[Bug target/112669] New: GCN: wrong 'LIBRARY_PATH' in presence of several different '-march=[...]' flags

[Bug target/112669] New: GCN: wrong 'LIBRARY_PATH' in presence of several different '-march=[...]' flags

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112669

            Bug ID: 112669
           Summary: GCN: wrong 'LIBRARY_PATH' in presence of several
                    different '-march=[...]' flags
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tschwinge at gcc dot gnu.org
                CC: ams at gcc dot gnu.org, jules at gcc dot gnu.org
  Target Milestone: ---
            Target: GCN

I've run into a weird issue when several different '-march=[...]' flags appear.
 This causes linking to fail: the linker tries to link in the wrong multilib's
libraries.  This happens, for example, if the user provides '-march=[...]' for
libgomp offloading testing, but a test cases also specifies a specific
'-march=[...]'.

The problem might perhaps be in GCN multilib setup, however it doesn't seem
related to the recent changes ("amdgcn: deprecate Fiji device and multilib"),
as I'm also reproducing the issue with previous GCC release branches.

The issue -- I suppose -- boils down to:

No '-march=[...]' flag appears, default paths:

    $ build-gcc-offload-amdgcn-amdhsa/gcc/xgcc -print-search-dirs | sed -n -e
"/^libraries: =/{s%[^=]\+=%%;s%$PWD%[...]%g;s%:%\n%g;p}" > default
    $ cat < default 
   
[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/
    [...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/
   
[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/../../../../../../amdgcn-amdhsa/lib/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/
   
[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/../../../../../../amdgcn-amdhsa/lib/
   
[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../amdgcn-amdhsa/lib/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/
    [...]/build-gcc-offload-amdgcn-amdhsa/gcc/../amdgcn-amdhsa/lib/

If one '-march=[...]' flag appears, we get those multilib paths prepended:

    $ build-gcc-offload-amdgcn-amdhsa/gcc/xgcc -print-search-dirs -march=gfx906
| sed -n -e "/^libraries: =/{s%[^=]\+=%%;s%$PWD%[...]%g;s%:%\n%g;p}" > gfx906
    $ diff -U1 default gfx906
    --- default     2023-11-22 11:47:14.021018613 +0100
    +++ gfx906      2023-11-22 11:47:21.856931965 +0100
    @@ -1 +1,7 @@
   
+[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/gfx906/
    +[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/gfx906/
   
+[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/../../../../../../amdgcn-amdhsa/lib/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/gfx906/
   
+[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/../../../../../../amdgcn-amdhsa/lib/gfx906/
   
+[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../amdgcn-amdhsa/lib/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/gfx906/
    +[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../amdgcn-amdhsa/lib/gfx906/
    
[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/

Similarly, if the same '-march=[...]' flag appears twice, we get those multilib
paths prepended:

    $ build-gcc-offload-amdgcn-amdhsa/gcc/xgcc -print-search-dirs -march=gfx908
-march=gfx908 | sed -n -e "/^libraries:
=/{s%[^=]\+=%%;s%$PWD%[...]%g;s%:%\n%g;p}" > gfx908
    $ diff -U1 default gfx908
    --- default     2023-11-22 11:47:14.021018613 +0100
    +++ gfx908      2023-11-22 11:47:34.760789347 +0100
    @@ -1 +1,7 @@
   
+[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/gfx908/
    +[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/gfx908/
   
+[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/../../../../../../amdgcn-amdhsa/lib/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/gfx908/
   
+[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/../../../../../../amdgcn-amdhsa/lib/gfx908/
   
+[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../amdgcn-amdhsa/lib/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/gfx908/
    +[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../amdgcn-amdhsa/lib/gfx908/
    
[...]/build-gcc-offload-amdgcn-amdhsa/gcc/../lib/gcc/x86_64-pc-linux-gnu/14.0.0/accel/amdgcn-amdhsa/

However, if several different '-march=[...]' flags appear, we're back to the
default:

    $ build-gcc-offload-amdgcn-amdhsa/gcc/xgcc -print-search-dirs -march=gfx906
-march=gfx908 | sed -n -e "/^libraries:
=/{s%[^=]\+=%%;s%$PWD%[...]%g;s%:%\n%g;p}" > gfx906,gfx908
    $ cmp default gfx906,gfx908 && echo 'no difference'
    no difference
    $ build-gcc-offload-amdgcn-amdhsa/gcc/xgcc -print-search-dirs -march=gfx908
-march=gfx906 | sed -n -e "/^libraries:
=/{s%[^=]\+=%%;s%$PWD%[...]%g;s%:%\n%g;p}" > gfx908,gfx906
    $ cmp default gfx908,gfx906 && echo 'no difference'
    no difference

Instead, the last '-march=[...]' should win.

