Shader coverage: test the coalescing exit analysis directly in slang-static-unit-test

[jvepsalainen-nv]

Problem Description

Four properties of the coverage coalescing analysis are only testable end-to-end today, and three of them cannot be expressed in Slang source at all. Each came up during review of #12683 and had to be recorded as untestable rather than covered.

The analysis lives in source/slang/slang-ir-coverage-instrument.cpp:

• CoverageFunctionExitAnalysis — decides whether a call can abandon the invocation, which is what splits a coalescing region.
• assignCoverageCounterSlots — groups markers into regions and picks which one emits the runtime probe.

What cannot be tested through a .slang test

1. GenericAsm counts as a normal exit. This is what stops a call to ordinary builtin math from splitting every region. It cannot be isolated from Slang source: sqrt, lerp and max all contain an ordinary return_val alongside their GenericAsm, so sawNormalExit is already true and flipping the clause would not change their classification. A regression here would defeat coalescing for essentially all numeric code with no test failing.

2. The probe sits at the last marker of a region. The design depends on this: reaching the probe must prove every earlier marker in the region executed. Within a straight-line block, first-marker and last-marker placement are count-equivalent, so every existing test passes either way. Nothing would catch a regression that moved it.

3. kIROp_Abort splits a region. It shares mayNotFallThrough with discard, which is covered, but is not reachable on its own from Slang source.

4. Mutual recursion is broken conservatively. Coalesce line coverage counters per straight-line region #12683 fixed a real soundness bug here: an optimistic cycle break escaped into other functions' memoized results, so one partner of a mutually recursive pair could be cached as "returns normally" when it does not, with the outcome depending on traversal order. The fix is covered only indirectly.

Preferred Solution

Test these directly in slang-static-unit-test, which exists for exactly this — non-exported compiler entry points exercised on hand-built IR. IRFixtureBuilder already offers what is needed: build a function whose only exit is a GenericAsm, or a mutually recursive pair where one partner never returns, and assert what the analysis concludes. None of those modules can be produced by the frontend.

For the probe-placement property, call assignCoverageCounterSlots on a known marker sequence and assert which marker carries the probe, rather than inferring it from a slot count.

Cost, and the reason this is not free

Both targets currently have internal linkage:

static IRFunc* getStaticallyResolvedCallee(...)   // slang-ir-coverage-instrument.cpp
struct CoverageFunctionExitAnalysis               // file-local
static void assignCoverageCounterSlots(...)

Static linking reaches non-exported symbols, not static-qualified ones, so this needs assignCoverageCounterSlots and the analysis declared in slang-ir-coverage-instrument.h. That widens a header for testability, which is worth weighing against the four properties gained — the codebase is otherwise wary of changes that exist to make tests pass.

An alternative worth considering first: keep the header unchanged and test through instrumentCoverage on hand-built IR, asserting on the emitted probes rather than calling the internals directly. That covers properties 1, 3 and 4 without a linkage change, though probably not 2.

Additional Context

Raised during review of #12683 (line coverage coalescing), where each of these was answered with "cannot be isolated by test" rather than a fix. The static suite landed in #12347 after that work started.

Test Plan

• GenericAsm-only exit is treated as returning normally.
• A function with no reachable return splits its caller's region; one with a conditional discard splits through the abandons-the-invocation path instead.
• kIROp_Abort mid-region splits it.
• Mutually recursive A/B where A never returns yields the same conservative answer whichever partner the walk reaches first.
• Given a marker sequence in one block, the last marker carries the probe.

[nv-slang-bot]

Triage — plan-ready, blocked on prerequisite. (Updated: approach verified against source; now awaiting a dependency.)

• Classification: enhancement (test-coverage) · IR-pass shader-coverage + CI/testing · low severity. Scope confirmed accurate; the four properties are correctly identified as not expressible from Slang source (property 2 — last-marker probe placement — has no coverage today).
• Blocked on Coalesce line coverage counters per straight-line region #12683 (still open). The symbols to be tested (CoverageFunctionExitAnalysis, assignCoverageCounterSlots, getStaticallyResolvedCallee) are introduced by Coalesce line coverage counters per straight-line region #12683 and don't exist on master yet. Implementation is prepared but will not land until Coalesce line coverage counters per straight-line region #12683 merges; if Coalesce line coverage counters per straight-line region #12683 is redesigned before merge (e.g. the DFS cycle-break becomes an iterated fixpoint) the plan is re-derived.
• Confirmed approach: expose only assignCoverageCounterSlots in slang-ir-coverage-instrument.h (drop its static) and assert on outSlots / outEmitsProbe / outCounterCount. Covers all four properties, including property 2 directly via outEmitsProbe — the one the blackbox-via-instrumentCoverage alternative cannot reach. CoverageFunctionExitAnalysis stays file-local (it's exercised through the region-join scan). This is confirmed mandatory, not cosmetic: slang-static-unit-test links the compiler statically, which reaches non-exported extern symbols but not static-qualified ones; the fallback is a thin non-static wrapper. Widening this internal header for testability is the sanctioned pattern the static suite (Add slang-static-unit-test for unit-testing non-exported compiler entry points #12347) exists for — unit-test-ir-dce.cpp calls the non-exported eliminateDeadCode the same way. No ABI concern (internal header, not include/).
• Implementer note: IRFixtureBuilder builds only void functions today — the GenericAsm-only exit, no-return / Abort function, mutually-recursive A/B pair, and single-block marker sequence need new builder methods (or direct IRBuilder use in the test). Each test should be revert-drilled (reintroduce the regression, confirm the test fails) before it's trusted.
• One question for @jvepsalainen-nv: slang-static-unit-test is gated on SLANG_LIB_TYPE=STATIC + SLANG_ENABLE_SLANG_RHI. Does a CI job build that configuration? If not, the new tests would compile but never run in CI — worth confirming so this coverage isn't dead weight.

Next step: ready-for-fix once #12683 merges (blocked-on-dependency until then). No code lands ahead of #12683.

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}