[Bug middle-end/115307] New: [avr] Don't expand isinf() like a built-in

[Bug middle-end/115307] New: [avr] Don't expand isinf() like a built-in

[gjl at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115307

            Bug ID: 115307
           Summary: [avr] Don't expand isinf() like a built-in
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: gjl at gcc dot gnu.org
  Target Milestone: ---

avr-gcc compiles

int isinff (float);

int call_isinff (float f)
{
    return isinff (f);
}

like:
call_isinff:
        push r12                 ;  63  [c=4 l=1]  pushqi1/0
        push r13                 ;  64  [c=4 l=1]  pushqi1/0
        push r14                 ;  65  [c=4 l=1]  pushqi1/0
        push r15                 ;  66  [c=4 l=1]  pushqi1/0
        push r28                 ;  67  [c=4 l=1]  pushqi1/0
/* prologue: function */
/* frame size = 0 */
/* stack size = 5 */
.L__stack_usage = 5
        movw r12,r22     ;  56  [c=4 l=2]  *movsf/0
        movw r14,r24
        clt      ;  57  [c=4 l=2]  *abssf2/1
        bld r15,7
        ldi r28,lo8(1)   ;  58  [c=4 l=1]  movqi_insn/1
        ldi r18,lo8(-1)  ;  59  [c=4 l=4]  *movsf/4
        ldi r19,lo8(-1)
        ldi r20,lo8(127)
        ldi r21,lo8(127)
        movw r24,r14     ;  60  [c=4 l=2]  *movsf/0
        movw r22,r12
        call __unordsf2  ;  10  [c=12 l=2]  call_value_insn/1
        cpse r24,__zero_reg__    ;  61  [c=0 l=1]  enable_interrupt-3
        rjmp .L2
        ldi r18,lo8(-1)  ;  52  [c=4 l=4]  *movsf/4
        ldi r19,lo8(-1)
        ldi r20,lo8(127)
        ldi r21,lo8(127)
        movw r24,r14     ;  53  [c=4 l=2]  *movsf/0
        movw r22,r12
        call __lesf2     ;  15  [c=12 l=2]  call_value_insn/1
        cp __zero_reg__,r24      ;  77  [c=16 l=1]  *swapped_tstqi
        brge .L2                 ;  78  [c=4 l=1]  branch
        ldi r28,0                ;  51  [c=4 l=1]  movqi_insn/0
.L2:
        ldi r24,lo8(1)   ;  45  [c=4 l=1]  movqi_insn/1
        eor r24,r28      ;  46  [c=4 l=1]  *xorqi3
        ldi r25,0                ;  50  [c=4 l=1]  movqi_insn/0
/* epilogue start */
        pop r28          ;  70  [c=4 l=1]  popqi
        pop r15          ;  71  [c=4 l=1]  popqi
        pop r14          ;  72  [c=4 l=1]  popqi
        pop r13          ;  73  [c=4 l=1]  popqi
        pop r12          ;  74  [c=4 l=1]  popqi
        ret              ;  75  [c=0 l=1]  return_from_epilogue

where a simple call to isinff would be mush better.

[Bug middle-end/115307] [avr] Don't expand isinf() like a built-in

[gjl at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115307

Georg-Johann Lay <gjl at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P4

[Bug target/115307] [avr] Don't expand isinf() like a built-in

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115307

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue is that we probably fold isinff early.  On x86 I see already in
.original:

  return !(ABS_EXPR <f> u<= 3.4028234663852885981170418348451692544e+38);

I think your option is to provide optabs for isinf<mode> but make expansion
of them always FAIL; (which is of course a quite ugly way)

[Bug tree-optimization/115307] [avr] Don't expand isinf() like a built-in

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115307

--- Comment #2 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Georg-Johann Lay <gjl@gcc.gnu.org>:

https://gcc.gnu.org/g:fabb545026f714b7d1cbe586f4c5bbf6430bdde3

commit r15-966-gfabb545026f714b7d1cbe586f4c5bbf6430bdde3
Author: Georg-Johann Lay <avr@gjlay.de>
Date:   Sat Jun 1 10:38:00 2024 +0200

    AVR: tree-optimization/115307 - Work around isinf bloat from early passes.

