[MemorySSA +LICM CFHoist] Solve PR40317.

[asbirlea]

Summary

MemorySSA needs updating each time an instruction is moved.
LICM and control flow hoisting re-hoists instructions, thus needing another update when re-moving those instructions.
Pending cleanup: the MSSA update is duplicated, should be moved inside moveInstructionBefore.

[asbirlea]

Add test-case.

llvm/trunk/lib/Transforms/Scalar/LICM.cpp

for (Instruction *I : reverse(HoistedInstructions)) {
             assert(DT->dominates(Dominator, HoistPoint->getParent()) &&
                    "New hoist point expected to dominate old hoist point");
           HoistPoint = Dominator->getTerminator();
         }
         LLVM_DEBUG(dbgs() << "LICM rehoisting to "
                           << HoistPoint->getParent()->getName()
                           << ": " << *I << "\n");
         moveInstructionBefore(*I, *HoistPoint, *SafetyInfo);
+        if (MSSAU)
+          if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>(
+                  MSSAU->getMemorySSA()->getMemoryAccess(I)))
+            MSSAU->moveToPlace(OldMemAcc, HoistPoint->getParent(),
+                               MemorySSA::End);
         HoistPoint = I;
         Changed = true;
       }
     }
   }
   if (MSSAU && VerifyMemorySSA)
     MSSAU->getMemorySSA()->verifyMemorySSA();
 

llvm/trunk/test/Transforms/LICM/pr40317.ll

+; RUN: opt -S -march=z13 -tbaa -licm -enable-mssa-loop-dependency -licm-control-flow-hoisting -verify-memoryssa < %s | FileCheck %s
+
+target datalayout = "E-m:e-i1:8:16-i8:8:16-i64:64-f128:64-v128:64-a:8:16-n32:64"
+target triple = "s390x-ibm-linux"
+
+%0 = type { %1, %1, i16, %2 }
+%1 = type <{ i16, i8, i32, i32, i32, i64, i64 }>
+%2 = type { i8, i16, i16, [2 x i8] }
+
+@0 = internal global %0 { %1 <{ i16 22437, i8 117, i32 2017322857, i32 900074563, i32 -1390364, i64 0, i64 0 }>, %1 <{ i16 0, i8 7, i32 -387299562, i32 925371866, i32 -1, i64 4826244575317081679, i64 1 }>, i16 8, %2 { i8 0, i16 0, i16 3, [2 x i8] undef } }, align 2
+@g_18 = external dso_local global i64, align 8
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.memset.p0i8.i64(i8* nocapture writeonly, i8, i64, i1) #1
+
+; CHECK-LABEL: @func_94
+; CHECK: bb:
+; CHECK: tail call void @llvm.memset.p0i8.i64
+; CHECK: load i32
+; CHECK: bb6.licm:
+; Function Attrs: noreturn nounwind
+define dso_local void @func_94(i16 %arg, i64* nocapture %arg1) local_unnamed_addr #3 {
+bb:
+  tail call void @llvm.memset.p0i8.i64(i8* align 8 undef, i8 0, i64 80, i1 false)
+  br label %bb3
+
+bb3:                                              ; preds = %bb13, %bb
+  %tmp5 = icmp eq i16 %arg, 0
+  br i1 %tmp5, label %bb6, label %bb13
+
+bb6:                                              ; preds = %bb3
+  %tmp7 = load i32, i32* getelementptr inbounds (%0, %0* @0, i64 0, i32 1, i32 2), align 1, !tbaa !11
+  %tmp8 = zext i32 %tmp7 to i64
+  %sext = shl i64 %tmp8, 56
+  %tmp10 = ashr exact i64 %sext, 56
+  store i64 %tmp10, i64* %arg1, align 8, !tbaa !12
+  br label %bb13
+
+bb13:                                             ; preds = %bb3, %bb6
+  br label %bb3
+}
+
+attributes #0 = { "use-soft-float"="false" }
+attributes #1 = { argmemonly nounwind }
+attributes #2 = { norecurse nounwind readnone "use-soft-float"="false" }
+attributes #3 = { noreturn nounwind "use-soft-float"="false" }
+
+!llvm.ident = !{!0}
+
+!0 = !{!"clang version 8.0.0 (http://llvm.org/git/clang.git e593a791f2cf19db84237b0b9d632e9966a00a39) (http://llvm.org/git/llvm.git fe0523d1bd7def3ef62cfb3dd37a8b1941aafa81)"}
+!1 = !{!2, !8, i64 46}
+!2 = !{!"S5", !3, i64 0, !3, i64 31, !4, i64 62, !9, i64 64}
+!3 = !{!"S2", !4, i64 0, !5, i64 2, !7, i64 3, !7, i64 7, !7, i64 11, !8, i64 15, !8, i64 23}
+!4 = !{!"short", !5, i64 0}
+!5 = !{!"omnipotent char", !6, i64 0}
+!6 = !{!"Simple C/C++ TBAA"}
+!7 = !{!"int", !5, i64 0}
+!8 = !{!"long", !5, i64 0}
+!9 = !{!"S3", !7, i64 0, !4, i64 2, !4, i64 4}
+!10 = !{!2, !7, i64 42}
+!11 = !{!2, !7, i64 34}
+!12 = !{!8, !8, i64 0}