Merge branch 'main' into users/chapuni/cov/single/unifyusers/chapuni/cov/single/unify

1 files changed, 10 insertions, 0 deletions

diff --git a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
index 7a7554c..c19bcfc 100644
--- a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -1650,6 +1650,13 @@ InstructionCost X86TTIImpl::getShuffleCost(
         return MatchingTypes ? TTI::TCC_Free : SubLT.first;
     }
 
+    // Attempt to match MOVSS (Idx == 0) or INSERTPS pattern. This will have
+    // been matched by improveShuffleKindFromMask as a SK_InsertSubvector of
+    // v1f32 (legalised to f32) into a v4f32.
+    if (LT.first == 1 && LT.second == MVT::v4f32 && SubLT.first == 1 &&
+        SubLT.second == MVT::f32 && (Index == 0 || ST->hasSSE41()))
+      return 1;
+
     // If the insertion isn't aligned, treat it like a 2-op shuffle.
     Kind = TTI::SK_PermuteTwoSrc;
   }
@@ -4797,9 +4804,12 @@ InstructionCost X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val,
     MVT MScalarTy = LT.second.getScalarType();
     auto IsCheapPInsrPExtrInsertPS = [&]() {
       // Assume pinsr/pextr XMM <-> GPR is relatively cheap on all targets.
+      // Inserting f32 into index0 is just movss.
       // Also, assume insertps is relatively cheap on all >= SSE41 targets.
       return (MScalarTy == MVT::i16 && ST->hasSSE2()) ||
              (MScalarTy.isInteger() && ST->hasSSE41()) ||
+             (MScalarTy == MVT::f32 && ST->hasSSE1() && Index == 0 &&
+              Opcode == Instruction::InsertElement) ||
              (MScalarTy == MVT::f32 && ST->hasSSE41() &&
               Opcode == Instruction::InsertElement);
     };