Distribute (A + B) * C to (A * C) + (B * C) to make use of NEON multiplier

accumulator forwarding:
vadd d3, d0, d1
vmul d3, d3, d2
=>
vmul d3, d0, d2
vmla d3, d1, d2


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@128665 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 80 insertions, 4 deletions

diff --git a/lib/Target/ARM/ARM.td b/lib/Target/ARM/ARM.td
index bf4315f..e690e18 100644
--- a/lib/Target/ARM/ARM.td
+++ b/lib/Target/ARM/ARM.td
@@ -51,6 +51,12 @@ def FeatureVFPOnlySP : SubtargetFeature<"fp-only-sp", "FPOnlySP", "true",
 // to just not use them.
 def FeatureHasSlowFPVMLx : SubtargetFeature<"slowfpvmlx", "SlowFPVMLx", "true",
                                          "Disable VFP / NEON MAC instructions">;
+
+// Cortex-A8 / A9 Advanced SIMD has multiplier accumulator forwarding.
+def FeatureVMLxForwarding : SubtargetFeature<"vmlx-forwarding",
+                                       "HasVMLxForwarding", "true",
+                                       "Has multiplier accumulator forwarding">;
+
 // Some processors benefit from using NEON instructions for scalar
 // single-precision FP operations.
 def FeatureNEONForFP : SubtargetFeature<"neonfp", "UseNEONForSinglePrecisionFP",
@@ -100,11 +106,12 @@ def ProcOthers  : SubtargetFeature<"others", "ARMProcFamily", "Others",
 def ProcA8      : SubtargetFeature<"a8", "ARMProcFamily", "CortexA8",
                                    "Cortex-A8 ARM processors",
                                    [FeatureSlowFPBrcc, FeatureNEONForFP,
-                                    FeatureHasSlowFPVMLx, FeatureT2XtPk]>;
+                                    FeatureHasSlowFPVMLx, FeatureVMLxForwarding,
+                                    FeatureT2XtPk]>;
 def ProcA9      : SubtargetFeature<"a9", "ARMProcFamily", "CortexA9",
                                    "Cortex-A9 ARM processors",
-                                   [FeatureHasSlowFPVMLx, FeatureT2XtPk,
-                                    FeatureFP16]>;
+                                   [FeatureHasSlowFPVMLx, FeatureVMLxForwarding,
+                                    FeatureT2XtPk, FeatureFP16]>;
 
 class ProcNoItin<string Name, list<SubtargetFeature> Features>
  : Processor<Name, GenericItineraries, Features>;
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 16b110f..5838181 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -5224,6 +5224,42 @@ static SDValue PerformSUBCombine(SDNode *N,
   return SDValue();
 }
 
+/// PerformVMULCombine
+/// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
+/// special multiplier accumulator forwarding.
+///   vmul d3, d0, d2
+///   vmla d3, d1, d2
+/// is faster than
+///   vadd d3, d0, d1
+///   vmul d3, d3, d2
+static SDValue PerformVMULCombine(SDNode *N,
+                                  TargetLowering::DAGCombinerInfo &DCI,
+                                  const ARMSubtarget *Subtarget) {
+  if (!Subtarget->hasVMLxForwarding())
+    return SDValue();
+
+  SelectionDAG &DAG = DCI.DAG;
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+  unsigned Opcode = N0.getOpcode();
+  if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
+      Opcode != ISD::FADD && Opcode != ISD::FSUB) {
+    Opcode = N0.getOpcode();
+    if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
+        Opcode != ISD::FADD && Opcode != ISD::FSUB)
+      return SDValue();
+    std::swap(N0, N1);
+  }
+
+  EVT VT = N->getValueType(0);
+  DebugLoc DL = N->getDebugLoc();
+  SDValue N00 = N0->getOperand(0);
+  SDValue N01 = N0->getOperand(1);
+  return DAG.getNode(Opcode, DL, VT,
+                     DAG.getNode(ISD::MUL, DL, VT, N00, N1),
+                     DAG.getNode(ISD::MUL, DL, VT, N01, N1));
+}
+
 static SDValue PerformMULCombine(SDNode *N,
                                  TargetLowering::DAGCombinerInfo &DCI,
                                  const ARMSubtarget *Subtarget) {
@@ -5236,6 +5272,8 @@ static SDValue PerformMULCombine(SDNode *N,
     return SDValue();
 
