https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118133
Bug ID: 118133
Summary: Consider alternative ways of writing
aarch64-simd-pragma-builtins.def
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: build
Severity: enhancement
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rsandifo at gcc dot gnu.org
Target Milestone: ---
Target: aarch64*-*-*
Created attachment 59927
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59927&action=edit
proof of concept patch
aarch64-simd-pragma-builtins.def uses a hierarchy of C preprocessor macros to
build up the individual definitions. That's probably going to become a bit
unwieldly, so it would be good to find an alternative method.
One challenge would be to this while staying within the permitted set of
required build tools.
One idea I had was to use m4. We don't currently require that to build gcc,
only to modify it, but that should be enough if we commit the autogenerated
output to the repository.
For example:
nonstreaming_only([FAMINMAX],
foreachn([[T],[Q]], [VHSDF],
[ENTRY([vamax,Q,_,T], binary, [[T,Q],[T,Q],[T,Q]], UNSPEC_FAMAX, DEFAULT)],
[ENTRY([vamin,Q,_,T], binary, [[T,Q],[T,Q],[T,Q]], UNSPEC_FAMIN, DEFAULT)],
),
)
nonstreaming_only([FP8],
foreachn([[T],[Q]], [VHSDF],
[ENTRY([vscale,Q,_,T], binary, [[T,Q],[T,Q],[signed(T),Q]],
UNSPEC_FSCALE, DEFAULT)],
),
)
could be used to generate:
#define REQUIRED_EXTENSIONS nonstreaming_only (AARCH64_FL_FAMINMAX)
ENTRY(vamax_f16, binary, f16, f16, f16, none, UNSPEC_FAMAX, DEFAULT)
ENTRY(vamin_f16, binary, f16, f16, f16, none, UNSPEC_FAMIN, DEFAULT)
ENTRY(vamaxq_f16, binary, f16q, f16q, f16q, none, UNSPEC_FAMAX, DEFAULT)
ENTRY(vaminq_f16, binary, f16q, f16q, f16q, none, UNSPEC_FAMIN, DEFAULT)
ENTRY(vamax_f32, binary, f32, f32, f32, none, UNSPEC_FAMAX, DEFAULT)
ENTRY(vamin_f32, binary, f32, f32, f32, none, UNSPEC_FAMIN, DEFAULT)
ENTRY(vamaxq_f32, binary, f32q, f32q, f32q, none, UNSPEC_FAMAX, DEFAULT)
ENTRY(vaminq_f32, binary, f32q, f32q, f32q, none, UNSPEC_FAMIN, DEFAULT)
ENTRY(vamaxq_f64, binary, f64q, f64q, f64q, none, UNSPEC_FAMAX, DEFAULT)
ENTRY(vaminq_f64, binary, f64q, f64q, f64q, none, UNSPEC_FAMIN, DEFAULT)
#undef REQUIRED_EXTENSIONS
#define REQUIRED_EXTENSIONS nonstreaming_only (AARCH64_FL_FP8)
ENTRY(vscale_f16, binary, f16, f16, s16, none, UNSPEC_FSCALE, DEFAULT)
ENTRY(vscaleq_f16, binary, f16q, f16q, s16q, none, UNSPEC_FSCALE, DEFAULT)
ENTRY(vscale_f32, binary, f32, f32, s32, none, UNSPEC_FSCALE, DEFAULT)
ENTRY(vscaleq_f32, binary, f32q, f32q, s32q, none, UNSPEC_FSCALE, DEFAULT)
ENTRY(vscaleq_f64, binary, f64q, f64q, s64q, none, UNSPEC_FSCALE, DEFAULT)
#undef REQUIRED_EXTENSIONS
I've attached a proof of concept patch.
