stillwater-sc / universal / 26919578707

85%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

perf(elreal,ci): cut sanitizer/Debug CI time ~20x (depth, ASan -O1, ctest -j) (#1047)

The ASan/UBSan jobs grew past an hour because the elreal transcendental tests
(el_math_trigonometry/hyperbolic/exponent/constants) run a heavy O(depth^4)
workload at -O0 with no test parallelism. Four independent, multiplicative levers,
none of which lose any asserted coverage:

1) Test generation depth (the dominant, quartic lever). The math functions default
   to depth 4 (~212 bits) for high-precision use, but the tests assert only loose
   host tolerances (1e-10/1e-5) plus structural identities and the 0-overlap
   invariant -- all satisfied with wide margin at far smaller depth (d ~ d*53 bits
   on a double host). Thread an explicit small test depth (LIBRARY DEFAULTS
   UNCHANGED): single-series fns (atan/sin/cos/tan/exp/log/sinh/cosh/tanh) at
   depth 2 (~106 bits); the deepest compositions (asin/acos, pow=exp(y*log(x))) at
   depth 3; constants generated at depth 6 instead of the 16/32 defaults, with one
   deliberately-deep pi check kept at 8. Full elreal Debug+assertions ctest drops
   178s -> 8.2s locally (constants alone 175s -> 8.3s), 32/32 pass on gcc + clang.

2) ASan builds at -O1 (CMakeLists.txt). AddressSanitizer's recommended setting:
   2-4x faster than -O0 on this template-heavy header-only library with
   equal-or-better detection (Debug -g retained for stacks). UBSan deliberately
   stays at -O0: at -O1 the optimizer can elide UB sites before instrumentation
   sees them, so -O0 preserves maximum UB-detection fidelity. UBSan still gets the
   depth + ctest -j speedups, just not the -O multiplier.

3) ctest -j 2 in sanitizers.yml (was serial). Matches the 2-vCPU ubuntu-latest
   runner; safe under ASan's ~2-3x shadow-memory multiplier in 7 GB (the sibling
   cmake.yml already runs ctest -j 4 on the same runner).

4) --timeout 180 in sanitizers.yml: caps a hung instrumented test at 180s instead
   of CTest's 1500s default (tail-latency insurance,... (continued)

Coverage Stats

1 of 1 new or added line in 1 file covered. (100.0%)

24 existing lines in 5 files now uncovered.

47552 of 56265 relevant lines covered (84.51%)

5996899.76 hits per line