You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
bench_audit hot loop emits 10 instructions per iteration; clang -O2 emits 7
for the same source. Quiet re-measurement 2026-11-16: t27b JIT about
1.9 ns/call, clang -O2 about 1.25 ns/call on this M-series Mac
(2026-11-15 "parity" was machine-load noise; corrected in release notes).
User scenario
Spec authors get about 1.5x slower loops than an optimizing C compiler on
dependency-chain-bound witnesses.
Requirements (from the asm diff, evidence in release notes)
SSA value-tracking: fold sub(add(x,c),c) back to x across the step and
the ifconv arm (clang: csel reads the OLD cur, no add/sub pair). Home:
PASS_SSA (STATUS_MISSING, DISP_BLOCKER in middle_end.t27).
Closed by the issue-shape gate (specs/policy/issue_shape.t27, #7498): an issue must have the template's sections. Edit the body; an edit that passes reopens it.
missing: ## User Scenarios -- at least one Given/When/Then scenario
missing: ## Requirements -- numbered obligations, FR-001: ... MUST ...
missing: ## Success Criteria -- an outcome a machine can settle: a number with a unit, a backticked file, exit 0, or a backticked command
Split recorded 2026-11-16: the LICM half of this issue is ready for merge as PR #14019 (branch claude/ssalicm-14012, spec licm_loop_imm + codegen; hot loop 10 -> 9 insns, A/A 203 files 0 regressions, spec tests 14/14).
The SSA half — folding sub(add(x,c),c) back to the old value so the ifconv arm reads the pre-increment register like clang's csel x21, x9, x20, hi — needs instruction scheduling / value-tracking across statements. That is PASS_SSA territory (STATUS_MISSING, DISP_BLOCKER in middle_end.t27); it stays open on this issue after the LICM PR lands. Interleaved bench on a loaded host: NEW 1.71 vs OLD 1.94 ns/call best-of-3; a firm wall-clock number needs the quiet host (lab runbook).
Reopened: the LICM half landed via PR #14019 (merge commit 3ec20ed on release/t27b-optimizer-parity-2026-11-15). The SSA half (fold sub(add(x,c),c) to the old value; PASS_SSA lane) stays open on this issue.
Closed by the issue-shape gate (specs/policy/issue_shape.t27, #7498): an issue must have the template's sections. Edit the body; an edit that passes reopens it.
missing: ## User Scenarios -- at least one Given/When/Then scenario
missing: ## Requirements -- numbered obligations, FR-001: ... MUST ...
missing: ## Success Criteria -- an outcome a machine can settle: a number with a unit, a backticked file, exit 0, or a backticked command
SSA half ready for merge: PR #14027 (branch claude/ssa-inc-14012-v2, based on master). INC-IF fold lands bench_audit at 8 hot-loop instructions -- 2 fewer than GCC -O2's 10; A/A 268 files 0 regressions; wall clock unchanged (~1.3 ns: the loop is dependency-chain bound, removed instructions were off the critical path).
Scope correction (owner, 2026-11-16): the landed SSA slice is the specific ssa_local_inc INC-IF fold plus ordered select emission — it is NOT the full PASS_SSA pass (no general value tracking, no live ranges). PASS_SSA stays STATUS_MISSING/DISP_BLOCKER in middle_end.t27; the disposition table should only change when the general pass lands. Also correcting the earlier wording: the 8-instruction state is a code-size win; the wall-clock <= clang criterion is not met (NEW ~1.36 vs clang ~1.2 ns, single quiet sample, no spread data).
Wording correction to the scope note above (owner, 2026-11-16): with PR #14027 still OPEN, the SSA slice is implemented on the branch (claude/ssa-inc-14012-v2) and PRESENTED as a PR — it is not landed, and must not be described as landed until that PR merges. What IS landed on master today: the LICM half (via the t27b release squash). The general PASS_SSA remains STATUS_MISSING / DISP_BLOCKER. And the <= clang criterion is unmet on both axes (instructions 8 > 7; wall clock NEW ~1.36 vs clang ~1.2 ns, single sample) — only the GCC -O2 instruction-count claim (10 -> 8) stands.
Reopened again: github-actions[bot] auto-closed this as NOT_PLANNED at 2026-11-15T19:45Z, 4 minutes after the manual reopen. The SSA PR #14027 is still OPEN; this issue stays open until it merges and the general PASS_SSA disposition question is resolved.
