Merge pull request #66328 from meg-gupta/fixtmprvalue

Disable TempRValueOpt when there are users of the temp object preceding the copy

Author: swift-ci

lib/SILOptimizer/Transforms/TempRValueElimination.cpp

@@ -547,9 +547,13 @@ void TempRValueOptPass::tryOptimizeCopyIntoTemp(CopyAddrInst *copyInst) {
   // only users that modify memory are the copy_addr [initialization] and
   // destroy_addr.
   InstructionSetWithSize loadInsts(getFunction());
+  // Set of tempObj users
+  InstructionSet userSet(getFunction());
   for (auto *useOper : tempObj->getUses()) {
     SILInstruction *user = useOper->getUser();
 
+    userSet.insert(user);
+
     if (user == copyInst)
       continue;
 
@@ -576,6 +580,20 @@ void TempRValueOptPass::tryOptimizeCopyIntoTemp(CopyAddrInst *copyInst) {
       return;
   }
 
+  // Check and return without optimization if we have any users of tempObj that
+  // precede the copyInst.
+  // This can happen with projections.
+  // TODO: We can enable this case if we clone the projections at "load" uses
+
+  // All instructions in userSet are in the same block as copyInst. collectLoads
+  // ensures of this.
+  for (SILInstruction &inst : llvm::make_range(copyInst->getParent()->begin(),
+                                               copyInst->getIterator())) {
+    if (userSet.contains(&inst)) {
+      return;
+    }
+  }
+
   AliasAnalysis *aa = getPassManager()->getAnalysis<AliasAnalysis>(getFunction());
 
   // Check if the source is modified within the lifetime of the temporary.

test/SILOptimizer/temp_rvalue_opt_ossa.sil

@@ -1847,3 +1847,21 @@ bb0:
   dealloc_stack %src : $*MOS
   return %instance_3 : $MOS
 }
+
+// CHECK-LABEL: sil [ossa] @dont_optimize_use_before_copy :
+// CHECK: copy_addr
+// CHECK-LABEL: } // end sil function 'dont_optimize_use_before_copy'
+sil [ossa] @dont_optimize_use_before_copy : $@convention(thin) <B> (@in_guaranteed GS<B>, @inout GS<B>) -> () {
+bb0(%0 : $*GS<B>, %1 : $*GS<B>):
+  %2 = alloc_stack $GS<B>
+  %3 = struct_element_addr %2 : $*GS<B>, #GS._value
+  copy_addr %1 to [init] %2 : $*GS<B>
+  %5 = load [trivial] %3 : $*Builtin.Int64
+  %6 = builtin "cmp_slt_Int64"(%5 : $Builtin.Int64, %5 : $Builtin.Int64) : $Builtin.Int1
+  copy_addr %2 to %1 : $*GS<B>
+  destroy_addr %2 : $*GS<B>
+  dealloc_stack %2 : $*GS<B>
+  %10 = tuple ()
+  return %10 : $()
+}
+