Count the number of extract uses

Author: arsenm

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -23811,6 +23811,8 @@ SDValue DAGCombiner::reduceBuildVecToShuffle(SDNode *N) {
   // value.
   unsigned OneConstExtractIndex = ~0u;
 
+  unsigned NumExtracts = 0;
+
   for (unsigned i = 0; i != NumElems; ++i) {
     SDValue Op = N->getOperand(i);
 
@@ -23847,7 +23849,10 @@ SDValue DAGCombiner::reduceBuildVecToShuffle(SDNode *N) {
         ExtractedFromVec.getValueType().getVectorElementType())
       return SDValue();
 
-    OneConstExtractIndex = ExtractIdx->getZExtValue();
+    if (OneConstExtractIndex == ~0u)
+      OneConstExtractIndex = ExtractIdx->getZExtValue();
+
+    ++NumExtracts;
 
     // Have we seen this input vector before?
     // The vectors are expected to be tiny (usually 1 or 2 elements), so using
@@ -23878,8 +23883,11 @@ SDValue DAGCombiner::reduceBuildVecToShuffle(SDNode *N) {
     // TODO: This should be more aggressive about skipping the shuffle formation
     // (e.g., always do this for VecIn[1]->hasOneUse())
     if (TLI.isOperationLegalOrCustom(ISD::EXTRACT_VECTOR_ELT, VT) &&
-        (VecIn[1].hasOneUse() &&
-         TLI.isExtractVecEltCheap(VT, OneConstExtractIndex)))
+        TLI.isTypeLegal(VT.getVectorElementType()) &&
+        // VecIn[1].hasOneUse() &&
+        NumExtracts == 1
+        //&& TLI.isExtractVecEltCheap(VT, OneConstExtractIndex))
+    )
       return SDValue();
 
     unsigned MaxIndex = 0;

llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll

@@ -452,11 +452,11 @@ define amdgpu_kernel void @byte8_inselt(ptr addrspace(1) %out, <8 x i8> %vec, i3
 ; GCN-NEXT:    s_and_b32 s6, s4, 0x1010101
 ; GCN-NEXT:    s_andn2_b64 s[2:3], s[2:3], s[4:5]
 ; GCN-NEXT:    s_or_b64 s[2:3], s[6:7], s[2:3]
-; GCN-NEXT:    v_mov_b32_e32 v3, s1
-; GCN-NEXT:    v_mov_b32_e32 v0, s2
-; GCN-NEXT:    v_mov_b32_e32 v1, s3
-; GCN-NEXT:    v_mov_b32_e32 v2, s0
-; GCN-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN-NEXT:    v_mov_b32_e32 v0, s0
+; GCN-NEXT:    v_mov_b32_e32 v2, s2
+; GCN-NEXT:    v_mov_b32_e32 v1, s1
+; GCN-NEXT:    v_mov_b32_e32 v3, s3
+; GCN-NEXT:    flat_store_dwordx2 v[0:1], v[2:3]
 ; GCN-NEXT:    s_endpgm
 entry:
   %v = insertelement <8 x i8> %vec, i8 1, i32 %sel