[DAG] Don't allow implicit truncation in extract_element(bitcast(scalar_to_vector(X))) -> trunc(srl(X,C)) fold

Limits #117900 to only fold when scalar_to_vector doesn't perform implicit truncation, as the scaled shift calculation doesn't currently account for this - this can be addressed in a future update.

Fixes #121306

Author: RKSimon

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -23088,8 +23088,11 @@ SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
     if (ExtractIndex == BCTruncElt && BCSrc.getValueType().isScalarInteger())
       return DAG.getAnyExtOrTrunc(BCSrc, DL, ScalarVT);
 
+    // TODO: Add support for SCALAR_TO_VECTOR implicit truncation.
     if (LegalTypes && BCSrc.getValueType().isInteger() &&
-        BCSrc.getOpcode() == ISD::SCALAR_TO_VECTOR) {
+        BCSrc.getOpcode() == ISD::SCALAR_TO_VECTOR &&
+        BCSrc.getScalarValueSizeInBits() ==
+            BCSrc.getOperand(0).getScalarValueSizeInBits()) {
       // ext_elt (bitcast (scalar_to_vec i64 X to v2i64) to v4i32), TruncElt -->
       // trunc i64 X to i32
       SDValue X = BCSrc.getOperand(0);

llvm/test/CodeGen/PowerPC/scalar_vector_test_5.ll

@@ -11,24 +11,35 @@
 define i8 @scalar_to_vector_half(ptr nocapture readonly %ad) {
 ; P9LE-LABEL: scalar_to_vector_half:
 ; P9LE:       # %bb.0: # %entry
-; P9LE-NEXT:    lhz r3, 0(r3)
+; P9LE-NEXT:    lxsihzx v2, 0, r3
+; P9LE-NEXT:    li r3, 0
+; P9LE-NEXT:    vsplth v2, v2, 3
+; P9LE-NEXT:    vextubrx r3, r3, v2
 ; P9LE-NEXT:    blr
 ;
 ; P9BE-LABEL: scalar_to_vector_half:
 ; P9BE:       # %bb.0: # %entry
-; P9BE-NEXT:    lhz r3, 0(r3)
-; P9BE-NEXT:    srwi r3, r3, 24
+; P9BE-NEXT:    lxsihzx v2, 0, r3
+; P9BE-NEXT:    li r3, 0
+; P9BE-NEXT:    vsplth v2, v2, 3
+; P9BE-NEXT:    vextublx r3, r3, v2
 ; P9BE-NEXT:    blr
 ;
 ; P8LE-LABEL: scalar_to_vector_half:
 ; P8LE:       # %bb.0: # %entry
 ; P8LE-NEXT:    lhz r3, 0(r3)
+; P8LE-NEXT:    mtfprd f0, r3
+; P8LE-NEXT:    mffprd r3, f0
+; P8LE-NEXT:    clrldi r3, r3, 56
 ; P8LE-NEXT:    blr
 ;
 ; P8BE-LABEL: scalar_to_vector_half:
 ; P8BE:       # %bb.0: # %entry
 ; P8BE-NEXT:    lhz r3, 0(r3)
-; P8BE-NEXT:    srwi r3, r3, 24
+; P8BE-NEXT:    sldi r3, r3, 48
+; P8BE-NEXT:    mtfprd f0, r3
+; P8BE-NEXT:    mffprd r3, f0
+; P8BE-NEXT:    rldicl r3, r3, 8, 56
 ; P8BE-NEXT:    blr
 entry:
     %0 = load <2 x i8>, ptr %ad, align 1