[PATCH] SHF_MERGE support for gcc (take 4)

[Jakub Jelinek]

On Mon, Sep 10, 2001 at 10:15:15AM -0700, Richard Henderson wrote:
> On Mon, Sep 10, 2001 at 05:37:07PM +0200, Jakub Jelinek wrote:
> > Bootstrapped on i386-redhat-linux, one regression (buggy testcase
> > execute/20000801-4.c).
> 
> The test case is not buggy.  How are you treating, say, "\0\0"?
> It should be the same.  I suspect that strings with embedded
> nulls simply cannot be merged.

My fault, I realized this too. I remembered this from when 20000801-4.c used to
fail and SHF_MERGE patch actuall made it XPASS or something like that.
Working on a fix.

> > +  /* Align the location counter as required by EXP's data type.  */
> > +  align = TYPE_ALIGN (TREE_TYPE (exp));
> > +#ifdef CONSTANT_ALIGNMENT
> > +  align = CONSTANT_ALIGNMENT (exp, align);
> 
> -Os should ignore extra alignment.

This was just code move, but I can certainly kill it for -Os.

> > +	echo '.section .rodata.str, "ams", @progbits, 1' > conftest.s
> 
> "s"?  That means short data on a number of targets...

Does it?
When I was adding this support to gas, the only port which was using
md_elf_section_letter was ppc (letter 'e'), rechecking now i370 uses 'e'
too.
Other than that obj-elf.c handles 'a' for SHF_ALLOC, 'w' for SHF_WRITE, 'x'
for SHF_EXECINSTR, 'm' for SHF_MERGE and 's' for SHF_STRINGS.

> >  #undef  EXTRA_SECTIONS
> > +#ifdef HAVE_GAS_SHF_MERGE
> > +#define EXTRA_SECTIONS in_const,				\
> > +  in_const_cst1, in_const_cst2, in_const_cst4,			\
> 
> No, just use named_section_flags.  This should go in varasm.c
> and not replicated in all relevant target headers.

Will try.

> > +      switch (align)							\
> > +	{								\
> > +	case 8: in = in_const_cst1; break;				\
> > +	case 16: in = in_const_cst2; break;				\
> > +	case 32: in = in_const_cst4; break;				\
> > +	case 64: in = in_const_cst8; break;				\
> > +	case 128: in = in_const_cst16; break;				\
> 
> What about an object larger than its alignment?  Won't merging
> wind up treating it as two objects?

GET_MODE_BITSIZE (mode) <= align should catch that.

	Jakub