[Bug tree-optimization/94092] Code size and performance degradations after -ftree-loop-distribute-patterns was enabled at -O[2s]+

[wilson at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94092

--- Comment #6 from Jim Wilson <wilson at gcc dot gnu.org> ---
Trying Alex's patch, it doesn't work on this testcase.

One problem is that the loop bound is unknown, so the memset size is estimated
as max 0x1fffffffc bytes.  The code calls
default_use_by_pieces_infrastructure_p which wants the total number of
instructions to be less than SET_RATIO which is 16.  The total number of
instructions is computed as 0x1fffffffc.  So the attempt fails.

There is a contributing problem here that the alignment of the dest was
computed as 8-bits, even though we have a pointer to int which has 32-bit
alignment on a strict alignment system.  That problem happens in
expand_builtin_memset_args which calls get_pointer_alignment on dest.  DEST is
a SSN_NAME, and the SSA_NAME pointer info says align is 0, so it defaults to
8-bit byte alignment.  I haven't tried to figure out where that went wrong.

The patch does work as advertised on libgcc soft-float code.  For an rv32i
build I see the memset calls in libgcc.a reduced from 31 to 10.