fatal error: error in backend: Failed to evaluate function length in SEH unwind info

[kalibera]

I've encountered a crash of clang from LLVM 17.01 when trying to build openblas 0.3.24 for Windows/aarch64 on a Linux machine. The crash files and full output with stacktrace are attached in crash.zip

The output includes:

fatal error: error in backend: Failed to evaluate function length in SEH unwind info

#5 0x00005579fe4a4f70 llvm::report_fatal_error(llvm::Twine const&, bool) (/home/tomas/toolchain/clang/toolchain_libs/mxe/usr/x86_64-pc-linux-gnu/bin/clang+0x1c72f70)
#8 0x00005579fe2627dd llvm::MCStreamer::emitWinCFIEndProc(llvm::SMLoc) (/home/tomas/toolchain/clang/toolchain_libs/mxe/usr/x86_64-pc-linux-gnu/bin/clang+0x1a307dd)
#11 0x00005579ff1be469 llvm::AsmPrinter::emitFunctionBody() (/home/tomas/toolchain/clang/toolchain_libs/mxe/usr/x86_64-pc-linux-gnu/bin/clang+0x298c469)
#14 0x00005579fdff2eea llvm::FPPassManager::runOnFunction(llvm::Function&) (/home/tomas/toolchain/clang/toolchain_libs/mxe/usr/x86_64-pc-linux-gnu/bin/clang+0x17c0eea)

[kalibera]

cc'ing @mstorsjo

[mstorsjo]

I can reproduce the error with clang 15 but not with 14, so either it's a regression, or the issue just happened to go unnoticed before.

[mstorsjo]

CC @efriedma-quic

I reduced the testcase to the following:

void a();
void b() {
  for (;;)
    a();
}

Compiled with the following command: clang -target aarch64-windows -O2 -mtune=cortex-a53 -c repro.c

The issue is that there's an alignment inserted in the middle of a function, which makes it impossible for the SEH unwind info generation to measure the size of the function (as it hasn't been laid out yet, so we don't know how large the alignment gap ends up). In this case, the output is:

$ clang -target aarch64-windows -O2 -mtune=cortex-a53 -S -o - repro.c
        .globl  b                               // -- Begin function b
        .p2align        4
b:                                      // @b
.seh_proc b
// %bb.0:                               // %entry
        str     x30, [sp, #-16]!                // 8-byte Folded Spill
        .seh_save_reg_x x30, 16
        .seh_endprologue
        .p2align        4, , 8
.LBB0_1:                                // %for.cond
                                        // =>This Inner Loop Header: Depth=1
        bl      a
        b       .LBB0_1
        .seh_endfunclet
        .seh_endproc

Here, the .p2align is what's causing the breakage.

I bisected this down to the commit 7d67671:

commit 7d676714fbf2dd4bcdd15bd63d6b43b467ceccb1
Author: Nicholas Guy <nicholas.guy@arm.com>
Date:   Tue Mar 22 16:35:12 2022 +0000

    [AArch64] Set MaxBytesForLoopAlignment for more targets
    
    Differential Revision: https://reviews.llvm.org/D122566

 llvm/lib/Target/AArch64/AArch64Subtarget.cpp        | 18 +++++++++++++++++-
 .../AArch64/aarch64-p2align-max-bytes-neoverse.ll   | 21 +++++++++++++++++----
 llvm/test/CodeGen/AArch64/merge-store-dependency.ll |  2 ++
 .../CodeGen/AArch64/preferred-function-alignment.ll |  2 +-
 4 files changed, 37 insertions(+), 6 deletions(-)

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

I've encountered a crash of clang from LLVM 17.01 when trying to build openblas 0.3.24 for Windows/aarch64 on a Linux machine. The crash files and full output with stacktrace are attached in [crash.zip](https://github.com/llvm/llvm-project/files/12674259/crash.zip)

The output includes:

fatal error: error in backend: Failed to evaluate function length in SEH unwind info

#5 0x00005579fe4a4f70 llvm::report_fatal_error(llvm::Twine const&, bool) (/home/tomas/toolchain/clang/toolchain_libs/mxe/usr/x86_64-pc-linux-gnu/bin/clang+0x1c72f70)
#8 0x00005579fe2627dd llvm::MCStreamer::emitWinCFIEndProc(llvm::SMLoc) (/home/tomas/toolchain/clang/toolchain_libs/mxe/usr/x86_64-pc-linux-gnu/bin/clang+0x1a307dd)
#11 0x00005579ff1be469 llvm::AsmPrinter::emitFunctionBody() (/home/tomas/toolchain/clang/toolchain_libs/mxe/usr/x86_64-pc-linux-gnu/bin/clang+0x298c469)
#14 0x00005579fdff2eea llvm::FPPassManager::runOnFunction(llvm::Function&) (/home/tomas/toolchain/clang/toolchain_libs/mxe/usr/x86_64-pc-linux-gnu/bin/clang+0x17c0eea)

[DavidSpickett]

@NickGuy-Arm

Edit: Though now I realise this landed about a year ago. Maybe they'll have some thoughts anyway.

[efriedma-quic]

This has also been reported as #47432 and #64504. The underlying backend limitation is known, but hard to fix... it requires a complicated rearrangement of the code to delay computing the length of a function until assembler relaxation. (I think there's some discussion of this elsewhere, but I can't seem to find it.)

[efriedma-quic]

See discussion on https://reviews.llvm.org/D61095

[mstorsjo]

This has also been reported as #47432 and #64504. The underlying backend limitation is known, but hard to fix... it requires a complicated rearrangement of the code to delay computing the length of a function until assembler relaxation. (I think there's some discussion of this elsewhere, but I can't seem to find it.)

Indeed... What's the practical way around it to recommend here? Should we avoid enabling loop alignment in the aarch64 backend, when SEH generation is enabled?

[hmartinez82]

🆙
Saw the same error when building with -mcpu=cortex-a55 in MSYS2's CLANGARM64 environment