[thotypous]

An x87 load/store or SSE conversion round trip quiets a signaling NaN. Kuna previously reduced it to a COPY: input 0x7f800001 returned unchanged instead of native 0x7fc00001.

Reproduce on an unmodified engine by adding only the independent test:

git fetch https://github.com/thotypous/kuna.git fix/preserve-float-quieting
git show FETCH_HEAD:decompiler/crates/kuna-cli/tests/float_nan_quieting.rs > decompiler/crates/kuna-cli/tests/float_nan_quieting.rs
printf '\n#[path = "float_nan_quieting.rs"]\nmod float_nan_quieting;\n' >> decompiler/crates/kuna-cli/tests/integration.rs
cd decompiler
cargo test --release -p kuna-cli --test integration float_format_roundtrips_preserve_native_nan_quieting

RuleFloatCast and SubfloatFlow now retain the conversion when its input might still be signaling. A bounded provenance walk permits existing folds after quieting operations or for non-NaN constants. nanignore all restores the previous behavior.

Keeping the IR conversion is insufficient: GCC also erased the old C cast pair with strict rounding flags. C emission therefore materializes a potentially quieting conversion in a volatile compound literal, evaluating its operand once. Already quiet inputs retain ordinary casts.

Validation: fresh x87 float32/float64 and SSE float32 fixtures, 4,108 patterns per format, four rounding modes, Clang/GCC at O0/O2, both nanignore none and compare; before failure and after pass. Also passed 39 rule tests, 17 cast-rendering tests, existing integer/intermediate rounding and native i386 return controls, and make check-spec. Full suite left to CI.

The nine affected corpus rendering assertions now require the retained volatile conversions. All 48 assertions in those four files pass locally; assertion names/counts and the recorded baseline are unchanged.

Also passed the 19 assertions in the four affected stage files. Their finite-only simplification controls now state nanignore all explicitly; strict NaN policies remain covered by native comparisons. No baseline counts were changed.

Updated nine integration rendering checks to require the volatile conversion expressions. All 56 tests across the six affected integration modules passed locally, including their generated-C recompilation checks. The first local pass lacked non-x86 specs for one module; rerunning that module with the installed complete spec tree passed all 16 tests.

[mahaloz]

[AUTOMATED] Pushed 34a4a43 plus two merges of origin/main (integration.rs and 09-emission.md conflicts).

Repro of the problem it addresses (i386, gcc -m32 -O0 -fno-pic -c):

void copyd(double *d, const double *s) { *d = *s; }

On the previous head, kuna decompile-all printed *a0 = (double)((volatile long double){ (long double)*a1 }); where main prints *a0 = *a1;. gcc emits fldl; fstpl for a plain copy, so every x87 float copy got the literal. Every widening or narrowing of an unknown float did too, (double)f used in arithmetic included. Across a 104-binary corpus that came to 178 changed functions (libm, libvorbis, libmpg123, libopus, libmp3lame, coreutils).

C does not promise to preserve signaling NaNs, so the quieting now applies only under --option nanignore none, the native-exact NaN mode (x87_integer_conversions_native already uses it). The default compare and all fold the round trip as before, and the datatest, stage and integration expectations this PR had rewritten are back to main's. Under none, only a widening whose result reaches a conversion back to its source format gets the volatile literal. Other conversions keep ordinary casts, because a C compiler cannot erase them. float_nan_quieting now checks the native comparison under none and checks that the default output has no literal.

Default output is byte-identical to the base on that corpus. The 14 remaining diffs also differ between two base runs. Under none the literal appears 0 times on libm, libvorbis, libvorbisfile, libmpg123, seq, du and numfmt.

[mahaloz]

[AUTOMATED] Pushed cdcc8d4 (+ merge of origin/main, 357711a). Under nanignore none, a widening that reaches its narrowing through a variable printed as v1 = ((volatile double){ (double)a1 }); return (float)v1;, and GCC 11/15 at -O2 drop a bare volatile compound literal in that position, so the regenerated code stopped quieting sNaN (e.g. void pickp(float*o,int c,const float*a,const float*b){double r; if(c) r=*a; else r=*b; *o=(float)r;} at -O0). The literal is now read as (*&(volatile double){ (double)x }), which GCC and Clang keep; float_nan_quieting gains a merged stack-slot round trip covering that shape. Default output is unchanged.

[mahaloz]

[AUTOMATED] Pushed d46b50b. Merged with main, implied_word_casts::implied_word_cast_keeps_float_member_bits_elf64 (#1244) failed. It ran under nanignore none, where the x87 flds; fstps pair after its word copy now keeps sNaN quieting instead of folding to a word COPY. That makes the pair a float reader, so the copy prints as items.real[0x10] = items.real[index];, which is the shape the default policy already prints when a copied word feeds float arithmetic (fadds). The test now runs under the default policy. That policy still folds the pair, so the test still checks the union bit-copy it was written for.