Bug 96923 - Failure to optimize a select-related bool pattern to or+not
Summary: Failure to optimize a select-related bool pattern to or+not
Status: RESOLVED DUPLICATE of bug 89263
Alias: None
Product: gcc
Classification: Unclassified
Component: tree-optimization (show other bugs)
Version: 11.0
: P3 enhancement
Target Milestone: ---
Assignee: Andrew Pinski
URL:
Keywords: missed-optimization
Depends on: 25290 100864
Blocks: 19987 20083 94898 105903
  Show dependency tree / graph
 
Reported: 2020-09-03 14:06 UTC by Gabriel Ravier
Modified: 2023-08-02 14:01 UTC (History)
2 users (show)

See Also:
Host:
Target:
Build:
Known to work:
Known to fail:
Last reconfirmed: 2021-06-01 00:00:00

Attachments
Patch which is in testing (needs testcases) (922 bytes, patch)
2021-06-01 21:40 UTC, Andrew Pinski 		Details | Diff
View All Add an attachment (proposed patch, testcase, etc.)

Note You need to log in before you can comment on or make changes to this bug.
Description Gabriel Ravier 2020-09-03 14:06:08 UTC 
_Bool f2(_Bool a, _Bool b)
{
    return !a ? !b : 0;
}

This can be optimized to `return !(a | b);`. This transformation is done by LLVM, but not by GCC.
Comment 1 Gabriel Ravier 2020-09-03 14:49:21 UTC 
_Bool f2(_Bool a, _Bool b)
{
    return a ? !b : 1;
}

This similar pattern can be optimized to `return !(a & b);`. This transformation is done by LLVM, but not by GCC.
Comment 2 Andrew Pinski 2020-09-06 23:47:26 UTC 
Note if we spell out the ?:, this would require PR 25290 too.
Comment 3 Andrew Pinski 2021-06-01 21:02:44 UTC 
Mine, I have a patch which implements this.
It needs https://gcc.gnu.org/pipermail/gcc-patches/2021-June/571616.html first.
I will be posting this once I write some testcases and do a bootstrap/test cycle.
Comment 4 Andrew Pinski 2021-06-01 21:40:22 UTC 
Created attachment 50905 [details]
Patch which is in testing (needs testcases)

As I said for the case in this PR, it needs https://gcc.gnu.org/pipermail/gcc-patches/2021-June/571616.html too.

If you change !b to b; that is "!a ? b : 0", this patch will optimize it.  The other patch is needed to allow to move !b out of the conditional.
Comment 5 Andrew Pinski 2021-06-01 23:19:05 UTC 
(In reply to Andrew Pinski from comment #4)
> Created attachment 50905 [details]
> Patch which is in testing (needs testcases)
> 
> As I said for the case in this PR, it needs
> https://gcc.gnu.org/pipermail/gcc-patches/2021-June/571616.html too.
> 
> If you change !b to b; that is "!a ? b : 0", this patch will optimize it. 
> The other patch is needed to allow to move !b out of the conditional.

This patch has one bug in it where we need a convert added.
BUT then we run into a different missed optimization issue.
  _51 = p2_22 <= prephitmp_122;
  _44 = (logical(kind=4)) _51;
  _37 = p2_22 > prephitmp_122;
  _49 = (logical(kind=4)) _37;
  _38 = _49 & found_p_63;
  _46 = _38 | _44;

This is really just:
_51 = p2_22 <= prephitmp_122;
_44 = (logical(kind=4)) _51;
_46 = found_p_63| _44;

That is we don't optimize:
(a & ~b) | b into a | b if ~b has been converted already.

The other thing I noticed is the cast should not be really needed but nothing removes it; I will look at that later.

Note I could rewrite the pattern to do the simplification of the constants manually but I want to try to avoid that.
Comment 6 Andrew Pinski 2021-06-01 23:49:10 UTC 
This depends on PR 100864 if I don't want to write out the 4 patterns.
Comment 7 Andrew Pinski 2023-05-24 00:01:24 UTC 
Ok, this in the end is a dup of bug 89263.

*** This bug has been marked as a duplicate of bug 89263 ***