[Bug target/112669] GCN: wrong 'LIBRARY_PATH' in presence of several different '-march=[...]' flags

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112669

--- Comment #1 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
Tracing through 'gcc/gcc.cc': 'build_search_list' -> 'for_each_path', I find:

For '-march=gfx908', we have:

    (gdb) print multilib_dir
    $3 = 0x82e6c0 "gfx908"
    (gdb) print multilib_os_dir
    $4 = 0x82e6c0 "gfx908"

For '-march=gfx906 -march=gfx908', we have:

    (gdb) print multilib_dir
    $3 = 0x0
    (gdb) print multilib_os_dir
    $4 = 0x0

These are:

    /* Subdirectory to use for locating libraries.  Set by
       set_multilib_dir based on the compilation options.  */

    static const char *multilib_dir;

    /* Subdirectory to use for locating libraries in OS conventions.  Set by
       set_multilib_dir based on the compilation options.  */

    static const char *multilib_os_dir;

Indeed, simpler:

    $ build-gcc-offload-amdgcn-amdhsa/gcc/xgcc -print-multi-directory
-march=gfx908
    gfx908
    $ build-gcc-offload-amdgcn-amdhsa/gcc/xgcc -print-multi-directory
-march=gfx906 -march=gfx908
    .

Instead of '.' (default), the latter should also print 'gfx908'.

I'll look into 'set_multilib_dir' etc.

[Bug target/112669] GCN: wrong 'LIBRARY_PATH' in presence of several different '-march=[...]' flags

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112669

Thomas Schwinge <tschwinge at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2023-11-22
             Status|UNCONFIRMED                 |ASSIGNED
     Ever confirmed|0                           |1
           Assignee|unassigned at gcc dot gnu.org      |tschwinge at gcc dot gnu.org

[Bug target/112669] GCN: wrong 'LIBRARY_PATH' in presence of several different '-march=[...]' flags

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112669

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

https://gcc.gnu.org/g:297fe5c166f1d920da6d5ef9c7e02846b0148575

commit r14-5884-g297fe5c166f1d920da6d5ef9c7e02846b0148575
Author: Thomas Schwinge <thomas@codesourcery.com>
Date:   Wed Nov 22 17:35:23 2023 +0100

    GCN: Tag '-march=[...]', '-mtune=[...]' as 'Negative' of themselves
[PR112669]

    Certain other command-line flags are mutually exclusive (random example:
GCN
    '-march=gfx906', '-march=gfx908').  If they're not appropriately marked up,
    this does disturb the multilib selection machinery, for example:

        $ build-gcc-offload-amdgcn-amdhsa/gcc/xgcc -print-multi-directory
-march=gfx906
        gfx906
        $ build-gcc-offload-amdgcn-amdhsa/gcc/xgcc -print-multi-directory
-march=gfx908
        gfx908
        $ build-gcc-offload-amdgcn-amdhsa/gcc/xgcc -print-multi-directory
-march=gfx906 -march=gfx908
        .

    In the last invocation, '-march=gfx900 -march=gfx906', for example, in
    'gcc/gcc.cc:set_multilib_dir' we see both flags -- which there doesn't
exist a
    matching multilib for, therefore we "fail" to the default ('.').  Tagges as
    'Negative', only the last flag survives, and we, for example, get the
expected:

        $ build-gcc-offload-amdgcn-amdhsa/gcc/xgcc -print-multi-directory
-march=gfx906 -march=gfx908
        gfx908

    I quickly found that the same also applies to GCN's '-mtune=[...]', but
I've
    not otherwise reviewed the GCN options.

            PR target/112669
            gcc/
            * config/gcn/gcn.opt (march=, mtune=): Tag as 'Negative' of
            themselves.

[Bug target/112669] GCN: wrong 'LIBRARY_PATH' in presence of several different '-march=[...]' flags

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112669

Thomas Schwinge <tschwinge at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #3 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
Fixed in master branch; not currently planning on backporting.