[ARM][ParallelDSP] Replace SExt uses

[samparker]

Summary

As loads are combined and widened, we replaced their sext users operands whereas we should have been replacing the uses of the sext. I've added a load of tests, with only a few of them originally causing assertion failures, the reset improve pattern coverage.

[efriedma]

How did we end up with this testing gap? I haven't been closely watching what tests were committed with previous patches.

Could you have caught this earlier by using AssertingVH in more places?

If you're confident this is the right patch, we can continue discussing post-commit; I'd like to unbreak the Android bot.

lib/Target/ARM/ARMParallelDSP.cpp

700

I don't understand how this makes a difference; the new SExt has to have the same operand and result type as the old one, I think, so making a new SExt followed by RAUW should be the same as changing the operand of the original. Is the old instruction in the wrong position? Or is there some sort of invalidation issue?

[samparker]

Because of the VecLd[0]/LHS typo, I had assumed that there was something wrong with the handling of the 'exchange' instructions. When I added that functionality, the tests weren't very varied so I assumed that I had missed an edge case with operand orderings. What the tests show is that there are some patterns which don't get optimised well, see the TODOs. The tests in overlapping are the kind of accesses that cause the problems and it wasn't something I had thought about. We test for sext loads with multiple users, but not in the case where multiple ones are widened.

lib/Target/ARM/ARMParallelDSP.cpp

700

The issue arises when a load is used to create more than one wide load. In the first instance, the sext would use the new load and the original load would have no users. In the second instance, when we enter this method, we'll hit the assertion because we try to look at the sext user of the original load.

[efriedma]

Okay, makes sense. LGTM

lib/Target/ARM/ARMParallelDSP.cpp

struct MulCandidate {
     MulCandidate(Instruction *I, Value *lhs, Value *rhs) :
       Root(I), LHS(lhs), RHS(rhs) { }
 
     bool HasTwoLoadInputs() const {
       return isa<LoadInst>(LHS) && isa<LoadInst>(RHS);
     }
 
     LoadInst *getBaseLoad() const {
-      return cast<LoadInst>(LHS);
+      return VecLd.front();
     }
   };
 
   /// Represent a sequence of multiply-accumulate operations with the aim to
   /// perform the multiplications in parallel.
   class Reduction {
     Instruction     *Root = nullptr;
     Value           *Acc = nullptr;
LoadInst* ARMParallelDSP::CreateWideLoad(MemInstList &Loads,
   // Make sure everything is in the correct order in the basic block.
   MoveBefore(Base->getPointerOperand(), VecPtr);
   MoveBefore(VecPtr, WideLoad);
 
   // From the wide load, create two values that equal the original two loads.
   // Loads[0] needs trunc while Loads[1] needs a lshr and trunc.
   // TODO: Support big-endian as well.
   Value *Bottom = IRB.CreateTrunc(WideLoad, Base->getType());
-  BaseSExt->setOperand(0, Bottom);
+  Value *NewBaseSExt = IRB.CreateSExt(Bottom, BaseSExt->getType());
+  BaseSExt->replaceAllUsesWith(NewBaseSExt);
 
   IntegerType *OffsetTy = cast<IntegerType>(Offset->getType());
   Value *ShiftVal = ConstantInt::get(LoadTy, OffsetTy->getBitWidth());
   Value *Top = IRB.CreateLShr(WideLoad, ShiftVal);
   Value *Trunc = IRB.CreateTrunc(Top, OffsetTy);
-  OffsetSExt->setOperand(0, Trunc);
+  Value *NewOffsetSExt = IRB.CreateSExt(Trunc, OffsetSExt->getType());
+  OffsetSExt->replaceAllUsesWith(NewOffsetSExt);
 
   WideLoads.emplace(std::make_pair(Base,
                                    make_unique<WidenedLoad>(Loads, WideLoad)));
   return WideLoad;
 }
 
 Pass *llvm::createARMParallelDSPPass() {
   return new ARMParallelDSP();
 }
 
 char ARMParallelDSP::ID = 0;
 
