Skip to content

[comgr][hotswap] Bound finite external PC call targets - #3606

Closed
harsh-amd wants to merge 1 commit into
ROCm:amd-stagingfrom
harsh-amd:users/harsh/hotswap-upstream-draft-04a-external-targets
Closed

[comgr][hotswap] Bound finite external PC call targets#3606
harsh-amd wants to merge 1 commit into
ROCm:amd-stagingfrom
harsh-amd:users/harsh/hotswap-upstream-draft-04a-external-targets

Conversation

@harsh-amd

@harsh-amd harsh-amd commented Jul 25, 2026 •
edited
Loading

Copy link
Copy Markdown

Scope

Bound calls whose complete target is proven but lies outside this object's
.text. The external address is not added to the local target set; the local
return continuation is protected. Proven register transfers are recorded
separately so unrelated instructions in the containing function remain
eligible for safe source relocation.

The bounded-return proof also treats an external call continuation entering a
helper's fallthrough chain as a second link-register provenance and rejects
that return conservatively. Immediate absolute, PC-relative, and canonical
get-PC/add register-call forms are covered. A returning finite call fails
closed unless its continuation is aligned and strictly inside .text; this
prevents TextSize or malformed/overflowed offsets from entering the local
target set.

Review range

  • Base: amd-staging at 50d16abb44335104f1f3b8ad8123d0b6aaa7f7d5
  • Head: f82d55cf79e6de8215445957e7169f06b7a79af8
  • This is a clean four-file feature diff with no cumulative prerequisite
    snapshot.
  • [comgr][hotswap] Close remaining offline corpus gaps #3598 remains unchanged as the integration reference.

Validation

  • HotswapMCTests --gtest_filter=CollectDirectBranchTargets.*: 26/26 passed.
  • Full HotswapMCTests: 130/130 passed.
  • Full supported COMGR LIT: 149/149 passed; 14 unsupported.
  • Exact-head ASAN/LeakSanitizer: focused 26/26 and full 130/130 passed with
    detect_leaks=1:halt_on_error=1:abort_on_error=1; no sanitizer findings.
  • Boundary unit coverage includes continuation == TextSize for
    materialized-register, immediate-absolute, and PC-relative external calls,
    plus misalignment and continuation overflow.
  • End-to-end LIT proves successful far rewrite, transformed disassembly,
    absence of unresolved-target fallback, and byte-level idempotency.
  • clang-format, no-added-auto, and git diff --check: clean.

Full corpus comparison

  • Corpus: 2,685 inputs; 2,558 success and 127 failure.
  • Transition matrix versus amd-staging: 2,558 SUCCESS -> SUCCESS, 127
    FAILURE -> FAILURE, zero regressions and zero improvements.
  • Evidence on mi300x:
    /home/harsh/hotswap-corpus-20260725-pr3606-f82d55cf-full2685
  • Built libamd_comgr.so SHA-256:
    fe81571b39fd14c052fa5519be8eb7004adedd1c830389339992496d588799e1
  • Manifest SHA-256:
    076ea013ba466139e76074fcbb31905a5041298e7644f6e8f590bf9353e4f5b9
  • Results SHA-256:
    67ff61bdcd3553047280f7381dea0ac6d8782eef9ca545724685a7c543122c18
  • Failures SHA-256:
    f51826f8183ae5b1c5de6b47ad9a88c4f0a25cca0b73c7cc9433a0d2b91aa300
  • Transitions SHA-256:
    ef79a3b151b11a5f44d5ba811ed03b3b2da60c33ac792bc7e4c26c5658155841

mi400-3 smoke

  • Dedicated container: hotswap-pr-review-root on
    ctheliosr-rck-g02-j19-13; four gfx1250 GPUs visible.
  • The installed /opt/therock/lib/libamd_comgr.so.3.3.0 resolved to the
    exact candidate with SHA-256
    fe81571b39fd14c052fa5519be8eb7004adedd1c830389339992496d588799e1.
  • Environment:
    ROCR_VISIBLE_DEVICES=0, HSA_HOTSWAP_VERBOSE=1,
    AMD_COMGR_HOTSWAP_ENTRY_TRAMPOLINES=0.
  • rocThrust CTest [SWDEV-457354] Fix flang-457354 test case #64 (merge_key_value.hip): 1/1 passed; embedded gtests:
    16/16 passed; 7.01 seconds.
  • Log:
    /home/harmenon/logs/hotswap-pr-review-20260725/rocthrust_test64_pr3606_f82d55cf.log
  • Log SHA-256:
    a3bd0d429be826d4e961298e0474497b5cff2b01eeeab9cfe3a2c33cfdb3605e

CI classification

The formatter, primary Linux and Windows builds, Linux and Windows COMGR tests,
SPIR-V translator tests, LLVM SPIR-V codegen tests, compiler/runtime stages,
and documentation build are green. Linux::release / Test rocm-examples
failed outside this COMGR change: HIP-Basic built and its 23 tests passed,
then the SPIR-V Applications/monte_carlo_pi build could not find
hipcub/iterator/counting_input_iterator.hpp. The larger multi-architecture
matrix is still running.

