.4.1 and I've
> noticed a few unresolved symbols on a project (u-boot) used to compile
> without any problem. After a quick investigation I've noticed the
> optimization flag -0s caused a few functions not to be in the .o files.
> Here's a code that reproduces the problem
>
> unsigned long get_version(void);
> int getc(void);
> int tstc(void);
> void putc(const char);
>
> void app_startup(char **);
>
> enum {
> XF_get_version ,
> XF_getc ,
> XF_tstc ,
> XF_putc ,
>  XF_MAX
> };
>
> typedef unsigned int size_t;
> typedef struct global_data
> {
>    void **jt;
> } gd_t;
>
> static void __attribute__((unused)) dummy(void)
> {
> asm volatile ( "        .globl " "get_version" "\n" "get_version" ":\n" "  
>     lwz     %%r11, %0(%%r29)\n" "   lwz     %%r11, %1(%%r11)\n" "   mtctr  
> %%r11\n" "      bctr\n" : : "i"(((size_t) &((gd_t *)0)->jt)),
> "i"(XF_get_version * sizeof(void *)) : "r11"); asm volatile ( "       
> .globl " "getc" "\n" "getc" ":\n" "     lwz     %%r11, %0(%%r29)\n" "   lwz
>     %%r11, %1(%%r11)\n" "   mtctr   %%r11\n" "      bctr\n" : :
> "i"(((size_t) &((gd_t *)0)->jt)), "i"(XF_getc * sizeof(void *)) : "r11");
> asm volatile ( "        .globl " "tstc" "\n" "tstc" ":\n" "     lwz    
> %%r11, %0(%%r29)\n" "   lwz     %%r11, %1(%%r11)\n" "   mtctr   %%r11\n" " 
>     bctr\n" : : "i"(((size_t) &((gd_t *)0)->jt)), "i"(XF_tstc * sizeof(void
> *)) : "r11"); asm volatile ( "        .globl " "putc" "\n" "putc" ":\n" "  
>   lwz     %%r11, %0(%%r29)\n" "   lwz     %%r11, %1(%%r11)\n" "   mtctr  
> %%r11\n" "      bctr\n" : : "i"(((size_t) &((gd_t *)0)->jt)), "i"(XF_putc *
> sizeof(void *)) : "r11"); }
>
> extern unsigned long __bss_start, _end;
>
> void app_startup(char **argv)
> {
>  unsigned long * cp = &__bss_start;
>  while (cp < &_end) {
>   *cp++ = 0;
>  }
>
> }
>
> Compiling it with optimization less then 2 (e.g. powerpc-linux-gcc  stubs.i
> -c -O1), I get this symbols
>
> nm -C --defined-only stub.o
>
> 00000044 T app_startup
> 00000000 t dummy
> 00000010 T getc
> 00000000 T get_version
> 00000030 T putc
> 00000020 T tstc
>
>
> Instead, using a different optimization (e.g. powerpc-linux-gcc  stubs.i -c
> -O2), I get this symbols
>
> nm -C --defined-only stub.o
> 00000000 T app_startup
>
> Using a previous version of gcc, I've not this different behaviour
>
> My questions are (as I'm not am assembler guru):
>
> Can you see anything wrong on this code that some new feature of the gcc
> try to overcome? Is it indeed a regression of gcc 3.4.1
>
>
>
>
> -------------------------------------------------
> WebMail Tele2 http://www.tele2.it
> -------------------------------------------------