 INITIALIZE_PASS_BEGIN(ARMParallelDSP, "arm-parallel-dsp",
                 "Transform functions to use DSP intrinsics", false, false)
 INITIALIZE_PASS_END(ARMParallelDSP, "arm-parallel-dsp",
                 "Transform functions to use DSP intrinsics", false, false)

test/CodeGen/ARM/ParallelDSP/blocks.ll

 ; RUN: opt -arm-parallel-dsp -mtriple=armv7-a -S %s -o - | FileCheck %s
 
 ; CHECK-LABEL: single_block
 ; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
 ; CHECK: [[A:%[^ ]+]] = load i32, i32* [[CAST_A]]
 ; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
 ; CHECK: [[B:%[^ ]+]] = load i32, i32* [[CAST_B]]
-; CHECK  call i32 @llvm.arm.smlad(i32 [[A]], i32 [[B]], i32 %acc)
+; CHECK: call i32 @llvm.arm.smlad(i32 [[A]], i32 [[B]], i32 %acc)
 define i32 @single_block(i16* %a, i16* %b, i32 %acc) {
 entry:
   %ld.a.0 = load i16, i16* %a
   %sext.a.0 = sext i16 %ld.a.0 to i32
   %ld.b.0 = load i16, i16* %b
   %sext.b.0 = sext i16 %ld.b.0 to i32
   %mul.0 = mul i32 %sext.a.0, %sext.b.0
   %addr.a.1 = getelementptr i16, i16* %a, i32 1
   %addr.b.1 = getelementptr i16, i16* %b, i32 1
   %ld.a.1 = load i16, i16* %addr.a.1
   %sext.a.1 = sext i16 %ld.a.1 to i32
   %ld.b.1 = load i16, i16* %addr.b.1
   %sext.b.1 = sext i16 %ld.b.1 to i32
   %mul.1 = mul i32 %sext.a.1, %sext.b.1
   %add = add i32 %mul.0, %mul.1
   %res = add i32 %add, %acc
   ret i32 %res
 }
 
 ; CHECK-LABEL: multi_block
 ; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
 ; CHECK: [[A:%[^ ]+]] = load i32, i32* [[CAST_A]]
 ; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
 ; CHECK: [[B:%[^ ]+]] = load i32, i32* [[CAST_B]]
-; CHECK  call i32 @llvm.arm.smlad(i32 [[A]], i32 [[B]], i32 0)
+; CHECK:  call i32 @llvm.arm.smlad(i32 [[A]], i32 [[B]], i32 0)
 define i32 @multi_block(i16* %a, i16* %b, i32 %acc) {
 entry:
   %ld.a.0 = load i16, i16* %a
   %sext.a.0 = sext i16 %ld.a.0 to i32
   %ld.b.0 = load i16, i16* %b
   %sext.b.0 = sext i16 %ld.b.0 to i32
   %mul.0 = mul i32 %sext.a.0, %sext.b.0
   %addr.a.1 = getelementptr i16, i16* %a, i32 1

