[llvm] [WebAssembly] Lower fmuladd to madd and nmadd (PR #161355)

[Sam Parker via llvm-commits]

================
@@ -1597,16 +1597,20 @@ multiclass SIMDMADD<Vec vec, bits<32> simdopA, bits<32> simdopS, list<Predicate>
            vec.prefix#".relaxed_nmadd\t$dst, $a, $b, $c",
            vec.prefix#".relaxed_nmadd", simdopS, reqs>;
 
-  def : Pat<(fadd_contract (vec.vt V128:$a), (fmul_contract (vec.vt V128:$b), (vec.vt V128:$c))),
-             (!cast<Instruction>("MADD_"#vec) V128:$a, V128:$b, V128:$c)>, Requires<[HasRelaxedSIMD]>;
+  def : Pat<(fadd_contract (fmul_contract (vec.vt V128:$a), (vec.vt V128:$b)), (vec.vt V128:$c)),
+            (!cast<Instruction>("MADD_"#vec) V128:$a, V128:$b, V128:$c)>, Requires<reqs>;
+  def : Pat<(fmuladd (vec.vt V128:$a), (vec.vt V128:$b), (vec.vt V128:$c)),
+             (!cast<Instruction>("MADD_"#vec) V128:$a, V128:$b, V128:$c)>, Requires<reqs>;
 
-  def : Pat<(fsub_contract (vec.vt V128:$a), (fmul_contract (vec.vt V128:$b), (vec.vt V128:$c))),
-             (!cast<Instruction>("NMADD_"#vec) V128:$a, V128:$b, V128:$c)>, Requires<[HasRelaxedSIMD]>;
+  def : Pat<(fsub_contract (vec.vt V128:$c), (fmul_contract (vec.vt V128:$a), (vec.vt V128:$b))),
+            (!cast<Instruction>("NMADD_"#vec) V128:$a, V128:$b, V128:$c)>, Requires<reqs>;
+  def : Pat<(fmuladd (fneg (vec.vt V128:$a)), (vec.vt V128:$b), (vec.vt V128:$c)),
+             (!cast<Instruction>("NMADD_"#vec) V128:$a, V128:$b, V128:$c)>, Requires<reqs>;
 }
 
 defm "" : SIMDMADD<F32x4, 0x105, 0x106, [HasRelaxedSIMD]>;
 defm "" : SIMDMADD<F64x2, 0x107, 0x108, [HasRelaxedSIMD]>;
-defm "" : SIMDMADD<F16x8, 0x14e, 0x14f, [HasFP16]>;
+defm "" : SIMDMADD<F16x8, 0x14e, 0x14f, [HasRelaxedSIMD, HasFP16]>;
----------------
sparker-arm wrote:

> HasFp16 implies the relaxed instructions are available.

That seems like an odd assumption to make. And, unfortunately, the FP16 proposal doesn't seem to have much information in it... but if it is a fused operation, then we can do lowering differently and use the FMA. I'll make the change.

https://github.com/llvm/llvm-project/pull/161355