[Bug target/100181] New: hot-cold partitioned code doesn't assemble

[Bug target/100181] New: hot-cold partitioned code doesn't assemble

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

            Bug ID: 100181
           Summary: hot-cold partitioned code doesn't assemble
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---

I see lots of libgomp fails with offloading for amdgcn like

./pr93515.xamdgcn-amdhsa.mkoffload.2.s:634:2: error: undefined label '.L33'
        s_branch        .L33
        ^
mkoffload: fatal error: /usr/bin//x86_64-suse-linux-accel-amdgcn-amdhsa-gcc-11
returned 1 exit status

which is because we compile to

        s_branch        .L33
        .section        .text.unlikely
        .type   main._omp_fn.4.cold, @function
main._omp_fn.4.cold:
.L33:
        s_getpc_b64     s[2:3]
        s_add_u32       s2, s2, abort@rel32@lo+4
        s_addc_u32      s3, s3, abort@rel32@hi+4
        s_swappc_b64    s[18:19], s[2:3]
        .text

and thus have a section crossing jump using a local label.  The amdgcn
assembler
I'm using doesn't like this, it is

> amdgcn-amdhsa-as --version
LLVM (http://llvm.org/):
  LLVM version 11.0.1
  Optimized build.
  Default target: x86_64-unknown-linux-gnu
  Host CPU: znver2

[Bug target/100181] hot-cold partitioned code doesn't assemble

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |assemble-failure
             Target|                            |gcn

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Example testcases that fail this way include

libgomp.c/../libgomp.c-c++-common/pr93515.c
libgomp.c/pr86660.c
libgomp.c/target-3.c
libgomp.c/target-4.c
libgomp.c/target-5.c
...

and more.  Altering -freorder-blocks-and-partition for amdgcn might be a
possible fix but might also be "interesting" to achieve (it's stuck in the
optimization node of each function).

[Bug target/100181] hot-cold partitioned code doesn't assemble

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 50648
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50648&action=edit
example assembly file

Example file that fails to assemble.

[Bug target/100181] hot-cold partitioned code doesn't assemble

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
llvm-mc from llvm 9.0.1 ICEs:

> llvm-mc-9.0.1 -o t.o pr93515.xamdgcn-amdhsa.mkoffload.2.s -triple=amdgcn--amdhsa -mcpu=fiji -filetype=obj
Stack dump:
0.      Program arguments: llvm-mc-9.0.1 -o t.o
pr93515.xamdgcn-amdhsa.mkoffload.2.s -triple=amdgcn--amdhsa -mcpu=fiji
-filetype=obj 
#0 0x00007ffff3d9e17d llvm::sys::PrintStackTrace(llvm::raw_ostream&)
(/usr/bin/../lib64/libLLVM.so.9+0xa9b17d)
#1 0x00007ffff3d9bc30 llvm::sys::RunSignalHandlers()
(/usr/bin/../lib64/libLLVM.so.9+0xa98c30)
#2 0x00007ffff3d9e5a4 (/usr/bin/../lib64/libLLVM.so.9+0xa9b5a4)
#3 0x00007ffff2baa4a0 __restore_rt (/lib64/libc.so.6+0x394a0)
#4 0x00007ffff57fe353
llvm::AMDGPUInstPrinter::printOperandAndFPInputMods(llvm::MCInst const*,
unsigned int, llvm::MCSubtargetInfo const&, llvm::raw_ostream&)
(/usr/bin/../lib64/libLLVM.so.9+0x24fb353)
#5 0x00007ffff57f1a6f llvm::AMDGPUInstPrinter::printInstruction(llvm::MCInst
const*, llvm::MCSubtargetInfo const&, llvm::raw_ostream&)
(/usr/bin/../lib64/libLLVM.so.9+0x24eea6f)
#6 0x0000000000480c60 
Segmentation fault (core dumped)

[Bug target/100181] hot-cold partitioned code doesn't assemble

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Depends on|                            |94278

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ah, looks like a duplicate of PR94278 then.


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94278
[Bug 94278] [amdgcn] Offloading build failures due to 'llvm-mc' SIGSEGV

[Bug target/100181] hot-cold partitioned code doesn't assemble

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|assemble-failure            |
         Depends on|94278                       |
                 CC|                            |ams at gcc dot gnu.org,
                   |                            |burnus at gcc dot gnu.org,
                   |                            |jules at gcc dot gnu.org,
                   |                            |tschwinge at gcc dot gnu.org

--- Comment #5 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
Isn't that a duplicate of what I once filed as PR94278 "[amdgcn] Offloading
build failures due to 'llvm-mc' SIGSEGV" -- with "improved error reporting":
error diagnostic instead of SIGSEGV?

PR94278 includes a patch that I've locally been using ever since, but Andrew
has not been able to reproduce that problem.


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94278
[Bug 94278] [amdgcn] Offloading build failures due to 'llvm-mc' SIGSEGV

[Bug target/100181] hot-cold partitioned code doesn't assemble

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |DUPLICATE
             Status|UNCONFIRMED                 |RESOLVED

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
Yes.  I'm giving the patch a try.

*** This bug has been marked as a duplicate of bug 94278 ***

[Bug target/100181] hot-cold partitioned code doesn't assemble

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

--- Comment #7 from Tobias Burnus <burnus at gcc dot gnu.org> ---
I can reproduce this one with LLVM trunk when compiled with '-save-temps -O2'
and then

llvm-mc --triple=amdgcn--amdhsa -mcpu=fiji -filetype=obj
--amdhsa-code-object-version=3 a.xamdgcn-amdhsa.mkoffload.1.s -o foo.o

a.xamdgcn-amdhsa.mkoffload.1.s:634:2: error: undefined label '.L33'

Removing the '.unlikely' from the .section name fixes this (or using -O1).
(I could not reproduce the LLVM 9 issue in PR94278 back then.)

