[llvm] r363597 - [MemorySSA] Don't use template when the clone is a simplified instruction.
Alina Sbirlea via llvm-commits
llvm-commits at lists.llvm.org
Mon Jun 17 11:58:40 PDT 2019
Author: asbirlea
Date: Mon Jun 17 11:58:40 2019
New Revision: 363597
URL: http://llvm.org/viewvc/llvm-project?rev=363597&view=rev
Log:
[MemorySSA] Don't use template when the clone is a simplified instruction.
Summary:
LoopRotate doesn't create a faithful clone of an instruction, it may
simplify it beforehand. Hence the clone of an instruction that has a
MemoryDef associated may not be a definition, but a use or not a memory
alternig instruction.
Don't rely on the template when the clone may be simplified.
Reviewers: george.burgess.iv
Subscribers: jlebar, Prazek, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D63355
Added:
llvm/trunk/test/Analysis/MemorySSA/loop-rotate-inv-template.ll
Modified:
llvm/trunk/include/llvm/Analysis/MemorySSAUpdater.h
llvm/trunk/lib/Analysis/MemorySSAUpdater.cpp
Modified: llvm/trunk/include/llvm/Analysis/MemorySSAUpdater.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Analysis/MemorySSAUpdater.h?rev=363597&r1=363596&r2=363597&view=diff
==============================================================================
--- llvm/trunk/include/llvm/Analysis/MemorySSAUpdater.h (original)
+++ llvm/trunk/include/llvm/Analysis/MemorySSAUpdater.h Mon Jun 17 11:58:40 2019
@@ -287,8 +287,14 @@ private:
// not necessarily be MemoryPhis themselves, they may be MemoryDefs. As such,
// the map is between MemoryPhis and MemoryAccesses, where the MemoryAccesses
// may be MemoryPhis or MemoryDefs and not MemoryUses.
+ // If CloneWasSimplified = true, the clone was exact. Otherwise, assume that
+ // the clone involved simplifications that may have: (1) turned a MemoryUse
+ // into an instruction that MemorySSA has no representation for, or (2) turned
+ // a MemoryDef into a MemoryUse or an instruction that MemorySSA has no
+ // representation for. No other cases are supported.
void cloneUsesAndDefs(BasicBlock *BB, BasicBlock *NewBB,
- const ValueToValueMapTy &VMap, PhiToDefMap &MPhiMap);
+ const ValueToValueMapTy &VMap, PhiToDefMap &MPhiMap,
+ bool CloneWasSimplified = false);
template <typename Iter>
void privateUpdateExitBlocksForClonedLoop(ArrayRef<BasicBlock *> ExitBlocks,
Iter ValuesBegin, Iter ValuesEnd,
Modified: llvm/trunk/lib/Analysis/MemorySSAUpdater.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/MemorySSAUpdater.cpp?rev=363597&r1=363596&r2=363597&view=diff
==============================================================================
--- llvm/trunk/lib/Analysis/MemorySSAUpdater.cpp (original)
+++ llvm/trunk/lib/Analysis/MemorySSAUpdater.cpp Mon Jun 17 11:58:40 2019
@@ -482,7 +482,8 @@ void MemorySSAUpdater::removeDuplicatePh
void MemorySSAUpdater::cloneUsesAndDefs(BasicBlock *BB, BasicBlock *NewBB,
const ValueToValueMapTy &VMap,
- PhiToDefMap &MPhiMap) {
+ PhiToDefMap &MPhiMap,
+ bool CloneWasSimplified) {
auto GetNewDefiningAccess = [&](MemoryAccess *MA) -> MemoryAccess * {
MemoryAccess *InsnDefining = MA;
if (MemoryUseOrDef *DefMUD = dyn_cast<MemoryUseOrDef>(InsnDefining)) {
@@ -512,10 +513,14 @@ void MemorySSAUpdater::cloneUsesAndDefs(
// instructions. This occurs in LoopRotate when cloning instructions
// from the old header to the old preheader. The cloned instruction may
// also be a simplified Value, not an Instruction (see LoopRotate).
+ // Also in LoopRotate, even when it's an instruction, due to it being
+ // simplified, it may be a Use rather than a Def, so we cannot use MUD as
+ // template. Calls coming from updateForClonedBlockIntoPred, ensure this.
if (Instruction *NewInsn =
dyn_cast_or_null<Instruction>(VMap.lookup(Insn))) {
MemoryAccess *NewUseOrDef = MSSA->createDefinedAccess(
- NewInsn, GetNewDefiningAccess(MUD->getDefiningAccess()), MUD);
+ NewInsn, GetNewDefiningAccess(MUD->getDefiningAccess()),
+ CloneWasSimplified ? nullptr : MUD);
MSSA->insertIntoListsForBlock(NewUseOrDef, NewBB, MemorySSA::End);
}
}
@@ -645,10 +650,13 @@ void MemorySSAUpdater::updateForClonedBl
// Defs from BB being used in BB will be replaced with the cloned defs from
// VM. The uses of BB's Phi (if it exists) in BB will be replaced by the
// incoming def into the Phi from P1.
+ // Instructions cloned into the predecessor are in practice sometimes
+ // simplified, so disable the use of the template, and create an access from
+ // scratch.
PhiToDefMap MPhiMap;
if (MemoryPhi *MPhi = MSSA->getMemoryAccess(BB))
MPhiMap[MPhi] = MPhi->getIncomingValueForBlock(P1);
- cloneUsesAndDefs(BB, P1, VM, MPhiMap);
+ cloneUsesAndDefs(BB, P1, VM, MPhiMap, /*CloneWasSimplified=*/true);
}
template <typename Iter>
Added: llvm/trunk/test/Analysis/MemorySSA/loop-rotate-inv-template.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Analysis/MemorySSA/loop-rotate-inv-template.ll?rev=363597&view=auto
==============================================================================
--- llvm/trunk/test/Analysis/MemorySSA/loop-rotate-inv-template.ll (added)
+++ llvm/trunk/test/Analysis/MemorySSA/loop-rotate-inv-template.ll Mon Jun 17 11:58:40 2019
@@ -0,0 +1,27 @@
+; RUN: opt -disable-output -loop-rotate -enable-mssa-loop-dependency -verify-memoryssa %s
+; REQUIRES: asserts
+
+; Function Attrs: nounwind
+define dso_local void @bar() local_unnamed_addr #0 align 32 {
+entry:
+ br label %looplabel.exit.i
+
+looplabel.exit.i: ; preds = %if.end.i, %entry
+ %0 = phi i1 (i32*, i32*)* [ @foo, %entry ], [ undef, %if.end.i ]
+ %call3.i.i = call zeroext i1 %0(i32* nonnull dereferenceable(16) undef, i32* nonnull undef)
+ br i1 %call3.i.i, label %if.end.i, label %label.exit
+
+if.end.i: ; preds = %looplabel.exit.i
+ %tobool.i = icmp eq i32* undef, null
+ br label %looplabel.exit.i
+
+label.exit: ; preds = %looplabel.exit.i
+ ret void
+}
+
+; Function Attrs: readonly
+declare dso_local i1 @foo(i32*, i32*) #1 align 32
+
+attributes #0 = { nounwind }
+attributes #1 = { readonly }
+
More information about the llvm-commits
mailing list