This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Profiling on S390
- From: Richard Henderson <rth at redhat dot com>
- To: Ulrich Weigand <Ulrich dot Weigand at de dot ibm dot com>
- Cc: Jan Hubicka <jh at suse dot cz>, gcc at gcc dot gnu dot org, aj at suse dot de
- Date: Wed, 23 Apr 2003 14:02:23 -0700
- Subject: Re: Profiling on S390
- References: <OF0AE7CC2A.662B03BB-ONC1256D11.006F098A@de.ibm.com>
On Wed, Apr 23, 2003 at 10:23:24PM +0200, Ulrich Weigand wrote:
> However, in the case of the profiler where saving/restoring CC
> is vital for *correctness*, shouldn't the fact that CCmode copies
> are *possible* (i.e. movcc works) be enough?
Well... if movcc exists, then fine. But if it doesn't, and
movsi doesn't support the flags register, then we're going
to generate an unreloadable insn.
OTOH, I suppose it's sufficient to abort in reload.
> If I were to introduce that capability, does the common code make
> any assumptions on just how those values are represented by integers?
> Does (subreg:SI (reg:CC ...) 0) or vice versa work?)
If a movcc pattern doesn't exist, then emit_move_insn_1 will subreg
to word_mode and try that move pattern. But no assumptions are made
wrt the contents of the bits.
r~