Mirror cstage's C-variadic call handling in the ww self-host: parse a
bare `...` param (decl.ww), skip param-keyed desugar for it to avoid a
nil-deref (check.ww), and emit AL = XMM-reg count plus CVTSS2SD
promotion of f32 args in the variadic tail (cgenutil.ww, cgenexpr.ww).
Closes the cat-A wwstage silent miscompile (AL=0, unpromoted f32 tail).
Parse/check/cgen are one atomic align-up (parse alone miscompiles, so
not bisect-splittable). 989_ffivariadic now runs dual-stage (cstage ww
+ wwstage ww_ww), 12/12; w6c==w6c_ww byte-identical. Byte-id alone is
blind here (the bootstrap calls no float-bearing C variadic), so the
ww_ww runtime rows are the real net.
C99 6.5.2.2p6 requires float->double promotion for floats in the
variadic region of a call; w6c emitted MOVSS (4B), so a C variadic's
va_arg(double) read 8B of garbage. Promote at push (CVTSS2SD) so the
8B slot holds a real double, covering the XMM-reg and >8-float spill
cases uniformly; fi/AL XMM-count is unchanged.
cstage only; the wwstage twin rides C2 (#10), which adds wwstage's
C-variadic-call codegen.
SysV §3.5.7 requires a variadic call to set AL = number of vector (XMM) regs used for the variable float args; the C callee gates its xmm-save-area stores on `test %al,%al`, so the old hardcoded XORQ AX,AX (AL=0) made va_arg(double) read garbage for any float-bearing C variadic call. Emit MOVQ $fi,AX (fi = the in-scope XMM cursor, ≤8); w6a has no MOVL-immediate encoding so MOVQ is the assemblable form and sets AL=fi identically. fi==0 keeps XORQ → byte-identical to pre-fix for no-float variadic calls. Runtime test 989_ffivariadic links a cc-compiled va_arg(double) fixture (zero relocs/undefined, w6l-linkable) and sweeps N=3/5/8 floats (N=2 is vacuous via stale-stack aliasing). C1 of the C-FFI-variadic align-up (USER ruling); C2 wwstage + C3 bodiless gate follow. ref/qbe/amd64/sysv.c:384. 454 green.
