Research-2170: Windows icx-cl and icpx strict FP spelling — 2026-10-07¶

Finding¶

On the gate commit of the MSVC warnings-as-errors work (job 112842753545 of run 37635551653), Windows MSVC+SYCL printed 2,268 of its 4,361 warnings as overriding '-ffp-model=precise' option with '-ffp-contract=off' [-Woverriding-option], from two places:

every C and C++ compile by icx-cl (1,955), from the project-wide vmaf_strict_fp_args of intel-llvm-cl, /fp:precise /Qfma-: icx-cl maps /fp:precise to -ffp-model=precise and /Qfma- to -ffp-contract=off;
every SYCL compile and device link by icpx (the MSVC build compiles its SYCL translation units and their explicit device link, ADR-1364, with the GNU-syntax icpx), from sycl_strict_fp_args, which kept the two-flag spelling on that build.

ADR-2170 removed the same override for the Unix icx and icpx by spelling -fno-fast-math -fcomplex-arithmetic=full between the model and the contraction flag. The question here is whether the same line, through /clang: for icx-cl and verbatim for the Windows icpx, sets the same floating-point state on the Windows target.

Method¶

oneAPI 2026.0.0 (icx-cl and icpx of the Linux installation, which compile for --target=x86_64-pc-windows-msvc without the Windows SDK as long as the translation unit includes no header). The Windows leg pins 2025.3.0.372; the CI run of the pull request is the check on that version.

icx-cl, old /fp:precise /Qfma- against new /fp:precise /clang:-fno-fast-math /clang:-fcomplex-arithmetic=full /clang:-ffp-contract=off: the -cc1 line of -###, the predefined macros (/E /clang:-dM), and the objects of a corpus of a * b + c in double and float, a / b + a * b, a dot-product loop, a float/double mix and complex multiply and divide (double and float), at /O2, /O3, /O3 /arch:AVX2 and /O3 /arch:CORE-AVX512.
icpx with -fsycl -fsycl-targets=spir64_gen,spir64, old -fp-model=precise -ffp-contract=off against new -fp-model=precise -fno-fast-math -fcomplex-arithmetic=full -ffp-contract=off, both followed by -foffload-fp32-prec-div -foffload-fp32-prec-sqrt: the -cc1 lines of the host and both device passes, the -fsycl-link expansion, the predefined macros, and the -fsycl-device-only bitcode of the same corpus.

Results¶

Check	icx-cl	icpx (SYCL)
Override warning	1 per compile old, 0 new	3 per compile and 1 per device link old, 0 new
`-cc1` lines	equal apart from `-fcomplex-arithmetic=full`	equal apart from `-fcomplex-arithmetic=full` (host, spir64_gen, spir64, device link)
Predefined macros	402, equal	970, equal
Output	objects byte-identical at all four settings	device bitcode byte-identical

The new line is the same arithmetic. -fcomplex-arithmetic=full restates the complex-arithmetic mode -fp-model=precise sets and -fno-fast-math would otherwise reset, as on Linux.

Consequence¶

core/src/meson.build spells the icx-cl strict line /fp:precise /clang:-fno-fast-math /clang:-fcomplex-arithmetic=full /clang:-ffp-contract=off and gives the MSVC build's icpx the same reset as Linux. core/test/test_strict_fp_compiler_args.py executes both and requires every icx-cl compile command to end its FP flags on /clang:-ffp-contract=off.