On Tue, Jan 13, 2015 at 12:45:27PM -0700, Jeff Law wrote:
> On 01/13/15 09:38, Segher Boessenkool wrote:
> >On Tue, Jan 13, 2015 at 05:18:19PM +0100, Jakub Jelinek wrote:
> >>3) on request from Richard (which Segher on IRC argues against), "memory"
> >> clobber also prevents CSE;
> >
> >As extend.texi used to say:
> >
> >"
> >If your assembler instructions access memory in an unpredictable
> >fashion, add @samp{memory} to the list of clobbered registers. This
> >causes GCC to not keep memory values cached in registers across the
> >assembler instruction and not optimize stores or loads to that memory.
> >You also should add the @code{volatile} keyword if the memory
> >affected is not listed in the inputs or outputs of the @code{asm}, as
> >the @samp{memory} clobber does not count as a side-effect of the
> >@code{asm}.
> >"
> >
> >so a "memory" clobber in a non-volatile asm should not prevent CSE.
> My reading of that paragraph is somewhat different.
>
> The key here is the memory clobber affects optimization of instructions
> around the asm while the volatile specifier affects the optimization of the
> ASM itself.
>
> A memory clobber must inhibit CSE of memory references on either side of the
> asm because the asm must be assumed to read or write memory in unpredictable
> ways.
>
> The volatile specifier tells the compiler that the asm itself must be
> preserved, even if dataflow shows the outputs as not used.
That is not necessarily in conflict.
My reading of Jeff's comment is that in
int a;
int
foo (void)
{
int b, c, d, e;
b = a;
asm ("..." : "=r" (c) : : "memory");
d = a;
asm ("..." : "=r" (e) : : "memory");
return b + d + 2 * (c + e);
}
we are not allowed to CSE d = a; into d = b;. CSE invalidate_from_clobbers
should ensure that already, even when we don't do anything special about
"memory" clobber in the patch. Another thing is if there is a store
in between the two non-volatile asms with "memory" clobber, here I'm not
sure if with the alternate patch we'd treat the "memory" clobber as use of
everything previously stored into memory (in this regard the posted version
is safe).
And finally there is the case of non-volatile asm with "memory" clobber with
no memory stores in between the two - the posted (safer) patch will not
allow to CSE the two, while in theory we could CSE them into just one asm.
Jakub
