[ARM] Use isFMAFasterThanFMulAndFAdd for scalars as well as MVE vectors

This adds extra scalar handling to isFMAFasterThanFMulAndFAdd, allowing
the target independent code to handle more folds in more situations (for
example if the fast math flags are present, but the global
AllowFPOpFusion option isnt). It also splits apart the HasSlowFPVMLx
into HasSlowFPVFMx, to allow VFMA and VMLA to be controlled separately
if needed.

Differential Revision: https://reviews.llvm.org/D72139

1 files changed, 9 insertions, 0 deletions

diff --git a/llvm/lib/Target/ARM/ARMSubtarget.h b/llvm/lib/Target/ARM/ARMSubtarget.h
index f582a92f656..6bdd021970e 100644
--- a/llvm/lib/Target/ARM/ARMSubtarget.h
+++ b/llvm/lib/Target/ARM/ARMSubtarget.h
@@ -203,6 +203,10 @@ protected:
   /// whether the FP VML[AS] instructions are slow (if so, don't use them).
   bool SlowFPVMLx = false;
 
+  /// SlowFPVFMx - If the VFP4 / NEON instructions are available, indicates
+  /// whether the FP VFM[AS] instructions are slow (if so, don't use them).
+  bool SlowFPVFMx = false;
+
   /// HasVMLxForwarding - If true, NEON has special multiplier accumulator
   /// forwarding to allow mul + mla being issued back to back.
   bool HasVMLxForwarding = false;
@@ -632,6 +636,11 @@ public:
 
   bool useMulOps() const { return UseMulOps; }
   bool useFPVMLx() const { return !SlowFPVMLx; }
+  bool useFPVFMx() const {
+    return !isTargetDarwin() && hasVFP4Base() && !SlowFPVFMx;
+  }
+  bool useFPVFMx16() const { return useFPVFMx() && hasFullFP16(); }
+  bool useFPVFMx64() const { return useFPVFMx() && hasFP64(); }
   bool hasVMLxForwarding() const { return HasVMLxForwarding; }
   bool isFPBrccSlow() const { return SlowFPBrcc; }
   bool hasFP64() const { return HasFP64; }