[PATCH] D135886: [MemCpyOpt] Move lifetime marker before call to enable call slot optimization

[Nikita Popov via Phabricator via llvm-commits]

nikic created this revision.
nikic added reviewers: fhahn, asbirlea, reames.
Herald added a subscriber: hiraditya.
Herald added a project: All.
nikic requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

Currently call slot optimization may be prevented because the lifetime markers for the destination only start after the call. In this case, rather than aborting the transform, we should move the lifetime.start before the call to enable the transform.


https://reviews.llvm.org/D135886

Files:
  llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
  llvm/test/Transforms/MemCpyOpt/lifetime.ll


Index: llvm/test/Transforms/MemCpyOpt/lifetime.ll
===================================================================
--- llvm/test/Transforms/MemCpyOpt/lifetime.ll
+++ llvm/test/Transforms/MemCpyOpt/lifetime.ll
@@ -55,9 +55,8 @@
 ; CHECK-LABEL: @call_slot_move_lifetime_start(
 ; CHECK-NEXT:    [[TMP:%.*]] = alloca i32, align 4
 ; CHECK-NEXT:    [[DST:%.*]] = alloca i32, align 4
-; CHECK-NEXT:    call void @call(ptr [[TMP]])
 ; CHECK-NEXT:    call void @llvm.lifetime.start.p0(i64 4, ptr [[DST]])
-; CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 4 [[DST]], ptr align 4 [[TMP]], i64 4, i1 false)
+; CHECK-NEXT:    call void @call(ptr [[DST]])
 ; CHECK-NEXT:    call void @llvm.lifetime.end.p0(i64 4, ptr [[DST]])
 ; CHECK-NEXT:    [[V:%.*]] = load i32, ptr [[DST]], align 4
 ; CHECK-NEXT:    ret i32 [[V]]
Index: llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
===================================================================
--- llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
+++ llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
@@ -331,16 +331,27 @@
 }
 
 // Check for mod or ref of Loc between Start and End, excluding both boundaries.
-// Start and End must be in the same block
+// Start and End must be in the same block.
+// If SkippedLifetimeStart is provided, skip over one clobbering lifetime.start
+// intrinsic and store it inside SkippedLifetimeStart.
 static bool accessedBetween(AliasAnalysis &AA, MemoryLocation Loc,
                             const MemoryUseOrDef *Start,
-                            const MemoryUseOrDef *End) {
+                            const MemoryUseOrDef *End,
+                            Instruction **SkippedLifetimeStart = nullptr) {
   assert(Start->getBlock() == End->getBlock() && "Only local supported");
   for (const MemoryAccess &MA :
        make_range(++Start->getIterator(), End->getIterator())) {
-    if (isModOrRefSet(AA.getModRefInfo(cast<MemoryUseOrDef>(MA).getMemoryInst(),
-                                       Loc)))
+    Instruction *I = cast<MemoryUseOrDef>(MA).getMemoryInst();
+    if (isModOrRefSet(AA.getModRefInfo(I, Loc))) {
+      auto *II = dyn_cast<IntrinsicInst>(I);
+      if (II && II->getIntrinsicID() == Intrinsic::lifetime_start &&
+          SkippedLifetimeStart && !*SkippedLifetimeStart) {
+        *SkippedLifetimeStart = I;
+        continue;
+      }
+
       return true;
+    }
   }
   return false;
 }
@@ -913,8 +924,9 @@
 
   // Check that nothing touches the dest of the copy between
   // the call and the store/memcpy.
+  Instruction *SkippedLifetimeStart = nullptr;
   if (accessedBetween(*AA, DestLoc, MSSA->getMemoryAccess(C),
-                      MSSA->getMemoryAccess(cpyStore))) {
+                      MSSA->getMemoryAccess(cpyStore), &SkippedLifetimeStart)) {
     LLVM_DEBUG(dbgs() << "Call Slot: Dest pointer modified after call\n");
     return false;
   }
@@ -1094,6 +1106,12 @@
     cast<AllocaInst>(cpyDest)->setAlignment(srcAlign);
   }
 
+  if (SkippedLifetimeStart) {
+    SkippedLifetimeStart->moveBefore(C);
+    MSSAU->moveBefore(MSSA->getMemoryAccess(SkippedLifetimeStart),
+                      MSSA->getMemoryAccess(C));
+  }
+
   // Update AA metadata
   // FIXME: MD_tbaa_struct and MD_mem_parallel_loop_access should also be
   // handled here, but combineMetadata doesn't support them yet


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D135886.467494.patch
Type: text/x-patch
Size: 3333 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20221013/ffc8c299/attachment.bin>