[llvm] [PseudoProbe] Extend to skip instrumenting probe into the dests of invoke (PR #79919)

[Lei Wang via llvm-commits]

================
@@ -173,21 +173,98 @@ SampleProfileProber::SampleProfileProber(Function &Func,
   BlockProbeIds.clear();
   CallProbeIds.clear();
   LastProbeId = (uint32_t)PseudoProbeReservedId::Last;
-  computeProbeIdForBlocks();
-  computeProbeIdForCallsites();
-  computeCFGHash();
+
+  DenseSet<BasicBlock *> BlocksToIgnoreProbe;
+  DenseSet<BasicBlock *> BlocksToIgnoreCall;
+  computeBlocksToIgnore(BlocksToIgnoreProbe, BlocksToIgnoreCall);
+
+  computeProbeIdForBlocks(BlocksToIgnoreProbe);
+  computeProbeIdForCallsites(BlocksToIgnoreCall);
+  computeCFGHash(BlocksToIgnoreProbe);
+}
+
+// Two purposes to compute the blocks to ignore:
+// 1. Reduce the IR size.
+// 2. Make the instrumentation(checksum mismatch) stable. e.g. the frondend may
+// generate unstable IR while optimizing nounwind attribute, some versions are
+// optimized with the call-to-invoke conversion, while other versions do not.
+// This discrepancy in probe ID could cause profile mismatching issues.
+// Note that those ignored blocks are either cold blocks or new split blocks
+// whose original blocks are instrumented, so it shouldn't degrade the profile
+// quailty.
+void SampleProfileProber::computeBlocksToIgnore(
+    DenseSet<BasicBlock *> &BlocksToIgnoreProbe,
+    DenseSet<BasicBlock *> &BlocksToIgnoreCall) {
+  // Ignore the cold EH blocks.
+  computeEHOnlyBlocks(*F, BlocksToIgnoreCall);
+  findUnreachableBlocks(BlocksToIgnoreCall);
+
+  BlocksToIgnoreProbe.insert(BlocksToIgnoreCall.begin(),
+                             BlocksToIgnoreCall.end());
+  findNewSplitBlocks(BlocksToIgnoreProbe);
+}
+
+void SampleProfileProber::findUnreachableBlocks(
+    DenseSet<BasicBlock *> &BlocksToIgnore) {
+  for (auto &BB : *F) {
+    if (&BB != &F->getEntryBlock() && pred_size(&BB) == 0)
+      BlocksToIgnore.insert(&BB);
+  }
+}
+
+// Basic block can be split into multiple blocks, e.g. due to the
+// call-to-invoke. If they are hotness-wise equal, we can optimize to only
+// instrument the leading block, ignore the other new split blocks.
----------------
wlei-llvm wrote:

Sounds good, changed to make it narrow and limit the mergeable into predecessor case only to the EH flow. 

https://github.com/llvm/llvm-project/pull/79919