A libgomp test demonstrated a problem with reduction code in the Power back end
(https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87064) for little-endian code.
The
define_insn_and_split at the heart of the problem depends on big-endian data
layout,
and should have been disabled when little-endian support was added around 2012.
We missed that. The pattern in question was used for V2DF mode; there is a
similar
pattern for V4SF mode that should also have been disabled. This patch disables
both patterns. Thanks to Jakub for analyzing the problem!
Bootstrapped and tested on powerpc64le-unknown-linux-gnu with no regressions.
The failing testcase now passes. Is this okay for trunk, and eventually for
backports to all supported releases?
Thanks,
Bill
2019-01-22 Jakub Jelinek <ja...@redhat.com>
Bill Schmidt <wschm...@linux.ibm.com>
* config/rs6000/vsx.md (*vsx_reduc_<VEC_reduc_name>_v2df_scalar):
Disable for little endian.
(*vsx_reduc_<VEC_reduc_name>_v4sf_scalar): Likewise.
Index: gcc/config/rs6000/vsx.md
===================================================================
--- gcc/config/rs6000/vsx.md (revision 268115)
+++ gcc/config/rs6000/vsx.md (working copy)
@@ -4351,7 +4351,7 @@
(match_dup 1))
(parallel [(const_int 1)])))
(clobber (match_scratch:DF 2 "=0,0,&wd,&wa"))]
- "VECTOR_UNIT_VSX_P (V2DFmode)"
+ "BYTES_BIG_ENDIAN && VECTOR_UNIT_VSX_P (V2DFmode)"
"#"
""
[(const_int 0)]
@@ -4378,7 +4378,7 @@
(clobber (match_scratch:V4SF 2 "=&wf,&wa"))
(clobber (match_scratch:V4SF 3 "=&wf,&wa"))
(clobber (match_scratch:V4SF 4 "=0,0"))]
- "VECTOR_UNIT_VSX_P (V4SFmode)"
+ "BYTES_BIG_ENDIAN && VECTOR_UNIT_VSX_P (V4SFmode)"
"#"
""
[(const_int 0)]
