https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84100
--- Comment #6 from Chris Hall <gcc at gmch dot uk> --- Created attachment 43325 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=43325&action=edit Summary of results using Compiler Explorer: for v5.4.0, 6.3.0 and 7.2.0. Setting -O2 or -O3 on the command line appears to imply -falign-functions=16 and -falign-loops=n, where n is at least 10 but less than 14. Setting -falign-functions=32 on the command line does the requested alignment, plus the implied -falign-loops=n for -O2 and -O3 -- except for -Os, which does no alignment Setting both -falign-functions=32 and -falign-loops=32 on the command line does both requested alignments -- except for -Os which does neither and -O0 which does only the function alignment. So far, so good. One might hope that using _Pragma() would have the same effect as the command line -falign-functions=32 and -falign-loops=32. Sadly, this is not the case. As reported, v7 rejects the pragmas. But I find that v5.4.0 and v6.3.0, (a) don't always do the same thing and (b) don't always align loops as requested. In particular: for -O2 and -O3, setting both align-functions=32 and align-loops=32 by _Pragma() will: align functions as requested, BUT loops are aligned to 8 (which is less than useful) !! The last set of results looks at compiling with command line -O0 through to -O3 (including -Og and -Os), with _Pragma() for -O3, align-functions=32 and align-loops=32. The _Pragma() for -O3 seems to override the command line as far as this small test is concerned, except for -fomit-frame-pointer. Sadly, align-loops=32 is still broken.