Closed by the issue-shape gate (specs/policy/issue_shape.t27, #7498): an issue must have the template's sections. Edit the body; an edit that passes reopens it.
missing: ## User Scenarios -- at least one Given/When/Then scenario
missing: ## Requirements -- numbered obligations, FR-001: ... MUST ...
missing: ## Success Criteria -- an outcome a machine can settle: a number with a unit, a backticked file, exit 0, or a backticked command
Activity
github-actions commentedon Nov 15, 2026
Closed by the issue-shape gate (specs/policy/issue_shape.t27, #7498): an issue must have the template's sections. Edit the body; an edit that passes reopens it.
FR-001: ... MUST ...exit 0, or a backticked commandt27b: correct bench_audit runtime claim -- quiet re-run shows clang -…
gHashTag commentedon Nov 15, 2026
Split recorded 2026-11-16: the LICM half of this issue is ready for merge as PR #14019 (branch
claude/ssalicm-14012, speclicm_loop_imm+ codegen; hot loop 10 -> 9 insns, A/A 203 files 0 regressions, spec tests 14/14).The SSA half — folding
sub(add(x,c),c)back to the old value so the ifconv arm reads the pre-increment register like clang'scsel x21, x9, x20, hi— needs instruction scheduling / value-tracking across statements. That is PASS_SSA territory (STATUS_MISSING, DISP_BLOCKER in middle_end.t27); it stays open on this issue after the LICM PR lands. Interleaved bench on a loaded host: NEW 1.71 vs OLD 1.94 ns/call best-of-3; a firm wall-clock number needs the quiet host (lab runbook).t27b: LICM loop consts live — hot loop 10→9 instructions, A/A 0 regre…
gHashTag commentedon Nov 15, 2026
Reopened: the LICM half landed via PR #14019 (merge commit 3ec20ed on release/t27b-optimizer-parity-2026-11-15). The SSA half (fold sub(add(x,c),c) to the old value; PASS_SSA lane) stays open on this issue.
github-actions commentedon Nov 15, 2026
Closed by the issue-shape gate (specs/policy/issue_shape.t27, #7498): an issue must have the template's sections. Edit the body; an edit that passes reopens it.
FR-001: ... MUST ...exit 0, or a backticked commandt27b: first-iteration peel — bench_audit to GCC -O2 parity; pass tele…
gHashTag commentedon Nov 15, 2026
SSA half ready for merge: PR #14027 (branch claude/ssa-inc-14012-v2, based on master). INC-IF fold lands bench_audit at 8 hot-loop instructions -- 2 fewer than GCC -O2's 10; A/A 268 files 0 regressions; wall clock unchanged (~1.3 ns: the loop is dependency-chain bound, removed instructions were off the critical path).
gHashTag commentedon Nov 16, 2026
Scope correction (owner, 2026-11-16): the landed SSA slice is the specific
ssa_local_incINC-IF fold plus ordered select emission — it is NOT the full PASS_SSA pass (no general value tracking, no live ranges). PASS_SSA stays STATUS_MISSING/DISP_BLOCKER in middle_end.t27; the disposition table should only change when the general pass lands. Also correcting the earlier wording: the 8-instruction state is a code-size win; the wall-clock <= clang criterion is not met (NEW ~1.36 vs clang ~1.2 ns, single quiet sample, no spread data).gHashTag commentedon Nov 16, 2026
Wording correction to the scope note above (owner, 2026-11-16): with PR #14027 still OPEN, the SSA slice is implemented on the branch (
claude/ssa-inc-14012-v2) and PRESENTED as a PR — it is not landed, and must not be described as landed until that PR merges. What IS landed on master today: the LICM half (via the t27b release squash). The general PASS_SSA remains STATUS_MISSING / DISP_BLOCKER. And the<= clangcriterion is unmet on both axes (instructions 8 > 7; wall clock NEW ~1.36 vs clang ~1.2 ns, single sample) — only the GCC -O2 instruction-count claim (10 -> 8) stands.gHashTag commentedon Nov 16, 2026
Reopened again: github-actions[bot] auto-closed this as NOT_PLANNED at 2026-11-15T19:45Z, 4 minutes after the manual reopen. The SSA PR #14027 is still OPEN; this issue stays open until it merges and the general PASS_SSA disposition question is resolved.
github-actions commentedon Nov 16, 2026
Closed by the issue-shape gate (specs/policy/issue_shape.t27, #7498): an issue must have the template's sections. Edit the body; an edit that passes reopens it.
FR-001: ... MUST ...exit 0, or a backticked commandMerge pull request #14027 from Trinity-S-AI-DNA/claude/ssa-inc-14012-v2