Optimize some special cases for SSE4a insertqi

Summary:
Since the upper 64 bits of the destination register are undefined when
performing this operation, we can substitute it and let the optimizer
figure out that only a copy is needed.

Also added range merging, if an instruction copies a range that can be
merged with a previous copied range.

Added test cases for both optimizations.

Reviewers: grosbach, nadav

CC: llvm-commits

Differential Revision: http://reviews.llvm.org/D3357

llvm-svn: 207055

2 files changed, 164 insertions, 0 deletions

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index f7ebb20b9fd..5d4b063d5e5 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -578,6 +578,73 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
     break;
   }
 
+  case Intrinsic::x86_sse4a_insertqi: {
+    // insertqi x, y, 64, 0 can just copy y's lower bits and leave the top
+    // ones undef
+    // TODO: eventually we should lower this intrinsic to IR
+    if (auto CIWidth = dyn_cast<ConstantInt>(II->getArgOperand(2))) {
+      if (auto CIStart = dyn_cast<ConstantInt>(II->getArgOperand(3))) {
+        if (CIWidth->equalsInt(64) && CIStart->isZero()) {
+          Value *Vec = II->getArgOperand(1);
+          Value *Undef = UndefValue::get(Vec->getType());
+          const uint32_t Mask[] = { 0, 2 };
+          return ReplaceInstUsesWith(
+              CI,
+              Builder->CreateShuffleVector(
+                  Vec, Undef, ConstantDataVector::get(
+                                  II->getContext(), ArrayRef<uint32_t>(Mask))));
+
+        } else if (auto Source =
+                       dyn_cast<IntrinsicInst>(II->getArgOperand(0))) {
+          if (Source->hasOneUse() &&
+              Source->getArgOperand(1) == II->getArgOperand(1)) {
+            // If the source of the insert has only one use and it's another
+            // insert (and they're both inserting from the same vector), try to
+            // bundle both together.
+            auto CISourceWidth =
+                dyn_cast<ConstantInt>(Source->getArgOperand(2));
+            auto CISourceStart =
+                dyn_cast<ConstantInt>(Source->getArgOperand(3));
+            if (CISourceStart && CISourceWidth) {
+              unsigned Start = CIStart->getZExtValue();
+              unsigned Width = CIWidth->getZExtValue();
+              unsigned End = Start + Width;
+              unsigned SourceStart = CISourceStart->getZExtValue();
+              unsigned SourceWidth = CISourceWidth->getZExtValue();
+              unsigned SourceEnd = SourceStart + SourceWidth;
+              unsigned NewStart, NewWidth;
+              bool ShouldReplace = false;
+              if (Start <= SourceStart && SourceStart <= End) {
+                NewStart = Start;
+                NewWidth = std::max(End, SourceEnd) - NewStart;
+                ShouldReplace = true;
+              } else if (SourceStart <= Start && Start <= SourceEnd) {
+                NewStart = SourceStart;
+                NewWidth = std::max(SourceEnd, End) - NewStart;
+                ShouldReplace = true;
+              }
+
+              if (ShouldReplace) {
+                Constant *ConstantWidth = ConstantInt::get(
+                    II->getArgOperand(2)->getType(), NewWidth, false);
+                Constant *ConstantStart = ConstantInt::get(
+                    II->getArgOperand(3)->getType(), NewStart, false);
+                Value *Args[4] = { Source->getArgOperand(0),
+                                   II->getArgOperand(1), ConstantWidth,
+                                   ConstantStart };
+                Module *M = CI.getParent()->getParent()->getParent();
+                Value *F =
+                    Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
+                return ReplaceInstUsesWith(CI, Builder->CreateCall(F, Args));
+              }
+            }
+          }
+        }
+      }
+    }
+    break;
+  }
+
   case Intrinsic::x86_avx_vpermilvar_ps:
   case Intrinsic::x86_avx_vpermilvar_ps_256:
   case Intrinsic::x86_avx_vpermilvar_pd:
diff --git a/llvm/test/Transforms/InstCombine/vec_demanded_elts.ll b/llvm/test/Transforms/InstCombine/vec_demanded_elts.ll
index f75ea032735..70370196e15 100644
--- a/llvm/test/Transforms/InstCombine/vec_demanded_elts.ll
+++ b/llvm/test/Transforms/InstCombine/vec_demanded_elts.ll
@@ -209,6 +209,103 @@ define <4 x float> @test_select(float %f, float %g) {
   ret <4 x float> %ret
 }
 
+; We should optimize these two redundant insertqi into one
+; CHECK: define <2 x i64> @testInsertTwice(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertTwice(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 32)
+; CHECK-NOT: insertqi
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 32)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 32)
+  ret <2 x i64> %2
+}
+
+; The result of this insert is the second arg, since the top 64 bits of
+; the result are undefined, and we copy the bottom 64 bits from the
+; second arg
+; CHECK: define <2 x i64> @testInsert64Bits(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsert64Bits(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: ret <2 x i64> %i
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 64, i8 0)
+  ret <2 x i64> %1
+}
+
+; Test the several types of ranges and ordering that exist for two insertqi
+; CHECK: define <2 x i64> @testInsertContainedRange(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertContainedRange(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 16)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertContainedRange_2(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertContainedRange_2(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 16)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 0)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertOverlappingRange(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertOverlappingRange(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 16)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertOverlappingRange_2(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertOverlappingRange_2(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 16)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 0)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertAdjacentRange(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertAdjacentRange(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 32, i8 0)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertAdjacentRange_2(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertAdjacentRange_2(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: %[[RES:.*]] = call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 48, i8 0)
+; CHECK: ret <2 x i64> %[[RES]]
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 32)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 32, i8 0)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertDisjointRange(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertDisjointRange(<2 x i64> %v, <2 x i64> %i) {
+; CHECK: tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
+; CHECK: tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
+  ret <2 x i64> %2
+}
+
+; CHECK: define <2 x i64> @testInsertDisjointRange_2(<2 x i64> %v, <2 x i64> %i)
+define <2 x i64> @testInsertDisjointRange_2(<2 x i64> %v, <2 x i64> %i) {
+; CHECK:  tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
+; CHECK:  tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
+  %1 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %v, <2 x i64> %i, i8 16, i8 0)
+  %2 = tail call <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64> %1, <2 x i64> %i, i8 16, i8 32)
+  ret <2 x i64> %2
+}
+
+
+; CHECK: declare <2 x i64> @llvm.x86.sse4a.insertqi
+declare <2 x i64> @llvm.x86.sse4a.insertqi(<2 x i64>, <2 x i64>, i8, i8) nounwind
+
 declare <4 x float> @llvm.x86.avx.vpermilvar.ps(<4 x float>, <4 x i32>)
 define <4 x float> @test_vpermilvar_ps(<4 x float> %v) {
 ; CHECK-LABEL: @test_vpermilvar_ps(