[Bug c/78853] New: aligned reads/writes (vmovdqa) emitted when no such guarantee can be made
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78853 Bug ID: 78853 Summary: aligned reads/writes (vmovdqa) emitted when no such guarantee can be made Product: gcc Version: 7.0 Status: UNCONFIRMED Severity: normal Priority: P3 Component: c Assignee: unassigned at gcc dot gnu.org Reporter: gonnet at google dot com Target Milestone: --- I have an example on https://godbolt.org/g/qZ4SuT, this produces incorrect code since at least gcc-4.9.1 (didn't test earlier). I have a function that swaps to pieces of memory using vectorized read/write instructions. When I pass it a 28-byte struct that has been padded to 32 bytes, it emits two vmovdqa reads and writes. The problem is that it does this even when it can't guarantee that the pointers provided to it are aligned correctly.
[Bug target/78853] aligned reads/writes (vmovdqa) emitted when no such guarantee can be made
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78853 --- Comment #3 from Pedro Gonnet --- OK, thanks for clarifying! The declaration of __m256i only specifies the attributes vector_size and may_alias (https://gcc.gnu.org/git/?p=gcc.git;a=blob;f=gcc/config/i386/avxintrin.h;hb=HEAD#l56), so I'm guessing vector_size implies alignment? This does not seem to be documented anywhere obvious. Also interesting is that icc-17 produces unaligned load/stores for the same code, but there is no amount of __attribute__((aligned(32))) that will make them aligned. In any case I'll use the __m*_u types to make an unaligned version of my memswap function.
