[PATCH] D67672: Don't use invalidated iterators in FlattenCFGPass

[Jakub Kuderski via Phabricator via llvm-commits]

kuhar created this revision.
kuhar added reviewers: dblaikie, tstellar, davide.
Herald added a subscriber: hiraditya.
Herald added a reviewer: grosser.
Herald added a project: LLVM.

FlattenCFG may erase unnecessary blocks, which also invalidates iterators to those erased blocks.
Before this patch, `iterativelyFlattenCFG` could try to increment a BB iterator after that BB has been removed and crash.

This patch makes FlattenCFGPass use `WeakVH` to skip over erased blocks.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D67672

Files:
  llvm/lib/Transforms/Scalar/FlattenCFGPass.cpp
  llvm/test/Transforms/Util/flattencfg.ll


Index: llvm/test/Transforms/Util/flattencfg.ll
===================================================================
--- llvm/test/Transforms/Util/flattencfg.ll
+++ llvm/test/Transforms/Util/flattencfg.ll
@@ -24,3 +24,33 @@
 exit:                               ; preds = %entry, %b0, %b1
   ret void
 }
+
+; CHECK-LABEL: @test_not_crash2
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    %0 = fcmp ult float %a
+; CHECK-NEXT:    %1 = fcmp ult float %b
+; CHECK-NEXT:    [[COND:%[a-z0-9]+]] = or i1 %0, %1
+; CHECK-NEXT:    br i1 [[COND]], label %bb4, label %bb3
+; CHECK:       bb3:
+; CHECK-NEXT:    br label %bb4
+; CHECK:       bb4:
+; CHECK-NEXT:    ret void
+define void @test_not_crash2(float %a, float %b) #0 {
+entry:
+  %0 = fcmp ult float %a, 1.000000e+00
+  br i1 %0, label %bb0, label %bb1
+
+bb3:                                               ; preds = %bb0
+  br label %bb4
+
+bb4:                                               ; preds = %bb0, %bb3
+  ret void
+
+bb1:                                               ; preds = %entry
+  br label %bb0
+
+bb0:                                               ; preds = %bb1, %entry
+  %1 = fcmp ult float %b, 1.000000e+00
+  br i1 %1, label %bb4, label %bb3
+}
+
Index: llvm/lib/Transforms/Scalar/FlattenCFGPass.cpp
===================================================================
--- llvm/lib/Transforms/Scalar/FlattenCFGPass.cpp
+++ llvm/lib/Transforms/Scalar/FlattenCFGPass.cpp
@@ -11,10 +11,12 @@
 //===----------------------------------------------------------------------===//
 
 #include "llvm/Analysis/AliasAnalysis.h"
-#include "llvm/Transforms/Utils/Local.h"
 #include "llvm/IR/CFG.h"
+#include "llvm/IR/ValueHandle.h"
 #include "llvm/Pass.h"
 #include "llvm/Transforms/Scalar.h"
+#include "llvm/Transforms/Utils/Local.h"
+
 using namespace llvm;
 
 #define DEBUG_TYPE "flattencfg"
@@ -52,15 +54,23 @@
 static bool iterativelyFlattenCFG(Function &F, AliasAnalysis *AA) {
   bool Changed = false;
   bool LocalChange = true;
+
+  // Use block handles instead of iterating over function blocks directly
+  // to avoid using iterators invalidated by erasing blocks.
+  std::vector<WeakVH> Blocks;
+  Blocks.reserve(F.size());
+  for (auto &BB : F)
+    Blocks.push_back(&BB);
+
   while (LocalChange) {
     LocalChange = false;
 
-    // Loop over all of the basic blocks and remove them if they are unneeded...
-    //
-    for (Function::iterator BBIt = F.begin(); BBIt != F.end();) {
-      if (FlattenCFG(&*BBIt++, AA)) {
-        LocalChange = true;
-      }
+    // Loop over all of the basic blocks and try to flatten them.
+    for (WeakVH &BlockHandle : Blocks) {
+      // Skip blocks erased by FlattenCFG.
+      if (auto *BB = cast_or_null<BasicBlock>(BlockHandle))
+        if (FlattenCFG(BB, AA))
+          LocalChange = true;
     }
     Changed |= LocalChange;
   }


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D67672.220545.patch
Type: text/x-patch
Size: 2841 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20190917/d07f4797/attachment.bin>