https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127577
--- Comment #4 from Sebastian Büttner <mail-gcc at bueddl dot de> --- I bisected between 14.2.0 and 14.3.0 and found the commit that triggered the regression to be: commit 67379c5b6ea4c87d69dea90ede51f33c9f5170c8 (HEAD) Author: Jakub Jelinek <[email protected]> Date: Wed Nov 6 10:21:09 2024 +0100 store-merging: Don't use sub_byte_op_p mode for empty_ctor_p unless necessary [PR117439] encode_tree_to_bitpos uses the more expensive sub_byte_op_p mode in which it has to allocate a buffer and do various extra work like shifting the bits etc. if bitlen or bitpos aren't multiples of BITS_PER_UNIT, or if bitlen doesn't have corresponding integer mode. The last case is explained later in the comments: /* The native_encode_expr machinery uses TYPE_MODE to determine how many bytes to write. This means it can write more than ROUND_UP (bitlen, BITS_PER_UNIT) / BITS_PER_UNIT bytes (for example write 8 bytes for a bitlen of 40). Skip the bytes that are not within bitlen and zero out the bits that are not relevant as well (that may contain a sign bit due to sign-extension). */ Now, we've later added empty_ctor_p support, either {} CONSTRUCTOR or {CLOBBER}, which doesn't use native_encode_expr at all, just memset, so that case doesn't need those fancy games unless bitlen or bitpos aren't multiples of BITS_PER_UNIT (unlikely, but let's pretend it is possible). The following patch makes us use the fast path even for empty_ctor_p which occupy full bytes, we can just memset that in the provided buffer and don't need to XALLOCAVEC another buffer. This patch in itself fixes the testcase from the PR (which was about using huge XALLLOCAVEC), but I want to do some other changes, to be posted in a next patch. 2024-11-06 Jakub Jelinek <[email protected]> PR tree-optimization/117439 * gimple-ssa-store-merging.cc (encode_tree_to_bitpos): For empty_ctor_p use !sub_byte_op_p even if bitlen doesn't have an integral mode. (cherry picked from commit aab572240a0752da74029ed9f8918e0b1628e8ba) diff --git a/gcc/gimple-ssa-store-merging.cc b/gcc/gimple-ssa-store-merging.cc index 7dba4a7a781..429c18c1c78 100644 --- a/gcc/gimple-ssa-store-merging.cc +++ b/gcc/gimple-ssa-store-merging.cc @@ -1933,14 +1933,15 @@ encode_tree_to_bitpos (tree expr, unsigned char *ptr, int bitlen, int bitpos, unsigned int total_bytes) { unsigned int first_byte = bitpos / BITS_PER_UNIT; - bool sub_byte_op_p = ((bitlen % BITS_PER_UNIT) - || (bitpos % BITS_PER_UNIT) - || !int_mode_for_size (bitlen, 0).exists ()); bool empty_ctor_p = (TREE_CODE (expr) == CONSTRUCTOR && CONSTRUCTOR_NELTS (expr) == 0 && TYPE_SIZE_UNIT (TREE_TYPE (expr)) - && tree_fits_uhwi_p (TYPE_SIZE_UNIT (TREE_TYPE (expr)))); + && tree_fits_uhwi_p (TYPE_SIZE_UNIT (TREE_TYPE (expr)))); + bool sub_byte_op_p = ((bitlen % BITS_PER_UNIT) + || (bitpos % BITS_PER_UNIT) + || (!int_mode_for_size (bitlen, 0).exists () + && !empty_ctor_p)); if (!sub_byte_op_p) { This could make sense (?) since the added `!empty_ctor_p` check causes selection of a different branch for this case. The commit itself is a cherry-pick of aab572240a0752da74029ed9f8918e0b1628e8ba and also has been applied to gcc 15 and 16 as well so it makes sense these are affected. This has also been cherry-picked as bf85c4df809 which affcts 13.4.0 and 13.5.0. I did not check 12.x.
