Hi Paul-Antoine,
Please find attached a revised version of the patch.
Compared to the previous iteration, I have:
* Rebased on top of Pan's work;
* Updated the cost model;
* Added a second pattern to handle the case where PLUS_MINUS operands
are swapped;
* Added compile and run tests.
I bootstrapped and regtested against rv64gcv.
We need to replace all "FR to VR" uses with the new function and IMHO
better done in a separate patch rather than folded into this one.
Also, please add f16 tests still. Not that I expect fallout but just for
completeness's sake.
Please CC patchworks...@rivosinc.com for the next version so we have CI
coverage.
+(define_insn_and_split "*<optab>_vf_<mode>"
+ [(set (match_operand:V_VLSF 0 "register_operand" "=vd")
+ (plus_minus:V_VLSF
+ (mult:V_VLSF
+ (vec_duplicate:V_VLSF
+ (match_operand:<VEL> 1 "register_operand" " f"))
+ (match_operand:V_VLSF 2 "register_operand" " 0"))
+ (match_operand:V_VLSF 3 "register_operand" " vr")))]
+ "TARGET_VECTOR && can_create_pseudo_p ()"
Nit: The constraints look unaligned here, or is that just my editor?
In the run test params you use
+/* { dg-additional-options "--param=gpr2vr-cost=0" } */
That should rather be fpr2vr-cost=.
Are you going to do the other fma variants in follow ups still?
--
Regards
Robin
