The fix for PR82621 makes us not split an I2 if one of the results of
those SETs is unused, since combine does not handle that properly.  But
this results in degradation for i386 (or more in general, for any
target that does not have patterns for parallels with an unused result
as a CLOBBER instead of a SET for that result).

This patch instead makes us not split only if one of the results is set
again before I3.  That fixes PR83156 and also fixes PR82621.

Unfortunately it undoes the nice optimisations that the previous patch
did, on powerpc.

Tested on powerpc64-linux {-m32,-m64}; committing to trunk.


Segher


2017-11-29  Segher Boessenkool  <seg...@kernel.crashing.org>

        PR rtl-optimization/83156
        PR rtl-optimization/82621
        * combine.c (try_combine): Don't split an I2 if one of the dests is
        set again before I3.  Allow unused dests.

---
 gcc/combine.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/gcc/combine.c b/gcc/combine.c
index 8462397..c578e47 100644
--- a/gcc/combine.c
+++ b/gcc/combine.c
@@ -3042,7 +3042,8 @@ try_combine (rtx_insn *i3, rtx_insn *i2, rtx_insn *i1, 
rtx_insn *i0,
       && can_split_parallel_of_n_reg_sets (i2, 2)
       && !reg_used_between_p (SET_DEST (XVECEXP (PATTERN (i2), 0, 0)), i2, i3)
       && !reg_used_between_p (SET_DEST (XVECEXP (PATTERN (i2), 0, 1)), i2, i3)
-      && !find_reg_note (i2, REG_UNUSED, 0))
+      && !reg_set_between_p  (SET_DEST (XVECEXP (PATTERN (i2), 0, 0)), i2, i3)
+      && !reg_set_between_p  (SET_DEST (XVECEXP (PATTERN (i2), 0, 1)), i2, i3))
     {
       /* If there is no I1, there is no I0 either.  */
       i0 = i1;
-- 
1.8.3.1

Reply via email to