[JumpThreading] Put a limit on the PHI nodes when duplicating a BB.
authorUsman Nadeem <mnadeem@quicinc.com>
Mon, 31 Oct 2022 22:08:11 +0000 (15:08 -0700)
committerUsman Nadeem <mnadeem@quicinc.com>
Mon, 31 Oct 2022 22:51:56 +0000 (15:51 -0700)
Do not duplicate a BB if it has a lot of PHI nodes.
If a threadable chain is too long then the number of duplicated PHI nodes
can add up, leading to a substantial increase in compile time when rewriting
the SSA.

Fixes https://github.com/llvm/llvm-project/issues/58203
Differential Revision: https://reviews.llvm.org/D136716

The threshold of 76 in this patch is reasonably high and reduces the compile
time of cldwat2m_macro.f90 in SPEC2017/cam4 from 80+min to <2min.

Change-Id: I153c89a8e0d89b206a5193dc1b908c67e320717e

llvm/lib/Transforms/Scalar/JumpThreading.cpp

index f9cd4cd..75bc4bb 100644 (file)
@@ -100,6 +100,11 @@ ImplicationSearchThreshold(
            "condition to use to thread over a weaker condition"),
   cl::init(3), cl::Hidden);
 
+static cl::opt<unsigned> PhiDuplicateThreshold(
+    "jump-threading-phi-threshold",
+    cl::desc("Max PHIs in BB to duplicate for jump threading"), cl::init(76),
+    cl::Hidden);
+
 static cl::opt<bool> PrintLVIAfterJumpThreading(
     "print-lvi-after-jump-threading",
     cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
@@ -518,8 +523,23 @@ static unsigned getJumpThreadDuplicationCost(const TargetTransformInfo *TTI,
                                              Instruction *StopAt,
                                              unsigned Threshold) {
   assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");
+
+  // Do not duplicate the BB if it has a lot of PHI nodes.
+  // If a threadable chain is too long then the number of PHI nodes can add up,
+  // leading to a substantial increase in compile time when rewriting the SSA.
+  unsigned PhiCount = 0;
+  Instruction *FirstNonPHI = nullptr;
+  for (Instruction &I : *BB) {
+    if (!isa<PHINode>(&I)) {
+      FirstNonPHI = &I;
+      break;
+    }
+    if (++PhiCount > PhiDuplicateThreshold)
+      return ~0U;
+  }
+
   /// Ignore PHI nodes, these will be flattened when duplication happens.
-  BasicBlock::const_iterator I(BB->getFirstNonPHI());
+  BasicBlock::const_iterator I(FirstNonPHI);
 
   // FIXME: THREADING will delete values that are just used to compute the
   // branch, so they shouldn't count against the duplication cost.