Remaining draft gate

Publish the instrumented source-coverage report, finish the required CI
classification, and demonstrate this correctness prerequisite's value through
a separately reviewed dependent corpus-reducing slice. Keep this PR draft
until those gates are recorded.

@harsh-amd
harsh-amd force-pushed the users/harsh/hotswap-upstream-draft-04a-external-targets branch 2 times, most recently from f84293d to c570b29 Compare July 26, 2026 00:02
@harsh-amd
harsh-amd force-pushed the users/harsh/hotswap-upstream-draft-04a-external-targets branch from c570b29 to f82d55c Compare July 26, 2026 03:44
This was referenced Jul 26, 2026
Closed
Merged
Closed
@lamb-j lamb-j added comgr Related to Code Object Manager hotswap Related to the Comgr Hotswap feature labels Jul 27, 2026
nirmie added a commit to harsh-amd/llvm-project that referenced this pull request Jul 27, 2026
@nirmie
comgr: address review feedback on reusable PC call analysis
3563c98 
Fix three soundness holes in the reusable PC-materialized call analysis and
tighten the surrounding code and tests, per review on ROCm#3576:

- Fail closed when the reaching-value worklist reaches an undecoded slot;
  previously the finite target set flowed through unknown bytes that could
  clobber the target pair or divert control flow.
- Recompute Resolved[I] on every reusable-call visit so a reconvergent path
  that makes the call Unknown clears an earlier finite result instead of
  leaving it stale.
- Insert each bounded s_set_pc_i64 return continuation into Info.Targets so
  the return's landing sites are protected from relocation, matching the
  forward call path (and ROCm#3606).
- Read the carry-form first addend through the MC operand
  (evaluateAbsoluteUint32Operand) instead of peeking at raw .text bytes; the
  disassembler already exposes the SOP2 literal as an immediate. Drop the
  now-unused Text parameter from the two inner helpers and the Endian.h
  include.
- Log the shared far-dispatch dispatcher-prefix overflow instead of returning
  false silently.

Tests: add undecoded_gap and reconverge_reload fail-closed cases to
hotswap-reusable-pc-call-targets.s; add plain and inline-immediate carry
materialization unit tests; update BoundsCanonicalSetPcReturn to expect the
now-protected return continuation.

Full corpus is byte-identical before and after (the bugs are only reachable by
malformed or hand-crafted input, not compiled output); all unit and lit suites
pass.
@nirmie nirmie mentioned this pull request Jul 27, 2026
Merged
nirmie added a commit to harsh-amd/llvm-project that referenced this pull request Jul 27, 2026
@nirmie
comgr: address review feedback on reusable PC call analysis
874be65 
Fix three soundness holes in the reusable PC-materialized call analysis and
tighten the surrounding code and tests, per review on ROCm#3576:

- Fail closed when the reaching-value worklist reaches an undecoded slot;
  previously the finite target set flowed through unknown bytes that could
  clobber the target pair or divert control flow.
- Recompute Resolved[I] on every reusable-call visit so a reconvergent path
  that makes the call Unknown clears an earlier finite result instead of
  leaving it stale.
- Insert each bounded s_set_pc_i64 return continuation into Info.Targets so
  the return's landing sites are protected from relocation, matching the
  forward call path (and ROCm#3606).
- Read the carry-form first addend through the MC operand
  (evaluateAbsoluteUint32Operand) instead of peeking at raw .text bytes; the
  disassembler already exposes the SOP2 literal as an immediate. Drop the
  now-unused Text parameter from the two inner helpers and the Endian.h
  include.
- Log the shared far-dispatch dispatcher-prefix overflow instead of returning
  false silently.

Tests: add undecoded_gap and reconverge_reload fail-closed cases to
hotswap-reusable-pc-call-targets.s; add plain and inline-immediate carry
materialization unit tests; update BoundsCanonicalSetPcReturn to expect the
now-protected return continuation.

