Allow generation of vmla.f32 instructions when targeting Cortex-A15. The patch also adds the VFP4 feature to Cortex-A15 and fixes the DontUseFusedMAC predicate so that we can still generate vmla.f32 instructions on non-darwin targets with VFP4.

llvm-svn: 187349

5 files changed, 31 insertions, 5 deletions

diff --git a/llvm/lib/Target/ARM/ARM.td b/llvm/lib/Target/ARM/ARM.td
index 46928dcb2ee..e5da3a54390 100644
--- a/llvm/lib/Target/ARM/ARM.td
+++ b/llvm/lib/Target/ARM/ARM.td
@@ -179,7 +179,7 @@ def ProcSwift   : SubtargetFeature<"swift", "ARMProcFamily", "Swift",
 // FIXME: It has not been determined if A15 has these features.
 def ProcA15      : SubtargetFeature<"a15", "ARMProcFamily", "CortexA15",
                                    "Cortex-A15 ARM processors",
-                                   [FeatureT2XtPk, FeatureFP16,
+                                   [FeatureT2XtPk, FeatureFP16, FeatureVFP4,
                                     FeatureAvoidPartialCPSR,
                                     FeatureTrustZone]>;
 def ProcR5      : SubtargetFeature<"r5", "ARMProcFamily", "CortexR5",
diff --git a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 31ce38e503c..4ca3af6b3ea 100644
--- a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -422,7 +422,7 @@ bool ARMDAGToDAGISel::hasNoVMLxHazardUse(SDNode *N) const {
   if (!CheckVMLxHazard)
     return true;
 
-  if (!Subtarget->isCortexA8() && !Subtarget->isLikeA9() &&
+  if (!Subtarget->isCortexA8() && !Subtarget->isCortexA9() &&
       !Subtarget->isSwift())
     return true;
 
diff --git a/llvm/lib/Target/ARM/ARMInstrInfo.td b/llvm/lib/Target/ARM/ARMInstrInfo.td
index f543e5d7ee8..c2434023f8f 100644
--- a/llvm/lib/Target/ARM/ARMInstrInfo.td
+++ b/llvm/lib/Target/ARM/ARMInstrInfo.td
@@ -262,7 +262,9 @@ def UseMulOps        : Predicate<"Subtarget->useMulOps()">;
 def UseFusedMAC      : Predicate<"(TM.Options.AllowFPOpFusion =="
                                  " FPOpFusion::Fast) && "
                                  "!Subtarget->isTargetDarwin()">;
-def DontUseFusedMAC  : Predicate<"!Subtarget->hasVFP4() || "
+def DontUseFusedMAC  : Predicate<"!(TM.Options.AllowFPOpFusion =="
+                                 " FPOpFusion::Fast &&"
+                                 " Subtarget->hasVFP4()) || "
                                  "Subtarget->isTargetDarwin()">;
 
 // VGETLNi32 is microcoded on Swift - prefer VMOV.
diff --git a/llvm/lib/Target/ARM/ARMTargetMachine.cpp b/llvm/lib/Target/ARM/ARMTargetMachine.cpp
index 354a7797dea..1ba78e4a984 100644
--- a/llvm/lib/Target/ARM/ARMTargetMachine.cpp
+++ b/llvm/lib/Target/ARM/ARMTargetMachine.cpp
@@ -169,7 +169,7 @@ bool ARMPassConfig::addPreRegAlloc() {
   // FIXME: temporarily disabling load / store optimization pass for Thumb1.
   if (getOptLevel() != CodeGenOpt::None && !getARMSubtarget().isThumb1Only())
     addPass(createARMLoadStoreOptimizationPass(true));
-  if (getOptLevel() != CodeGenOpt::None && getARMSubtarget().isLikeA9())
+  if (getOptLevel() != CodeGenOpt::None && getARMSubtarget().isCortexA9())
     addPass(createMLxExpansionPass());
   // Since the A15SDOptimizer pass can insert VDUP instructions, it can only be
   // enabled when NEON is available.
diff --git a/llvm/test/CodeGen/ARM/a15-mla.ll b/llvm/test/CodeGen/ARM/a15-mla.ll
index 25f6de4762d..b233cc27c4b 100644
--- a/llvm/test/CodeGen/ARM/a15-mla.ll
+++ b/llvm/test/CodeGen/ARM/a15-mla.ll
@@ -1,7 +1,7 @@
 ; RUN: llc < %s  -march=arm -float-abi=hard -mcpu=cortex-a15 -mattr=+neon,+neonfp | FileCheck %s
 
 ; This test checks that the VMLxForwarting feature is disabled for A15.
-; CHECK: fun_a
+; CHECK: fun_a:
 define <4 x i32> @fun_a(<4 x i32> %x, <4 x i32> %y) nounwind{
   %1 = add <4 x i32> %x, %y
 ; CHECK-NOT: vmul
@@ -10,3 +10,27 @@ define <4 x i32> @fun_a(<4 x i32> %x, <4 x i32> %y) nounwind{
   %3 = add <4 x i32> %y, %2
   ret <4 x i32> %3
 }
+
+; This tests checks that VMLA FP patterns can be matched in instruction selection when targeting
+; Cortex-A15.
+; CHECK: fun_b:
+define <4 x float> @fun_b(<4 x float> %x, <4 x float> %y, <4 x float> %z) nounwind{
+; CHECK: vmla.f32
+  %t = fmul <4 x float> %x, %y
+  %r = fadd <4 x float> %t, %z
+  ret <4 x float> %r
+}
+
+; This tests checks that FP VMLA instructions are not expanded into separate multiply/addition
+; operations when targeting Cortex-A15.
+; CHECK: fun_c:
+define <4 x float> @fun_c(<4 x float> %x, <4 x float> %y, <4 x float> %z, <4 x float> %u, <4 x float> %v) nounwind{
+; CHECK: vmla.f32
+  %t1 = fmul <4 x float> %x, %y
+  %r1 = fadd <4 x float> %t1, %z
+; CHECK: vmla.f32
+  %t2 = fmul <4 x float> %u, %v
+  %r2 = fadd <4 x float> %t2, %r1
+  ret <4 x float> %r2
+}
+