In the tree codes and optabs, the "hi" in a vector hi/lo pair means
"most significant" and the "lo" means "least significant", with
sigificance following GCC's normal endian expectations. Thus on
big-endian targets, the hi part handles the first half of the elements
in memory order and the lo part handles the second half.
For tree codes, supportable_widening_operation first chooses hi/lo
pairs based on little-endian order and then uses:
if (BYTES_BIG_ENDIAN && c1 != VEC_WIDEN_MULT_EVEN_EXPR)
std::swap (c1, c2);
to adjust. However, the handling for internal functions was missing
an equivalent fixup. This led to several execution failures in vect.exp
on aarch64_be-elf.
If the hi/lo code fails, the internal function handling goes on to try
even/odd. But I couldn't see anything obvious that would put the even/
odd results back into the right order later, so there might be a latent
bug there too.
Tested on aarch64-linux-gnu & aarch64_be-elf. OK to install?
Richard
gcc/
PR tree-optimization/118891
* tree-vect-stmts.cc (supportable_widening_operation): Swap the
hi and lo internal functions on big-endian targets.
---
gcc/tree-vect-stmts.cc | 2 ++
1 file changed, 2 insertions(+)
diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index 95406b4e3a3..5767a351fdf 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -14410,6 +14410,8 @@ supportable_widening_operation (vec_info *vinfo,
internal_fn lo, hi, even, odd;
lookup_hilo_internal_fn (ifn, &lo, &hi);
+ if (BYTES_BIG_ENDIAN)
+ std::swap (lo, hi);
*code1 = as_combined_fn (lo);
*code2 = as_combined_fn (hi);
optab1 = direct_internal_fn_optab (lo, {vectype, vectype});
--
2.43.0
