[Bug c++/114921] New: Optimization flags cause _Float16 to __bf16 casting to do nothing

[Bug c++/114921] New: Optimization flags cause _Float16 to __bf16 casting to do nothing

[mat.mcroci at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114921

            Bug ID: 114921
           Summary: Optimization flags cause _Float16 to __bf16 casting to
                    do nothing
           Product: gcc
           Version: 13.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: mat.mcroci at gmail dot com
  Target Milestone: ---

When compiling the following code (assuming it's in a file called test.cpp)

#include <stdlib.h>
#include <cstdint>
#include <array>
#include <iostream>

#define SIZE 8

typedef _Float16 T;
//typedef volatile float T;

void fp16tobf16(_Float16 * f) {
        __bf16 * b = reinterpret_cast<__bf16*>(f);
        for(int i=0; i<SIZE; i++){
            T temp = f[i];
            b[i] = (__bf16) temp;
        }
}

with:

g++ -c -S -O2 -std=c++23 -o test.s test.cpp

using gcc 13.2.0 compiled with binutils 2.42 using spack, the assembler does
not do any casting and simply returns. You can also check this behaviour on
godbolt. The resulting assembly code is simply

fp16tobf16(_Float16*):
        ret

When using -O1 or -O0, the assembler generates the command __extendhfbf2 which
causes a separate issue, which is related to this bug:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114907

P.s. I reported this bug in binutils already, but they suggested I reported it
here instead, see

https://sourceware.org/bugzilla/show_bug.cgi?id=31685

[Bug c++/114921] Optimization flags cause _Float16 to __bf16 casting to do nothing

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114921

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
           Keywords|                            |wrong-code
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2024-05-02
             Status|UNCONFIRMED                 |ASSIGNED

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  We vectorize the loop to

  <bb 2> [local count: 119292720]:
  vect_temp_9.6_3 = MEM <vector(4) _Float16> [(_Float16 *)f_7(D)];
  vect__4.7_9 = VIEW_CONVERT_EXPR<vector(4) __bf16>(vect_temp_9.6_3);
  MEM <vector(4) __bf16> [(__bf16 *)f_7(D)] = vect__4.7_9;
  vect_temp_9.6_17 = MEM <vector(4) _Float16> [(_Float16 *)f_7(D) + 8B];
  vect__4.7_18 = VIEW_CONVERT_EXPR<vector(4) __bf16>(vect_temp_9.6_17);
  MEM <vector(4) __bf16> [(__bf16 *)f_7(D) + 8B] = vect__4.7_18;

likely because the vectorizer thinks this is a noop conversion, it handles
it via vectorizable_assignment.

[Bug c++/114921] Optimization flags cause _Float16 to __bf16 casting to do nothing

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114921

--- Comment #2 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:87e35da16df74cd1c4729a55d94e7bc592487f48

commit r15-124-g87e35da16df74cd1c4729a55d94e7bc592487f48
Author: Richard Biener <rguenther@suse.de>
Date:   Thu May 2 13:55:15 2024 +0200

    tree-optimization/114921 - _Float16 -> __bf16 isn't noop

    The vectorizer handles a _Float16 to __bf16 conversion through
    vectorizable_assignment, thinking it's a noop.  The following
    fixes this by requiring the same vector component mode when
    checking for CONVERT_EXPR_CODE_P, being stricter than for
    VIEW_CONVERT_EXPR.

            PR tree-optimization/114921
            * tree-vect-stmts.cc (vectorizable_assignment): Require
            same vector component modes for input and output for
            CONVERT_EXPR_CODE_P.

[Bug tree-optimization/114921] Optimization flags cause _Float16 to __bf16 casting to do nothing

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114921

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

https://gcc.gnu.org/g:fa7e05d90bb672b58424ddbe37a05a0fd0949b35

commit r14-10159-gfa7e05d90bb672b58424ddbe37a05a0fd0949b35
Author: Richard Biener <rguenther@suse.de>
Date:   Thu May 2 13:55:15 2024 +0200

    tree-optimization/114921 - _Float16 -> __bf16 isn't noop

    The vectorizer handles a _Float16 to __bf16 conversion through
    vectorizable_assignment, thinking it's a noop.  The following
    fixes this by requiring the same vector component mode when
    checking for CONVERT_EXPR_CODE_P, being stricter than for
    VIEW_CONVERT_EXPR.

            PR tree-optimization/114921
            * tree-vect-stmts.cc (vectorizable_assignment): Require
            same vector component modes for input and output for
            CONVERT_EXPR_CODE_P.

    (cherry picked from commit 87e35da16df74cd1c4729a55d94e7bc592487f48)

[Bug tree-optimization/114921] Optimization flags cause _Float16 to __bf16 casting to do nothing

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114921

--- Comment #4 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:d0d6dcc019cd32eebf85d625f56e0f7573938319

commit r15-223-gd0d6dcc019cd32eebf85d625f56e0f7573938319
Author: Richard Biener <rguenther@suse.de>
Date:   Mon May 6 12:03:09 2024 +0200

    tree-optimization/114921 - _Float16 -> __bf16 isn't noop fixup

    The following further strengthens the check which convert expressions
    we allow to vectorize as simple copy by resorting to
    tree_nop_conversion_p on the vector components.

            PR tree-optimization/114921
            * tree-vect-stmts.cc (vectorizable_assignment): Use
            tree_nop_conversion_p to identify converts we can vectorize
            with a simple assignment.

[Bug tree-optimization/114921] Optimization flags cause _Float16 to __bf16 casting to do nothing

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114921

--- Comment #5 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-14 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:9e971c671ded9647beb0a1c5b9430b4e64060862

commit r14-10255-g9e971c671ded9647beb0a1c5b9430b4e64060862
Author: Richard Biener <rguenther@suse.de>
Date:   Mon May 6 12:03:09 2024 +0200

    tree-optimization/114921 - _Float16 -> __bf16 isn't noop fixup

    The following further strengthens the check which convert expressions
    we allow to vectorize as simple copy by resorting to
    tree_nop_conversion_p on the vector components.

            PR tree-optimization/114921
            * tree-vect-stmts.cc (vectorizable_assignment): Use
            tree_nop_conversion_p to identify converts we can vectorize
            with a simple assignment.

    (cherry picked from commit d0d6dcc019cd32eebf85d625f56e0f7573938319)