[Bug target/100181] hot-cold partitioned code doesn't assemble

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

--- Comment #8 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
(In reply to Tobias Burnus from comment #7)
> (I could not reproduce the LLVM 9 issue in PR94278 back then.)

Hmm, but didn't you say in the LLVM issue
<https://bugs.llvm.org/show_bug.cgi?id=45887> (liked to in _See Also_ in
PR94278) that you _did_ reproduce this back then?  (Just making sure that it's
not actually a different issue that we're discussing now?)

[Bug target/100181] hot-cold partitioned code doesn't assemble

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

--- Comment #9 from Tobias Burnus <burnus at gcc dot gnu.org> ---
(In reply to Thomas Schwinge from comment #8)
> (In reply to Tobias Burnus from comment #7)
> > (I could not reproduce the LLVM 9 issue in PR94278 back then.)
> 
> Hmm, but didn't you say in the LLVM issue
> <https://bugs.llvm.org/show_bug.cgi?id=45887> that you _did_ reproduce this back then?

I partially mixed it up - I could not reproduce it with a self-build LLVM but
could do so with the distro LLVM ones.

Our/my LLVM bug 45887 is similar to
https://bugs.llvm.org/show_bug.cgi?id=41914, which was fixed by
https://reviews.llvm.org/D79943 and the latter adds the check which causes the
error message: "undefined label '.L33'"

However, the other bug (41914) and the the testcase in the fix (D79943) were
about a label which was truly missing as opposed to one which is just in a
differently named .section.

 * * *

Obviously, with D79943 at place (LLVM 11, 12, ...), I can reproduce the issue
with -O2 (with < -O0 there is only .text) as then 's_branch .L33' is in
.section .text but '.L33:' is in '.section .text.unlikely'.

Whether it is intended that cross-.section s_branch are no longer permitted or
whether it happened by chance by this patch or on purpose, I don't know.

Sections themselves are supported – and the  LVM testsuite for GCN has
testcases with '.section' (under llvm/test/MC/AMDGPU/*.s) – but in the
testsuite one at a time.

GCC guards the .text.unlikely generation by (→ varasm.c's
default_function_section) by

  if (!flag_reorder_functions
      || !targetm_common.have_named_sections)
    return NULL;

while the patch in PR 94278 comment 3 uses
   flag_reorder_blocks_and_partition = 0

[Bug target/100181] hot-cold partitioned code doesn't assemble

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

Tobias Burnus <burnus at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2021-07-19
     Ever confirmed|0                           |1
             Status|RESOLVED                    |NEW
         Resolution|DUPLICATE                   |---

--- Comment #10 from Tobias Burnus <burnus at gcc dot gnu.org> ---
Remove duplicate as this PR describes the issue more clearly than PR94278 (I
will late mark the latte as duplicate of this PR.)

LLVM main (which will become LLVM 13) has FIXES for this issue:

- [AMDGPU] Handle s_branch to another section.
  https://reviews.llvm.org/D105760 / commit b205f2bb8938

- [lld][AMDGPU] Handle R_AMDGPU_REL16 relocation.
  https://reviews.llvm.org/D105761 / commit fb9c5c3dce27

[Bug target/100181] hot-cold partitioned code doesn't assemble

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

--- Comment #11 from Tobias Burnus <burnus at gcc dot gnu.org> ---
*** Bug 94278 has been marked as a duplicate of this bug. ***

[Bug target/100181] hot-cold partitioned code doesn't assemble

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

Tobias Burnus <burnus at gcc dot gnu.org> changed:

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

--- Comment #12 from Tobias Burnus <burnus at gcc dot gnu.org> ---
I think with all the fixes which went into LLVM main and are now also in
release LLVM 13.0 (including a late fix, which is only in LLVM 13.0.1) this
issue is FIXED.

Thus, close as FIXED – if new issues shows up again, I think it makes sense to
open a new PR.

[Bug target/100181] hot-cold partitioned code doesn't assemble

[ams at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

--- Comment #13 from Andrew Stubbs <ams at gcc dot gnu.org> ---
I've updated the LLVM version documentation at
https://gcc.gnu.org/wiki/Offloading#For_AMD_GCN:

It's LLVM 9 or 13.0.1 now (nothing in between), and will be 13.0.1+ for the
next release (dropping LLVM 9 because we'll want to add newer device support to
GCC soonish).

[Bug target/100181] hot-cold partitioned code doesn't assemble

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100181

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           See Also|https://gcc.gnu.org/bugzill |
                   |a/show_bug.cgi?id=97827,    |
                   |https://bugs.llvm.org/show_ |
                   |bug.cgi?id=41914,           |
                   |https://gcc.gnu.org/bugzill |
                   |a/show_bug.cgi?id=93515,    |
                   |https://gcc.gnu.org/bugzill |
                   |a/show_bug.cgi?id=86660     |

--- Comment #14 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
Confirming that this issue has been "resolved" ;-) via the recent commit
r13-742-g8086230e7ac619c0b0eeb6e15df7975ac214725f "amdgcn: Remove LLVM 9
assembler/linker support": "The minimum required LLVM version is now 13.0.1,
and is enforced by configure."; I've thus removed my "[gcn] [WIP] Force-disable
'-freorder-blocks-and-partition' [PR94278]" hack from local trees.