c(x86-64 SysV): converting x87 long double to _Float16 is refused ("does not support this runtime conversion to binary16")

[davidgmbb]

Observed

_Float16 h(long double a) { return a; }

$ build/Release/ide cc -c -target x86_64-linux h.c
cc: error: h.c:1:35: in function 'h': C IR lowering does not support this runtime conversion to binary16

Clang 18 (--target=x86_64-linux-gnu) compiles it to one call to __truncxfhf2, which libgcc has provided since GCC 12. The refusal is deliberate: the binary16 arm of c_ir_emit_cast in src/buster/lib/compiler/frontend/c/c_gen.c refuses wider-than-64-bit sources, because routing them through __truncsfhf2 would double-round. The parser mirrors it in c_parse_scalar_conversion_message (c_parse.c). No direct x87 entry point is wired in. Complex long double → _Float16 goes through the same path.

Expected

Lower (_Float16)x87_value to __truncxfhf2(long double) with the x87 operand in its usual memory/stack position, as Clang does. Also relax the parser mirror, which #1720 made target-aware for the AAPCS64 binary128 case (__trunctfhf2). Widening _Float16 → x87 should be checked at the same time; it is exact through binary32.

Details

• Revision c3f93e8 (branch of aarch64: lower binary128 long double through the compiler runtime #1720, which does not change the x87 path); Linux x86-64.
• Linking also needs libgcc, which buster's native link does not provide (driver/link: _Float16 conversions emit calls to libgcc helpers (__truncdfhf2, __extendhfsf2) that Buster's native link never provides #1272).

Done when

The function above compiles in all x86-64 SysV allocators with zero fallback, and a fixture checks the rounded binary16 bits against Clang for ties, subnormals, overflow to infinity and NaN.