[llvm] [SLP] Let the caller tell canConvertToFMA which operand holds the fmul (PR #218754)

[Dmitry Sidorov via llvm-commits]

================
@@ -7,35 +7,42 @@
 ; Reduced from a real-world molecular docking kernel where independent
 ; scalar fcmps are combined with a loop-invariant scalar condition via
 ; logical select, feeding into float selects stored to consecutive memory.
+;
+; FIXME: The operand aware fma pricing keeps the one-use fmuls scalar and the
----------------
MrSidims wrote:

https://github.com/llvm/llvm-project/pull/222467 is addressing this.

https://github.com/llvm/llvm-project/pull/218754