Add support for llvm.arm.neon.vmull* intrinsics to InstCombine. Fixes

<rdar://problem/11291436>.

This is a second attempt at a fix for this, the first was r155468. Thanks
to Chandler, Bob and others for the feedback that helped me improve this.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@155866 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 119 insertions, 0 deletions

diff --git a/lib/Transforms/InstCombine/InstCombineCalls.cpp b/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 77e4727..a3dc77d 100644
--- a/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -694,6 +694,57 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
     break;
   }
 
+  case Intrinsic::arm_neon_vmulls:
+  case Intrinsic::arm_neon_vmullu: {
+    Value *Arg0 = II->getArgOperand(0);
+    Value *Arg1 = II->getArgOperand(1);
+
+    // Handle mul by zero first:
+    if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) {
+      return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType()));
+    }
+
+    // Check for constant LHS & RHS - in this case we just simplify.
+    bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu);
+    VectorType *NewVT = cast<VectorType>(II->getType());
+    unsigned NewWidth = NewVT->getElementType()->getIntegerBitWidth();
+    if (ConstantDataVector *CV0 = dyn_cast<ConstantDataVector>(Arg0)) {
+      if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) {
+        VectorType* VT = cast<VectorType>(CV0->getType());
+        SmallVector<Constant*, 4> NewElems;
+        for (unsigned i = 0; i < VT->getNumElements(); ++i) {
+          APInt CV0E =
+            (cast<ConstantInt>(CV0->getAggregateElement(i)))->getValue();
+          CV0E = Zext ? CV0E.zext(NewWidth) : CV0E.sext(NewWidth);
+          APInt CV1E =
+            (cast<ConstantInt>(CV1->getAggregateElement(i)))->getValue();
+          CV1E = Zext ? CV1E.zext(NewWidth) : CV1E.sext(NewWidth);
+          NewElems.push_back(
+            ConstantInt::get(NewVT->getElementType(), CV0E * CV1E));
+        }
+        return ReplaceInstUsesWith(CI, ConstantVector::get(NewElems));
+      }
+
+      // Couldn't simplify - cannonicalize constant to the RHS.
+      std::swap(Arg0, Arg1);
+    }
+
+    // Handle mul by one:
+    if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) {
+      if (ConstantInt *Splat =
+            dyn_cast_or_null<ConstantInt>(CV1->getSplatValue())) {
+        if (Splat->isOne()) {
+          if (Zext)
+            return CastInst::CreateZExtOrBitCast(Arg0, II->getType());
+          // else    
+          return CastInst::CreateSExtOrBitCast(Arg0, II->getType());
+        }
+      }
+    }
+
+    break;
+  }
+
   case Intrinsic::stackrestore: {
     // If the save is right next to the restore, remove the restore.  This can
     // happen when variable allocas are DCE'd.
diff --git a/test/Transforms/InstCombine/2012-04-23-Neon-Intrinsics.ll b/test/Transforms/InstCombine/2012-04-23-Neon-Intrinsics.ll
new file mode 100644
index 0000000..0907c49
--- /dev/null
+++ b/test/Transforms/InstCombine/2012-04-23-Neon-Intrinsics.ll
@@ -0,0 +1,68 @@
+target datalayout = "e-p:32:32:32-i1:8:32-i8:8:32-i16:16:32-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:32:64-v128:32:128-a0:0:32-n32-S32"
+target triple = "thumbv7-apple-ios0"
+
+; RUN: opt -S -instcombine < %s | FileCheck %s
+
+define <4 x i32> @mulByZero(<4 x i16> %x) nounwind readnone ssp {
+entry:
+  %a = tail call <4 x i32> @llvm.arm.neon.vmulls.v4i32(<4 x i16> %x, <4 x i16> zeroinitializer) nounwind
+  ret <4 x i32> %a
+; CHECK: entry:
+; CHECK-NEXT: ret <4 x i32> zeroinitializer
+}
+
+define <4 x i32> @mulByOne(<4 x i16> %x) nounwind readnone ssp {
+entry:
+  %a = tail call <4 x i32> @llvm.arm.neon.vmulls.v4i32(<4 x i16> %x, <4 x i16> <i16 1, i16 1, i16 1, i16 1>) nounwind
+  ret <4 x i32> %a
+; CHECK: entry:
+; CHECK-NEXT: %a = sext <4 x i16> %x to <4 x i32>
+; CHECK-NEXT: ret <4 x i32> %a
+}
+
+define <4 x i32> @constantMul() nounwind readnone ssp {
+entry:
+  %a = tail call <4 x i32> @llvm.arm.neon.vmulls.v4i32(<4 x i16> <i16 3, i16 3, i16 3, i16 3>, <4 x i16> <i16 2, i16 2, i16 2, i16 2>) nounwind
+  ret <4 x i32> %a
+; CHECK: entry:
+; CHECK-NEXT: ret <4 x i32> <i32 6, i32 6, i32 6, i32 6>
+}
+
+define <4 x i32> @constantMulS() nounwind readnone ssp {
+entry:
+  %b = tail call <4 x i32> @llvm.arm.neon.vmulls.v4i32(<4 x i16> <i16 -1, i16 -1, i16 -1, i16 -1>, <4 x i16> <i16 1, i16 1, i16 1, i16 1>) nounwind
+  ret <4 x i32> %b
+; CHECK: entry:
+; CHECK-NEXT: ret <4 x i32> <i32 -1, i32 -1, i32 -1, i32 -1>
+}
+
+define <4 x i32> @constantMulU() nounwind readnone ssp {
+entry:
+  %b = tail call <4 x i32> @llvm.arm.neon.vmullu.v4i32(<4 x i16> <i16 -1, i16 -1, i16 -1, i16 -1>, <4 x i16> <i16 1, i16 1, i16 1, i16 1>) nounwind
+  ret <4 x i32> %b
+; CHECK: entry:
+; CHECK-NEXT: ret <4 x i32> <i32 65535, i32 65535, i32 65535, i32 65535>
+}
+
+define <4 x i32> @complex1(<4 x i16> %x) nounwind readnone ssp {
+entry:
+  %a = tail call <4 x i32> @llvm.arm.neon.vmulls.v4i32(<4 x i16> <i16 2, i16 2, i16 2, i16 2>, <4 x i16> %x) nounwind
+  %b = add <4 x i32> zeroinitializer, %a
+  ret <4 x i32> %b
+; CHECK: entry:
+; CHECK-NEXT: %a = tail call <4 x i32> @llvm.arm.neon.vmulls.v4i32(<4 x i16> <i16 2, i16 2, i16 2, i16 2>, <4 x i16> %x) nounwind
+; CHECK-NEXT: ret <4 x i32> %a
+}
+
+define <4 x i32> @complex2(<4 x i32> %x) nounwind readnone ssp {
+entry:
+  %a = tail call <4 x i32> @llvm.arm.neon.vmulls.v4i32(<4 x i16> <i16 3, i16 3, i16 3, i16 3>, <4 x i16> <i16 2, i16 2, i16 2, i16 2>) nounwind
+  %b = add <4 x i32> %x, %a
+  ret <4 x i32> %b  
+; CHECK: entry:
+; CHECK-NEXT: %b = add <4 x i32> %x, <i32 6, i32 6, i32 6, i32 6>
+; CHECK-NEXT: ret <4 x i32> %b
+}
+
+declare <4 x i32> @llvm.arm.neon.vmulls.v4i32(<4 x i16>, <4 x i16>) nounwind readnone
+declare <4 x i32> @llvm.arm.neon.vmullu.v4i32(<4 x i16>, <4 x i16>) nounwind readnone