[DAGCombine] Handle promotion of shift with both operands the same

[nikic]

Summary

When promoting a shift, make sure we only fetch the second operand after promoting the first. Load promotion may replace users of the old load, and we don't want to be left with a dangling reference to the old load instruction.

The crashing test case is from https://reviews.llvm.org/D126689#3553212.

[craig.topper]

LGTM

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
   // what's the right type to promote it to.
   if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
     assert(PVT != VT && "Don't know what type to promote to!");
 
     LLVM_DEBUG(dbgs() << "\nPromoting "; Op.dump(&DAG));
 
     bool Replace = false;
     SDValue N0 = Op.getOperand(0);
-    SDValue N1 = Op.getOperand(1);
     if (Opc == ISD::SRA)
       N0 = SExtPromoteOperand(N0, PVT);
     else if (Opc == ISD::SRL)
       N0 = ZExtPromoteOperand(N0, PVT);
     else
       N0 = PromoteOperand(N0, PVT, Replace);
 
     if (!N0.getNode())
       return SDValue();
 
     SDLoc DL(Op);
+    SDValue N1 = Op.getOperand(1);
     SDValue RV =
         DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, N0, N1));
 
     if (Replace)
       ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
 
     // Deal with Op being deleted.
     if (Op && Op.getOpcode() != ISD::DELETED_NODE)

llvm/test/CodeGen/X86/promote-sra-by-itself.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu < %s | FileCheck %s
+
+declare void @use(<1 x i16>)
+
+define i16 @basic(i16* %p) {
+; CHECK-LABEL: basic:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movswl (%rdi), %eax
+; CHECK-NEXT:    movl %eax, %ecx
+; CHECK-NEXT:    sarl %cl, %eax
+; CHECK-NEXT:    # kill: def $ax killed $ax killed $eax
+; CHECK-NEXT:    retq
+  %v = load i16, i16* %p
+  %s = ashr i16 %v, %v
+  ret i16 %s
+}
+
+define void @crash(i1 %cond, <1 x i16>* %p) {
+; CHECK-LABEL: crash:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    .p2align 4, 0x90
+; CHECK-NEXT:  .LBB1_1: # %loop
+; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    testb $1, %dil
+; CHECK-NEXT:    jne .LBB1_1
+; CHECK-NEXT:  # %bb.2: # %exit
+; CHECK-NEXT:    retq
+  br label %loop
+
+loop:
+  %v = load <1 x i16>, <1 x i16>* %p, align 2
+  %ins = insertelement <4 x double> zeroinitializer, double 0.000000e+00, i32 0
+  %cmp = fcmp uge <4 x double> %ins, zeroinitializer
+  %ashr = ashr <1 x i16> %v, %v
+  %shuf = shufflevector <4 x i1> %cmp, <4 x i1> zeroinitializer, <4 x i32> zeroinitializer
+  br i1 %cond, label %loop, label %exit
+
+exit:
+  %use1 = add <4 x i1> %shuf, zeroinitializer
+  %use2 = add <1 x i16> %ashr, zeroinitializer
+  ret void
+}