netdev: card.resource stubs list every scratch register the call may destroy (#70, A of 3) - #71
Merged
Merged
Conversation
…destroy The five card.resource stubs in netdev_pcmcia.c passed their argument registers (a1, and a0/d1 where used) as inputs only. The library ABI lets every LVO destroy d0/d1/a0/a1, but GCC was told those registers still held their values after the jsr. This was REAL in anxnet.device at main 7108650. pc_misc_control loads d1 = 10 for CardMiscControl (jsr -48 at 0x68aa), and GCC reused d1 as the value argument of netdev_diag_note(ANXDIAG_PC_IOMODE) at 0x68b0, so the diagnostic IOMODE value was whatever card.resource left in d1. The other four stubs were latent: today's code rewrites or never reads the register. Each argument register is now a "+r" operand; unused scratch registers remain clobbers. After the change, 0x68b0 is `pea 0xa` (the constant), and the OwnCard / ReleaseCard / CardResetCard / CopyTuple sites compile to the same instructions, only shifted by 2 bytes. The #70 scanner no longer flags netdev_pcmcia.c. anxnet.device 7d62fbe4 -> 9e8e9011 (default preset). Part of #70 (shipped sites, PR A of 3). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
tinic
force-pushed
the
fix/lvo-clobber-pcmcia
branch
from
September 26, 2026 02:30
367c60d to
8c300b0
Compare
tinic
added a commit
that referenced
this pull request
Sep 26, 2026
…call may destroy (#73) installnetprobe.c's three card.resource stubs (OwnCard, ReleaseCard, CopyTuple) passed their argument registers as asm inputs only. The library ABI lets every LVO destroy d0/d1/a0/a1. This was REAL at main 66e0bf5. probe_own_card loads a1 = &handle (0x968) for OwnCard (jsr -6 at 0x96e). On the owned path, the first CopyTuple(0x20) (jsr -72 at 0x98e) was issued without reloading a1, so it used whatever OwnCard left there. After the change, 0x986 is `lea 0x18,a1` before the CopyTuple, which reloads the handle. The other two stubs were latent. Each argument register is now a "+r" operand; unused scratch registers remain clobbers. The #70 scanner no longer flags installnetprobe.c. With #71 and #72 this closes the 16 shipped sites in #70. Part of #70 (shipped sites, PR C of 3). Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This was referenced Sep 26, 2026
tinic
added a commit
that referenced
this pull request
Sep 26, 2026
…ers; CI gate (#70) (#86) An LVO may destroy d0, d1, a0 and a1. An extended asm statement that lists one of them only as an input lets GCC reuse a value the call has overwritten, which is how #68 passed a stale recvfrom address. tools/check-lvo-clobbers.sh (python in check-lvo-clobbers.py) scans src/ and tests/ for asm statements whose template calls through a6 and reports each one whose outputs and clobbers leave out a scratch register. It reads register declarations through object-like macros (BSD_SCRATCH, NETDEV_REG_A1) from the file and its "..." includes, and skips top-level asm. Output is key=value, exit 1 on any site. It is wired into stage_host next to check-lvo-matrix.sh, registered in check-gates-wired.sh, and added to the shellcheck list. Validation: 58c5c4f (the tree the issue scanned): 415 statements, 31 flagged. All 16 shipped sites match the issue's missing sets exactly, and tests/sockopt/sockopt_test.c:332 (pre-#68) is flagged missing=a1. src/tools/toolsock.c alone: 22 statements, 0 flagged. The issue's 77 included 46 false positives: 43 BSD_SCRATCH macro sites, apidrill's TH() macro, and the top-level asm in genet.c and netdev_cache.c. The 14 real test sites on main are fixed: a register that is also an input becomes an output (dummy "=r", or "+r" as in #71), one that is not becomes a clobber. udpdrill's recvfrom and getsockopt helpers had the #68 shape (a1). before (7fa3df3): statements=415 files=50 flagged=14 after: statements=415 files=50 flagged=0 Fixes #70 Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part of #70, shipped-site fixes, PR A of 3:
src/netdev/netdev_pcmcia.c→ anxnet.device.The five card.resource stubs passed their argument registers (a1, plus a0/d1 where used) as asm inputs only. Any LVO may destroy d0/d1/a0/a1, but GCC was told they survived the call.
REAL today (main 7108650, anxnet.device 7d62fbe4):
After (anxnet.device 9e8e9011):
The other four stubs (OwnCard -6, ReleaseCard -12, CardResetCard -66, CopyTuple -72) were latent. Before and after they compile to the same instructions, shifted by 2 bytes.
Change: each argument register is now a
"+r"operand, and unused scratch registers stay clobbers. There is a comment at the stub header, a CHANGELOG entry, and no other code change.Proof:
tools/ci.sh hostin CI conditions: rc 0, 464/464.🤖 Generated with Claude Code