   EVT VT = N->getValueType(0);
+  if (VT.is64BitVector() || VT.is128BitVector())
+    return PerformVMULCombine(N, DCI, Subtarget);
   if (VT != MVT::i32)
     return SDValue();
 
diff --git a/lib/Target/ARM/ARMSubtarget.h b/lib/Target/ARM/ARMSubtarget.h
index 76c1c3f..e024182 100644
--- a/lib/Target/ARM/ARMSubtarget.h
+++ b/lib/Target/ARM/ARMSubtarget.h
@@ -61,6 +61,10 @@ protected:
   /// whether the FP VML[AS] instructions are slow (if so, don't use them).
   bool SlowFPVMLx;
 
+  /// HasVMLxForwarding - If true, NEON has special multiplier accumulator
+  /// forwarding to allow mul + mla being issued back to back.
+  bool HasVMLxForwarding;
+
   /// SlowFPBrcc - True if floating point compare + branch is slow.
   bool SlowFPBrcc;
 
@@ -182,6 +186,7 @@ protected:
   bool hasT2ExtractPack() const { return HasT2ExtractPack; }
   bool hasDataBarrier() const { return HasDataBarrier; }
   bool useFPVMLx() const { return !SlowFPVMLx; }
+  bool hasVMLxForwarding() const { return HasVMLxForwarding; }
   bool isFPBrccSlow() const { return SlowFPBrcc; }
   bool isFPOnlySP() const { return FPOnlySP; }
   bool prefers32BitThumb() const { return Pref32BitThumb; }
diff --git a/test/CodeGen/ARM/vmul.ll b/test/CodeGen/ARM/vmul.ll
index 80ba9be..1fd6581 100644
--- a/test/CodeGen/ARM/vmul.ll
+++ b/test/CodeGen/ARM/vmul.ll
@@ -1,4 +1,4 @@
-; RUN: llc < %s -march=arm -mattr=+neon | FileCheck %s
+; RUN: llc < %s -march=arm -mcpu=cortex-a8 | FileCheck %s
 
 define <8 x i8> @vmuli8(<8 x i8>* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vmuli8:
@@ -466,3 +466,29 @@ entry:
 declare <16 x i8> @llvm.arm.neon.vld1.v16i8(i8*, i32) nounwind readonly
 
 declare void @llvm.arm.neon.vst1.v8i16(i8*, <8 x i16>, i32) nounwind
+
+; Take advantage of the Cortex-A8 multiplier accumulator forward.
+
+%struct.uint8x8_t = type { <8 x i8> }
+
+define void @distribue2(%struct.uint8x8_t* nocapture %dst, i8* %src, i32 %mul) nounwind {
+entry:
+; CHECK: distribue2
+; CHECK-NOT: vadd.i8
+; CHECK: vmul.i8
+; CHECK: vmla.i8
+  %0 = trunc i32 %mul to i8
+  %1 = insertelement <8 x i8> undef, i8 %0, i32 0
+  %2 = shufflevector <8 x i8> %1, <8 x i8> undef, <8 x i32> zeroinitializer
+  %3 = tail call <16 x i8> @llvm.arm.neon.vld1.v16i8(i8* %src, i32 1)
+  %4 = bitcast <16 x i8> %3 to <2 x double>
+  %5 = extractelement <2 x double> %4, i32 1
+  %6 = bitcast double %5 to <8 x i8>
+  %7 = extractelement <2 x double> %4, i32 0
+  %8 = bitcast double %7 to <8 x i8>
+  %9 = add <8 x i8> %6, %8
+  %10 = mul <8 x i8> %9, %2
+  %11 = getelementptr inbounds %struct.uint8x8_t* %dst, i32 0, i32 0
+  store <8 x i8> %10, <8 x i8>* %11, align 8
+  ret void
+}