https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90811
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Alexander Monakov from comment #3)
> Thanks. The change in the attached patch looks good to me, but I must admit
> I don't see how the testcase triggers the problem. Basically, it's not
> obvious how the controlling if condition becomes true:
>
> > if (!CONST_INT_P (size) || UINTVAL (align) > GET_MODE_SIZE (DImode))
>
> (I don't expect that loop body to have variable-sized or over-aligned
> objects)
From the dumps it looks like overaligned objects.
Debugging this in the system compiler, as I don't have nvptx offloading enabled
build of current trunk.
lower_rec_simd_input_clauses is called for d with:
<var_decl 0x7ffff7ffbd80 d
type <array_type 0x7fffeaab5f18
type <integer_type 0x7fffeaab5c78 long long int readonly
sizes-gimplified DI
size <integer_cst 0x7fffea976cf0 constant 64>
unit-size <integer_cst 0x7fffea976d08 constant 8>
align:64 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type
0x7fffeaab5c78 precision:64 min <integer_cst 0x7fffea976fa8
-9223372036854775808> max <integer_cst 0x7fffea976fd8 9223372036854775807>>
sizes-gimplified BLK
size <integer_cst 0x7fffeaab8288 constant 320>
unit-size <integer_cst 0x7fffeaab85a0 constant 40>
align:64 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type
0x7fffeaab5f18
domain <integer_type 0x7fffeaab5e70 type <integer_type 0x7fffea98e000
sizetype>
sizes-gimplified DI size <integer_cst 0x7fffea976cf0 64> unit-size
<integer_cst 0x7fffea976d08 8>
align:64 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type
0x7fffeaab5e70 precision:64 min <integer_cst 0x7fffea976d20 0> max <integer_cst
0x7fffea976f48 4>>>
readonly addressable used read BLK pr90811.c:15:23 size <integer_cst
0x7fffeaab8288 320> unit-size <integer_cst 0x7fffeaab85a0 40>
align:64 warn_if_not_align:0 context <function_decl 0x7fffeaab7700 main>>
so that var only needs 64-bit alignment.
For non-simt, when we create
tree atype = build_array_type_nelts (TREE_TYPE (new_var), sctx->max_vf);
tree avar = create_tmp_var_raw (atype);
I'm quite sure that kicks in the i386.c (ix86_data_alignment):
/* x86-64 ABI requires arrays greater than 16 bytes to be aligned
to 16byte boundary. */
if (TARGET_64BIT)
{
if ((opt ? AGGREGATE_TYPE_P (type) : TREE_CODE (type) == ARRAY_TYPE)
&& TYPE_SIZE (type)
&& TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST
&& wi::geu_p (wi::to_wide (TYPE_SIZE (type)), 128)
&& align < 128)
return 128;
}
kicks in that case; not sure where we create a temporary with ARRAY_TYPE in the
simt case. Might be nice to force we don't add excessive alignment to those
through DECL_USER_ALIGN and copying alignment from the type or something
similar, unless the user variable is already DECL_USER_ALIGN.
