https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113572
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rsandifo at gcc dot gnu.org
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
vect_all_same before r14-5880-g31e9074977bb7de83fa5d28d286323987d5d87f2
wouldn't ICE on this, that did
int nelts = vector_cst_encoded_nelts (v);
tree first_el = VECTOR_CST_ENCODED_ELT (v, 0);
for (i = 0; i < nelts; i += step)
if (!operand_equal_p (VECTOR_CST_ENCODED_ELT (v, i), first_el, 0))
return false;
and so i was never higher than vector_cst_encoded_nelts,
but the current generalized vector_cst_all_same doesn't guarantee that:
unsigned int lcm = least_common_multiple (step, VECTOR_CST_NPATTERNS (v));
unsigned int nelts = lcm * VECTOR_CST_NELTS_PER_PATTERN (v);
tree first_el = VECTOR_CST_ENCODED_ELT (v, 0);
for (unsigned int i = 0; i < nelts; i += step)
if (!operand_equal_p (VECTOR_CST_ENCODED_ELT (v, i), first_el, 0))
return false;
because least_common_multiple of something and VECTOR_CST_NPATTERNS will often
be
higher than VECTOR_CST_NPATTERNS (unless VECTOR_CST_NPATTERNS is a multiple of
step).
So, if that part is right, I think we want to use VECTOR_CST_ELT instead of
VECTOR_CST_ENCODED_ELT, like:
--- gcc/config/aarch64/aarch64-sve-builtins.cc.jj 2024-01-12
13:47:20.815429012 +0100
+++ gcc/config/aarch64/aarch64-sve-builtins.cc 2024-01-24 20:58:33.720677634
+0100
@@ -3474,7 +3474,7 @@ vector_cst_all_same (tree v, unsigned in
unsigned int nelts = lcm * VECTOR_CST_NELTS_PER_PATTERN (v);
tree first_el = VECTOR_CST_ENCODED_ELT (v, 0);
for (unsigned int i = 0; i < nelts; i += step)
- if (!operand_equal_p (VECTOR_CST_ENCODED_ELT (v, i), first_el, 0))
+ if (!operand_equal_p (VECTOR_CST_ELT (v, i), first_el, 0))
return false;
return true;
which fixes the ICE.
On the testcase, VECTOR_CST_NELTS_PER_PATTERN is 2, VECTOR_CST_NPATTERNS is 1
and
step is 8, so lcm is 8 and nelts 16, while the vector has only 2 encoded elts.
