[PATCH v2 0/2] bpf: fix gotol with a signed offset, and run the spacing test

[Vineet Gupta]

In the pseudo-C dialect `gotol +1' is silently assembled as `goto l+1'
leading to libbpf load time failures. LLVM obviously supports this so gas
needs fixing.

Ran into this when trying to enable certain selftests which are
currently clang built only while GCC can build them fine. Moreover those
not run tests are declared passing too [1].

Patch 1: Orthogonal issue discovered in the process. Directed tests for
exactly this exist: spacing-pseudoc.{s,d} but are never run.

Patch 2 fixes the actual mis-assembly.

As part of v1 review [2] it was suggested to simply use 'W%' to mandate
at least whitespace i.e. `goto%W%d16'. However that does not work.
gas strips the whitespace next to a sign in the operand field before
md_assemble sees the line, so rejects valid insn: `if rX > N goto +M'
The registered testsuite does not catch either, as it only ever
writes conditional jumps with label operands. So v2 retains the v1
approach to reject a template whose literal text stops in the
middle of a name, letting a later template match the whole mnemonic.

Tested on bpf-unknown-none: 59 expected passes, no failures.

[1] https://lore.kernel.org/bpf/20260821201316.1059129-1-vineet.gupta@linux.dev/
[2] https://gcc.gnu.org/pipermail/bpf/2026-August/000192.html

Vineet Gupta (2):
  gas: bpf: actually run the pseudo-C spacing test
  PR 34558: bpf: don't mis-assemble `gotol' with signed offset

 gas/config/tc-bpf.c                           | 28 ++++++++++++++++++-
 gas/testsuite/gas/bpf/bpf.exp                 |  3 ++
 .../gas/bpf/jump-gotol-signed-pseudoc.d       | 15 ++++++++++
 .../gas/bpf/jump-gotol-signed-pseudoc.s       | 10 +++++++
 gas/testsuite/gas/bpf/spacing-pseudoc.d       |  2 +-
 5 files changed, 56 insertions(+), 2 deletions(-)
 create mode 100644 gas/testsuite/gas/bpf/jump-gotol-signed-pseudoc.d
 create mode 100644 gas/testsuite/gas/bpf/jump-gotol-signed-pseudoc.s

-- 
2.53.0-Meta