test/CodeGen/ARM/ParallelDSP/exchange.ll

+; RUN: opt -arm-parallel-dsp -mtriple=armv7-a -S %s -o - | FileCheck %s
+
+; CHECK-LABEL: exchange_1
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: call i32 @llvm.arm.smladx(i32 [[LD_A]], i32 [[LD_B]]
+define i32 @exchange_1(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.a.0, %sext.b.1
+  %mul.1 = mul i32 %sext.a.1, %sext.b.0
+  %add = add i32 %mul.0, %mul.1
+  %res = add i32 %add, %acc
+  ret i32 %res
+}
+
+; CHECK-LABEL: exchange_2
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: call i32 @llvm.arm.smladx(i32 [[LD_A]], i32 [[LD_B]]
+define i32 @exchange_2(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.b.1, %sext.a.0
+  %mul.1 = mul i32 %sext.b.0, %sext.a.1
+  %add = add i32 %mul.0, %mul.1
+  %res = add i32 %add, %acc
+  ret i32 %res
+}
+
+; CHECK-LABEL: exchange_3
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: call i32 @llvm.arm.smladx(i32 [[LD_B]], i32 [[LD_A]]
+define i32 @exchange_3(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.a.0, %sext.b.1
+  %mul.1 = mul i32 %sext.a.1, %sext.b.0
+  %add = add i32 %mul.1, %mul.0
+  %res = add i32 %add, %acc
+  ret i32 %res
+}
+
+; CHECK-LABEL: exchange_4
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: call i32 @llvm.arm.smladx(i32 [[LD_B]], i32 [[LD_A]]
+define i32 @exchange_4(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.b.1, %sext.a.0
+  %mul.1 = mul i32 %sext.b.0, %sext.a.1
+  %add = add i32 %mul.1, %mul.0
+  %res = add i32 %add, %acc
+  ret i32 %res
+}
+
+; CHECK-LABEL: exchange_multi_use_1
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: [[GEP:%[^ ]+]] = getelementptr i16, i16* %a, i32 2
+; CHECK: [[CAST_A_2:%[^ ]+]] = bitcast i16* [[GEP]] to i32*
+; CHECK: [[LD_A_2:%[^ ]+]] = load i32, i32* [[CAST_A_2]]
+; CHECK: [[X:%[^ ]+]] = call i32 @llvm.arm.smladx(i32 [[LD_A]], i32 [[LD_B]], i32 %acc
+; CHECK: call i32 @llvm.arm.smlad(i32 [[LD_A_2]], i32 [[LD_B]], i32 [[X]])
+define i32 @exchange_multi_use_1(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.a.0, %sext.b.1
+  %mul.1 = mul i32 %sext.a.1, %sext.b.0
+  %add = add i32 %mul.0, %mul.1
+  %addr.a.2 = getelementptr i16, i16* %a, i32 2
+  %addr.a.3 = getelementptr i16, i16* %a, i32 3
+  %ld.a.2 = load i16, i16* %addr.a.2
+  %ld.a.3 = load i16, i16* %addr.a.3
+  %sext.a.2 = sext i16 %ld.a.2 to i32
+  %sext.a.3 = sext i16 %ld.a.3 to i32
+  %mul.2 = mul i32 %sext.a.3, %sext.b.1
+  %mul.3 = mul i32 %sext.a.2, %sext.b.0
+  %add.1 = add i32 %mul.2, %mul.3
+  %add.2 = add i32 %add, %add.1
+  %res = add i32 %add.2, %acc
+  ret i32 %res
+}
+
+; CHECK-LABEL: exchange_multi_use_2
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: [[GEP:%[^ ]+]] = getelementptr i16, i16* %a, i32 2
+; CHECK: [[CAST_A_2:%[^ ]+]] = bitcast i16* [[GEP]] to i32*
+; CHECK: [[LD_A_2:%[^ ]+]] = load i32, i32* [[CAST_A_2]]
+; CHECK: [[X:%[^ ]+]] = call i32 @llvm.arm.smlad(i32 [[LD_A]], i32 [[LD_B]], i32 %acc
+; CHECK: call i32 @llvm.arm.smladx(i32 [[LD_B]], i32 [[LD_A_2]], i32 [[X]])
+define i32 @exchange_multi_use_2(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.a.0, %sext.b.0
+  %mul.1 = mul i32 %sext.a.1, %sext.b.1
+  %add = add i32 %mul.0, %mul.1
+  %addr.a.2 = getelementptr i16, i16* %a, i32 2
+  %addr.a.3 = getelementptr i16, i16* %a, i32 3
+  %ld.a.2 = load i16, i16* %addr.a.2
+  %ld.a.3 = load i16, i16* %addr.a.3
