https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93512
Bug ID: 93512
Summary: Introduce rotate_truncation_mask
Product: gcc
Version: unknown
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: dje at gcc dot gnu.org
CC: jakub at gcc dot gnu.org, segher at gcc dot gnu.org
Depends on: 93505
Target Milestone: ---
<jakub> i386.md has ;; Avoid useless masking of count operand.
<jakub> (define_insn_and_split "*<rotate_insn><mode>3_mask"
<jakub> and (define_insn_and_split "*<rotate_insn><mode>3_mask_1"
<jakub> for this purpose
<segher> yeah... but that catches not all cases at all. but the common ones,
sure
<jakub> the other option is to say that in RTL all rotate counts are well
defined and so effectively have implicit ROTATE_COUNT_TRUNCATED
<jakub> or introduce rotate_truncation_mask with a default that will truncate
and let weirdo targets opt out
<segher> yes, i think that would help all targets
<jakub> in any case, that looks like GCC 11 material to me
<jakub> because one would need to check all code that works with rotates, do
the masking, ensure simplify-rtx.c simplifies it etc.
<jakub> would need to be a property of ROTATE/ROTATERT, because e.g. for
library implementation if it is done using shifts we don't want to introduce UB
there
<jakub> so with the combine.c change you've acked perhaps we could defer the
rest for GCC11 (and at that point change even that hunk to do % prec
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93505
[Bug 93505] [8/9/10 Regression] wrong code or ICE with __builtin_bswap64() and
rotation at -Og
