[Bug middle-end/44505] [4.6 Regression] gcc.c-torture/execute/frame-address.c

["hubicka at ucw dot cz" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #4 from hubicka at ucw dot cz  2010-06-16 16:13 -------
Subject: Re:  [4.6 Regression]
        gcc.c-torture/execute/frame-address.c

> Confirmed on powerpc-linux.  check_fa tail calls check_fa_mid, ignoring the
> fact that check_fa_mid is passed the address of a check_fa local var.

This seems like valid transform to me.  check_fa_mid is recognized as const so
caller know that even if local variable is passed, it is not going to read it.
So i would say that the testcase is semi-invalid at least.

What about adding a asm statement clobbering memory to check_fa_work
to avoid the failure?

On the other hand I was worng to assume that built_in_frame_address does not
make function not-const or pure since it returns address of local frame
that is useless after returning from function.  With parameter greater than
0 it is not true.
I wonder if we care enough about builtin_eh_frame to make it worthwhile to
special case this in ipa-pure-const? It does not 100% fit into the framework
since the calling statement is available only to non-WPA compilation.

Honza


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44505