isnormal-pre-sm69
Status: stub. The full-length analysis is queued for a v1.0.x patch release per ADR 0018, section 5, criterion #6. The companion rule page at docs/rules/isnormal-pre-sm69.md contains the canonical detection logic + GPU reasoning.
TL;DR
isnormal is a portable numerical-classification primitive that mirrors the C99 isnormal semantic. Before SM 6.9, authors hand-rolled the equivalent check using bit-cast tricks: asuint(x) & 0x7F800000 against the IEEE exponent field, with separate tests for zero and subnormal. The hand-rolled form costs 4-6 instructions, plus a wave-wide bit-mask construction; the SM 6.9 intrinsic compiles to a single per-lane test on every IHV (NVIDIA Ada Lovelace, AMD RDNA 3/4, Intel Xe-HPG) because the floating-point classification logic is exposed by the underlying ISA.
What the rule fires on
A call to the isnormal intrinsic in a shader compiled against a target shader model older than SM 6.9. The isnormal intrinsic — which returns true when the argument is a normal IEEE float (not zero, not subnormal, not infinity, not NaN) — was added in SM 6.9. Earlier targets do not implement it and DXC issues a hard compile error. Slang reflection provides the target SM; the rule reads the target and fires when isnormal is called against any pre-SM-6.9 target.
See the What it detects section of the rule page for the full pattern definition.
Why it matters
The full GPU-mechanism analysis lives in the Why it matters on a GPU section of the companion rule page.
Examples
The bad / good code snippets are kept canonical on the rule page; see isnormal-pre-sm69.md -> Examples.
See also
- Rule page -- canonical reference + change log.
- math overview -- broader context.
- ADR 0018 -- v1.0 readiness plan.
This is a v1.0-ship stub. Full analysis pending; track issue link TBD.