+  %sext.a.2 = sext i16 %ld.a.2 to i32
+  %sext.a.3 = sext i16 %ld.a.3 to i32
+  %mul.2 = mul i32 %sext.b.0, %sext.a.3
+  %mul.3 = mul i32 %sext.b.1, %sext.a.2
+  %add.1 = add i32 %mul.2, %mul.3
+  %add.2 = add i32 %add, %add.1
+  %res = add i32 %add.2, %acc
+  ret i32 %res
+}
+
+; TODO: Why aren't two intrinsics generated?
+; CHECK-LABEL: exchange_multi_use_3
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: [[GEP:%[^ ]+]] = getelementptr i16, i16* %a, i32 2
+; CHECK: [[CAST_A_2:%[^ ]+]] = bitcast i16* [[GEP]] to i32*
+; CHECK: [[LD_A_2:%[^ ]+]] = load i32, i32* [[CAST_A_2]]
+; CHECK-NOT: call i32 @llvm.arm.smlad
+; CHECK: [[X:%[^ ]+]] = call i32 @llvm.arm.smladx(i32 [[LD_B]], i32 [[LD_A_2]], i32 0
+define i32 @exchange_multi_use_3(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %addr.a.2 = getelementptr i16, i16* %a, i32 2
+  %addr.a.3 = getelementptr i16, i16* %a, i32 3
+  %ld.a.2 = load i16, i16* %addr.a.2
+  %ld.a.3 = load i16, i16* %addr.a.3
+  %sext.a.2 = sext i16 %ld.a.2 to i32
+  %sext.a.3 = sext i16 %ld.a.3 to i32
+  %mul.2 = mul i32 %sext.b.0, %sext.a.3
+  %mul.3 = mul i32 %sext.b.1, %sext.a.2
+  %mul.0 = mul i32 %sext.a.0, %sext.b.0
+  %mul.1 = mul i32 %sext.a.1, %sext.b.1
+  %add = add i32 %mul.0, %mul.1
+  %add.1 = add i32 %mul.2, %mul.3
+  %sub = sub i32 %add, %add.1
+  %res = add i32 %acc, %sub
+  ret i32 %res
+}
+
+; TODO: Why isn't smladx generated too?
+; CHECK-LABEL: exchange_multi_use_4
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: [[X:%[^ ]+]] = call i32 @llvm.arm.smlad(i32 [[LD_A]], i32 [[LD_B]], i32 0
+; CHECK-NOT: call i32 @llvm.arm.smlad
+define i32 @exchange_multi_use_4(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %addr.a.2 = getelementptr i16, i16* %a, i32 2
+  %addr.a.3 = getelementptr i16, i16* %a, i32 3
+  %ld.a.2 = load i16, i16* %addr.a.2
+  %ld.a.3 = load i16, i16* %addr.a.3
+  %sext.a.2 = sext i16 %ld.a.2 to i32
+  %sext.a.3 = sext i16 %ld.a.3 to i32
+  %mul.2 = mul i32 %sext.b.0, %sext.a.3
+  %mul.3 = mul i32 %sext.b.1, %sext.a.2
+  %mul.0 = mul i32 %sext.a.0, %sext.b.0
+  %mul.1 = mul i32 %sext.a.1, %sext.b.1
+  %add.1 = add i32 %mul.2, %mul.3
+  %add = add i32 %mul.0, %mul.1
+  %sub = sub i32 %add, %add.1
+  %res = add i32 %acc, %sub
+  ret i32 %res
+}
+
+; CHECK-LABEL: exchange_swap
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: call i32 @llvm.arm.smladx(i32 [[LD_B]], i32 [[LD_A]]
+define i32 @exchange_swap(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.a.1, %sext.b.0
+  %mul.1 = mul i32 %sext.a.0, %sext.b.1
+  %add = add i32 %mul.0, %mul.1
+  %res = add i32 %add, %acc
+  ret i32 %res
+}
+
+; CHECK-LABEL: exchange_swap_2
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: call i32 @llvm.arm.smladx(i32 [[LD_A]], i32 [[LD_B]]
+define i32 @exchange_swap_2(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.a.1, %sext.b.0
+  %mul.1 = mul i32 %sext.a.0, %sext.b.1
+  %add = add i32 %mul.1, %mul.0
+  %res = add i32 %add, %acc
+  ret i32 %res
+}
+
+; CHECK-LABEL: exchange_swap_3
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: call i32 @llvm.arm.smladx(i32 [[LD_A]], i32 [[LD_B]]
+define i32 @exchange_swap_3(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.b.0, %sext.a.1
+  %mul.1 = mul i32 %sext.b.1, %sext.a.0
+  %add = add i32 %mul.1, %mul.0
+  %res = add i32 %add, %acc
+  ret i32 %res
+}