            PR tree-optimization/115307
    gcc/
            * config/avr/avr.md (SFDF): New mode iterator.
            (isinf<mode>2) [sf, df]: New expanders.

    gcc/testsuite/
            * gcc.target/avr/torture/pr115307-isinf.c: New test.

[Bug tree-optimization/115307] [avr] Don't expand isinf() like a built-in

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115307

--- Comment #3 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-14 branch has been updated by Georg-Johann Lay
<gjl@gcc.gnu.org>:

https://gcc.gnu.org/g:9d08c55f7c99329f4289ad3a4157c2d8d8a78d8c

commit r14-10266-g9d08c55f7c99329f4289ad3a4157c2d8d8a78d8c
Author: Georg-Johann Lay <avr@gjlay.de>
Date:   Sat Jun 1 10:38:00 2024 +0200

    AVR: tree-optimization/115307 - Work around isinf bloat from early passes.

            PR tree-optimization/115307
    gcc/
            * config/avr/avr.md (SFDF): New mode iterator.
            (isinf<mode>2) [sf, df]: New expanders.

    gcc/testsuite/
            * gcc.target/avr/torture/pr115307-isinf.c: New test.

    (cherry picked from commit fabb545026f714b7d1cbe586f4c5bbf6430bdde3)

[Bug tree-optimization/115307] [avr] Don't expand isinf() like a built-in

[gjl at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115307

--- Comment #4 from Georg-Johann Lay <gjl at gcc dot gnu.org> ---
The isinf part is fixied in v14.2+, but there is also isnan etc. which don't
currently habe an optabs entry, so defining a failing expander won't work.

[Bug tree-optimization/115307] [avr] Don't expand isinf() like a built-in

[gjl at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115307

--- Comment #5 from Georg-Johann Lay <gjl at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #1)
> The issue is that we probably fold isinff early.  On x86 I see already in
> .original:
> 
>   return !(ABS_EXPR <f> u<= 3.4028234663852885981170418348451692544e+38);

That looks incorrect.  The man 3 page for isinf states that it returns -1 for
-Inf, but that expression returns 1 for -Inf (same for the expanded avr code
above).

[Bug tree-optimization/115307] [avr] Don't expand isinf() like a built-in

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115307

--- Comment #6 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-13 branch has been updated by Georg-Johann Lay
<gjl@gcc.gnu.org>:

https://gcc.gnu.org/g:c57d73f4cd5ca61327406fc2521a2235dd49d12e

commit r13-8817-gc57d73f4cd5ca61327406fc2521a2235dd49d12e
Author: Georg-Johann Lay <avr@gjlay.de>
Date:   Sat Jun 1 10:38:00 2024 +0200

    AVR: tree-optimization/115307 - Work around isinf bloat from early passes.

            PR tree-optimization/115307
    gcc/
            * config/avr/avr.md (SFDF): New mode iterator.
            (isinf<mode>2) [sf, df]: New expanders.

    gcc/testsuite/
            * gcc.target/avr/torture/pr115307-isinf.c: New test.

    (cherry picked from commit fabb545026f714b7d1cbe586f4c5bbf6430bdde3)

[Bug tree-optimization/115307] [avr] Don't expand isinf() like a built-in

[gjl at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115307

Georg-Johann Lay <gjl at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P4                          |P3

[Bug tree-optimization/115307] [avr] Don't expand isinf() like a built-in

[jsm28 at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115307

--- Comment #7 from Joseph S. Myers <jsm28 at gcc dot gnu.org> ---
Note the difference between __builtin_isinf (C standard semantics, returns
unspecified nonzero value for an infinity) and __builtin_isinf_sign (stricter
semantics returning 1 for +Inf and -1 for -Inf).