Full corpus is byte-identical before and after (the bugs are only reachable by
malformed or hand-crafted input, not compiled output); all unit and lit suites
pass.
nirmie added a commit to harsh-amd/llvm-project that referenced this pull request Jul 27, 2026
@nirmie
comgr: address review feedback on reusable PC call analysis
979d238 
Fix three soundness holes in the reusable PC-materialized call analysis and
tighten the surrounding code and tests, per review on ROCm#3576:

- Fail closed when the reaching-value worklist reaches an undecoded slot;
  previously the finite target set flowed through unknown bytes that could
  clobber the target pair or divert control flow.
- Recompute Resolved[I] on every reusable-call visit so a reconvergent path
  that makes the call Unknown clears an earlier finite result instead of
  leaving it stale.
- Insert each bounded s_set_pc_i64 return continuation into Info.Targets so
  the return's landing sites are protected from relocation, matching the
  forward call path (and ROCm#3606).
- Read the carry-form first addend through the MC operand
  (evaluateAbsoluteUint32Operand) instead of peeking at raw .text bytes; the
  disassembler already exposes the SOP2 literal as an immediate. Drop the
  now-unused Text parameter from the two inner helpers and the Endian.h
  include.
- Log the shared far-dispatch dispatcher-prefix overflow instead of returning
  false silently.

Tests: add undecoded_gap and reconverge_reload fail-closed cases to
hotswap-reusable-pc-call-targets.s; add plain and inline-immediate carry
materialization unit tests; update BoundsCanonicalSetPcReturn to expect the
now-protected return continuation.

Full corpus is byte-identical before and after (the bugs are only reachable by
malformed or hand-crafted input, not compiled output); all unit and lit suites
pass.
SourishW added a commit to harsh-amd/llvm-project that referenced this pull request Jul 28, 2026
@SourishW
[comgr][hotswap] rebase ROCm#3610 onto current ROCm#3576 and ROCm#3606
25c4380 
Preserve Harsh Menon�[27m's original ROCm#3610 commit and apply the validated rebase delta as one follow-up commit.
SourishW added a commit to harsh-amd/llvm-project that referenced this pull request Jul 28, 2026
@SourishW
[comgr][hotswap] rebase ROCm#3610 onto current ROCm#3576 and ROCm#3606
3ef92e3 
Preserve Harsh Menon�[27m's original ROCm#3610 commit and apply the validated rebase delta as one follow-up commit.
SourishW added a commit to harsh-amd/llvm-project that referenced this pull request Jul 28, 2026
@SourishW
[comgr][hotswap] rebase ROCm#3610 onto current ROCm#3576 and ROCm#3606
abc4013 
Preserve Harsh Menon�[27m's original ROCm#3610 commit and apply the validated rebase delta as one follow-up commit.
SourishW added a commit to harsh-amd/llvm-project that referenced this pull request Jul 28, 2026
@SourishW
[comgr][hotswap] rebase ROCm#3610 onto current ROCm#3576 and ROCm#3606
81967d5 
Preserve Harsh Menon�[27m's original ROCm#3610 commit and apply the validated rebase delta as one follow-up commit.
SourishW added a commit to harsh-amd/llvm-project that referenced this pull request Jul 28, 2026
@SourishW
[comgr][hotswap] rebase ROCm#3610 onto current ROCm#3576 and ROCm#3606
9c95d90 
Preserve Harsh Menon�[27m's original ROCm#3610 commit and apply the validated rebase delta as one follow-up commit.
SourishW added a commit to harsh-amd/llvm-project that referenced this pull request Jul 28, 2026
@SourishW
[comgr][hotswap] rebase ROCm#3610 onto current ROCm#3576 and ROCm#3606
7df2ede 
Preserve Harsh Menon�[27m's original ROCm#3610 commit and apply the validated rebase delta as one follow-up commit.
SourishW added a commit to harsh-amd/llvm-project that referenced this pull request Jul 28, 2026
@SourishW
comgr: finalize ROCm#3576 and ROCm#3606 integration
0d9b2f3
SourishW added a commit to harsh-amd/llvm-project that referenced this pull request Jul 28, 2026
@SourishW
comgr: finalize ROCm#3576 and ROCm#3606 integration
b8b7c13
SourishW added a commit to harsh-amd/llvm-project that referenced this pull request Jul 28, 2026
@SourishW
comgr: finalize ROCm#3576 and ROCm#3606 integration
8aaf7e6
@harsh-amd harsh-amd closed this Aug 3, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@lamb-j lamb-j Awaiting requested review from lamb-j lamb-j will be requested when the pull request is marked ready for review lamb-j is a code owner
@chinmaydd chinmaydd Awaiting requested review from chinmaydd chinmaydd will be requested when the pull request is marked ready for review chinmaydd is a code owner

Assignees

No one assigned

Labels

comgr Related to Code Object Manager hotswap Related to the Comgr Hotswap feature

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@harsh-amd @lamb-j