test/CodeGen/ARM/ParallelDSP/overlapping.ll

+; RUN: opt -arm-parallel-dsp -mtriple=armv7-a -S %s -o - | FileCheck %s
+
+; CHECK-LABEL: overlap_1
+; CHECK: [[GEP_A:%[^ ]+]] = getelementptr i16, i16* %a, i32 1
+; CHECK: [[GEP_B:%[^ ]+]] = getelementptr i16, i16* %b, i32 1
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* [[GEP_A]] to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* [[GEP_B]] to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: call i32 @llvm.arm.smlad(i32 [[LD_A]], i32 [[LD_B]], i32 %acc
+define i32 @overlap_1(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.a.0, %sext.b.0
+  %mul.1 = mul i32 %sext.a.1, %sext.b.1
+  %addr.a.2 = getelementptr i16, i16* %a, i32 2
+  %addr.b.2 = getelementptr i16, i16* %b, i32 2
+  %ld.a.2 = load i16, i16* %addr.a.2
+  %ld.b.2 = load i16, i16* %addr.b.2
+  %sext.a.2 = sext i16 %ld.a.2 to i32
+  %sext.b.2 = sext i16 %ld.b.2 to i32
+  %mul.2 = mul i32 %sext.a.2, %sext.b.2
+  %add = add i32 %mul.0, %mul.1
+  %add.1 = add i32 %mul.1, %mul.2
+  %add.2 = add i32 %add.1, %add
+  %res = add i32 %add.2, %acc
+  ret i32 %res
+}
+
+; CHECK-LABEL: overlap_2
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: call i32 @llvm.arm.smlad(i32 [[LD_A]], i32 [[LD_B]], i32 %acc
+define i32 @overlap_2(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.a.0, %sext.b.0
+  %mul.1 = mul i32 %sext.a.1, %sext.b.1
+  %addr.a.2 = getelementptr i16, i16* %a, i32 2
+  %addr.b.2 = getelementptr i16, i16* %b, i32 2
+  %ld.a.2 = load i16, i16* %addr.a.2
+  %ld.b.2 = load i16, i16* %addr.b.2
+  %sext.a.2 = sext i16 %ld.a.2 to i32
+  %sext.b.2 = sext i16 %ld.b.2 to i32
+  %mul.2 = mul i32 %sext.b.2, %sext.a.2
+  %add = add i32 %mul.0, %mul.1
+  %add.1 = add i32 %mul.1, %mul.2
+  %add.2 = add i32 %add, %add.1
+  %res = add i32 %add.2, %acc
+  ret i32 %res
+}
+
+; CHECK-LABEL: overlap_3
+; CHECK: [[GEP_B:%[^ ]+]] = getelementptr i16, i16* %b, i32 1
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: [[CAST_B_1:%[^ ]+]] = bitcast i16* [[GEP_B]] to i32*
+; CHECK: [[LD_B_1:%[^ ]+]] = load i32, i32* [[CAST_B_1]]
+; CHECK: [[GEP_A:%[^ ]+]] = getelementptr i16, i16* %a, i32 2
+; CHECK: [[CAST_A_2:%[^ ]+]] = bitcast i16* [[GEP_A]] to i32*
+; CHECK: [[LD_A_2:%[^ ]+]] = load i32, i32* [[CAST_A_2]]
+; CHECK: [[SMLAD:%[^ ]+]] = call i32 @llvm.arm.smlad(i32 [[LD_A_2]], i32 [[LD_B_1]], i32 %acc)
+; CHECK: call i32 @llvm.arm.smlad(i32 [[LD_A]], i32 [[LD_B]], i32 [[SMLAD]])
+define i32 @overlap_3(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.a.0, %sext.b.0
+  %mul.1 = mul i32 %sext.a.1, %sext.b.1
