[Bug tree-optimization/67328] New: range test rather than single bit test for code testing enum values

[Bug tree-optimization/67328] New: range test rather than single bit test for code testing enum values

[amodra at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67328

            Bug ID: 67328
           Summary: range test rather than single bit test for code
                    testing enum values
           Product: gcc
           Version: 5.1.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: amodra at gmail dot com
  Target Milestone: ---

The attached file generates ideal code, a test of a single bit, for the
condition in test_pic or test_exe depending on -DALT.  gcc ought to be able to
optimise to a single bit test for both functions whether or not -DALT.  It
looks like gcc incorrectly prefers a range test.

[Bug tree-optimization/67328] range test rather than single bit test for code testing enum values

[amodra at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67328

--- Comment #1 from Alan Modra <amodra at gmail dot com> ---
Created attachment 36247
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=36247&action=edit
testcase from binutils/include/bfdlink.h

[Bug tree-optimization/67328] range test rather than single bit test for code testing enum values

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67328

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
   Last reconfirmed|                            |2015-08-25
     Ever confirmed|0                           |1

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Note this is already fold-const.c:optimize_bit_field_compare at work.  With
-DALT (non-working code) we get

;; Function test_pic (null)
;; enabled by -tree-original


{
  if ((BIT_FIELD_REF <*info, 8, 0> & 3) + 254 <= 1)

and

;; Function test_exe (null)
;; enabled by -tree-original


{
  if ((SAVE_EXPR <BIT_FIELD_REF <*info, 8, 0> & 3>) == 0 || (SAVE_EXPR
<BIT_FIELD_REF <*info, 8, 0> & 3>) == 2)

from it.  Without -DALT

;; Function test_pic (null)
;; enabled by -tree-original


{
  if ((SAVE_EXPR <BIT_FIELD_REF <*info, 8, 0> & 3>) == 3 || (SAVE_EXPR
<BIT_FIELD_REF <*info, 8, 0> & 3>) == 1)

;; Function test_exe (null)
;; enabled by -tree-original


{
  if ((BIT_FIELD_REF <*info, 8, 0> & 3) <= 1)


I see more that a single bit test for both cases btw, mostly because we
need to mask the padding.  Not sure what optimal code you expect here.

[Bug tree-optimization/67328] range test rather than single bit test for code testing enum values

[amodra at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67328

Alan Modra <amodra at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |NEW

--- Comment #3 from Alan Modra <amodra at gmail dot com> ---
Maybe adding #if ALT only confused the issue.  For -UALT -O1 on x86_64 I get

test_pic:
        testb   $1, (%rdi)
        je      .L1
        addl    $1, result(%rip)
.L1:
        rep ret

test_exe:
        movzbl  (%rdi), %eax
        andl    $3, %eax
        cmpb    $1, %al
        ja      .L3
        addl    $1, result(%rip)
.L3:
        rep ret

For test_exe I would like to see the following equivalent and better optimised
code.

test_exe:
        testb   $2, (%rdi)
        jne     .L3
        addl    $1, result(%rip)
.L3:
        rep ret

The optimisation to a bit test for test_pic happens in reassoc1, or at least
that's where the process starts.