const-elim-1.c

[Zack Weinberg]

Mark Mitchell <mark@codesourcery.com> writes:

> Zack --
>
> gcc.dg/const-elim-1.c is a test that you added back in May.
>
> It fails on arm-none-elf, and also on i686-pc-linux-gnu, unless you
> use -mtune=pentiumpro.
>
> The issue in both cases relates to this code:
>
> 	const struct S S = { 23, { 1.414, 1.618 }, 0 };
>
> If the compiler chooses to do a block copy from a read-only data into
> "S", then the compiler is unable to eliminate the block copy, and is
> therefore unable to eliminate the constant pool.  The compiler is
> unable to eliminate the block copy because the code at the end of
> init_propagate_block_info that handles SETs to MEMs doesn't fire for
> these block copies.  Ideally, that code would be made smarter, but
> really the elimination should really happen in tree-ssa code, not in
> RTL code anyhow.

Hmm, no, the bug is in the code that decides that 'auto' overrides
'const' in this case; we needn't generate a copy at all.  I will look
into this.

zw