+  %addr.a.2 = getelementptr i16, i16* %a, i32 2
+  %addr.b.2 = getelementptr i16, i16* %b, i32 2
+  %addr.a.3 = getelementptr i16, i16* %a, i32 3
+  %ld.a.2 = load i16, i16* %addr.a.2
+  %ld.b.2 = load i16, i16* %addr.b.2
+  %ld.a.3 = load i16, i16* %addr.a.3
+  %sext.a.2 = sext i16 %ld.a.2 to i32
+  %sext.b.2 = sext i16 %ld.b.2 to i32
+  %sext.a.3 = sext i16 %ld.a.3 to i32
+  %mul.2 = mul i32 %sext.a.2, %sext.b.1
+  %mul.3 = mul i32 %sext.a.3, %sext.b.2
+  %add = add i32 %mul.0, %mul.1
+  %add.1 = add i32 %mul.2, %mul.3
+  %add.2 = add i32 %add.1, %add
+  %res = add i32 %add.2, %acc
+  ret i32 %res
+}
+
+; CHECK-LABEL: overlap_4
+; CHECK: [[GEP_B:%[^ ]+]] = getelementptr i16, i16* %b, i32 1
+; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
+; CHECK: [[LD_A:%[^ ]+]] = load i32, i32* [[CAST_A]]
+; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
+; CHECK: [[LD_B:%[^ ]+]] = load i32, i32* [[CAST_B]]
+; CHECK: [[CAST_B_1:%[^ ]+]] = bitcast i16* [[GEP_B]] to i32*
+; CHECK: [[LD_B_1:%[^ ]+]] = load i32, i32* [[CAST_B_1]]
+; CHECK: [[GEP_A:%[^ ]+]] = getelementptr i16, i16* %a, i32 2
+; CHECK: [[CAST_A_2:%[^ ]+]] = bitcast i16* [[GEP_A]] to i32*
+; CHECK: [[LD_A_2:%[^ ]+]] = load i32, i32* [[CAST_A_2]]
+; CHECK: [[SMLAD:%[^ ]+]] = call i32 @llvm.arm.smladx(i32 [[LD_A_2]], i32 [[LD_B_1]], i32 %acc)
+; CHECK: call i32 @llvm.arm.smlad(i32 [[LD_A]], i32 [[LD_B]], i32 [[SMLAD]])
+define i32 @overlap_4(i16* %a, i16* %b, i32 %acc) {
+entry:
+  %addr.a.1 = getelementptr i16, i16* %a, i32 1
+  %addr.b.1 = getelementptr i16, i16* %b, i32 1
+  %ld.a.0 = load i16, i16* %a
+  %sext.a.0 = sext i16 %ld.a.0 to i32
+  %ld.b.0 = load i16, i16* %b
+  %ld.a.1 = load i16, i16* %addr.a.1
+  %ld.b.1 = load i16, i16* %addr.b.1
+  %sext.a.1 = sext i16 %ld.a.1 to i32
+  %sext.b.1 = sext i16 %ld.b.1 to i32
+  %sext.b.0 = sext i16 %ld.b.0 to i32
+  %mul.0 = mul i32 %sext.a.0, %sext.b.0
+  %mul.1 = mul i32 %sext.a.1, %sext.b.1
+  %addr.a.2 = getelementptr i16, i16* %a, i32 2
+  %addr.b.2 = getelementptr i16, i16* %b, i32 2
+  %addr.a.3 = getelementptr i16, i16* %a, i32 3
+  %ld.a.2 = load i16, i16* %addr.a.2
+  %ld.b.2 = load i16, i16* %addr.b.2
+  %ld.a.3 = load i16, i16* %addr.a.3
+  %sext.a.2 = sext i16 %ld.a.2 to i32
+  %sext.b.2 = sext i16 %ld.b.2 to i32
+  %sext.a.3 = sext i16 %ld.a.3 to i32
+  %mul.2 = mul i32 %sext.b.2, %sext.a.2
+  %mul.3 = mul i32 %sext.b.1, %sext.a.3
+  %add = add i32 %mul.0, %mul.1
+  %add.1 = add i32 %mul.2, %mul.3
+  %add.2 = add i32 %add.1, %add
+  %res = add i32 %add.2, %acc
+  ret i32 %res
+}