[llvm] [PGO] Add llvm.loop.estimated_trip_count metadata (PR #152775)

[Joel E. Denny via llvm-commits]

================
@@ -834,43 +859,99 @@ static std::optional<unsigned> getEstimatedTripCount(BranchInst *ExitingBranch,
     return std::numeric_limits<unsigned>::max();
 
   // Estimated trip count is one plus estimated exit count.
-  return ExitCount + 1;
+  uint64_t TC = ExitCount + 1;
+  LLVM_DEBUG(dbgs() << "estimateLoopTripCount: Estimated trip count of " << TC
+                    << " for " << DbgLoop(L) << "\n");
+  return TC;
 }
 
 std::optional<unsigned>
 llvm::getLoopEstimatedTripCount(Loop *L,
                                 unsigned *EstimatedLoopInvocationWeight) {
-  // Currently we take the estimate exit count only from the loop latch,
-  // ignoring other exiting blocks.  This can overestimate the trip count
-  // if we exit through another exit, but can never underestimate it.
-  // TODO: incorporate information from other exits
-  if (BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L)) {
-    uint64_t ExitWeight;
-    if (std::optional<uint64_t> EstTripCount =
-            getEstimatedTripCount(LatchBranch, L, ExitWeight)) {
-      if (EstimatedLoopInvocationWeight)
-        *EstimatedLoopInvocationWeight = ExitWeight;
-      return *EstTripCount;
-    }
+  // If EstimatedLoopInvocationWeight, we do not support this loop if
+  // getExpectedExitLoopLatchBranch returns nullptr.
+  //
+  // FIXME: Also, this is a stop-gap solution for nested loops.  It avoids
+  // mistaking LLVMLoopEstimatedTripCount metadata to be for an outer loop when
+  // it was created for an inner loop.  The problem is that loop metadata is
+  // attached to the branch instruction in the loop latch block, but that can be
+  // shared by the loops.  The solution is to attach loop metadata to loop
+  // headers instead, but that would be a large change to LLVM.
+  //
+  // Until that happens, we work around the problem as follows.
+  // getExpectedExitLoopLatchBranch (which also guards
+  // setLoopEstimatedTripCount) will not recognize the same latch for both loops
+  // unless the latch exits both loops and has only two successors.  However, to
+  // exit both loops, the latch must have at least three successors: the inner
+  // loop header, the outer loop header (exit for the inner loop), and an exit
+  // for the outer loop.
+  BranchInst *ExitingBranch = getExpectedExitLoopLatchBranch(L);
+  if (!ExitingBranch)
+    return std::nullopt;
+
+  // If requested, either compute *EstimatedLoopInvocationWeight or return
+  // nullopt if cannot.
+  //
+  // TODO: Eventually, once all passes have migrated away from setting branch
+  // weights to indicate estimated trip counts, this function will drop the
+  // EstimatedLoopInvocationWeight parameter.
+  if (EstimatedLoopInvocationWeight) {
+    uint64_t LoopWeight = 0, ExitWeight = 0; // Inits expected to be unused.
+    if (!extractBranchWeights(*ExitingBranch, LoopWeight, ExitWeight))
+      return std::nullopt;
+    if (L->contains(ExitingBranch->getSuccessor(1)))
+      std::swap(LoopWeight, ExitWeight);
+    if (!ExitWeight)
+      return std::nullopt;
+    *EstimatedLoopInvocationWeight = ExitWeight;
   }
-  return std::nullopt;
+
+  // Return the estimated trip count from metadata unless the metadata is
+  // missing or has no value.
+  if (auto TC = getOptionalIntLoopAttribute(L, LLVMLoopEstimatedTripCount)) {
+    LLVM_DEBUG(dbgs() << "getLoopEstimatedTripCount: "
+                      << LLVMLoopEstimatedTripCount << " metadata has trip "
+                      << "count of " << *TC << " for " << DbgLoop(L) << "\n");
+    return TC;
+  }
+
+  // Estimate the trip count from latch branch weights.
+  return estimateLoopTripCount(L);
 }
 
-bool llvm::setLoopEstimatedTripCount(Loop *L, unsigned EstimatedTripCount,
-                                     unsigned EstimatedloopInvocationWeight) {
-  // At the moment, we currently support changing the estimate trip count of
-  // the latch branch only.  We could extend this API to manipulate estimated
-  // trip counts for any exit.
+bool llvm::setLoopEstimatedTripCount(
+    Loop *L, unsigned EstimatedTripCount,
+    std::optional<unsigned> EstimatedloopInvocationWeight) {
+  // If EstimatedLoopInvocationWeight, we do not support this loop if
+  // getExpectedExitLoopLatchBranch returns nullptr.
+  //
+  // FIXME: See comments in getLoopEstimatedTripCount for why this is required
+  // here regardless of EstimatedLoopInvocationWeight.
   BranchInst *LatchBranch = getExpectedExitLoopLatchBranch(L);
   if (!LatchBranch)
     return false;
 
+  // Set the metadata.
+  addStringMetadataToLoop(L, LLVMLoopEstimatedTripCount, EstimatedTripCount);
+
+  // At the moment, we currently support changing the estimated trip count in
+  // the latch branch's branch weights only.  We could extend this API to
+  // manipulate estimated trip counts for any exit.
+  //
+  // TODO: Eventually, once all passes have migrated away from setting branch
+  // weights to indicate estimated trip counts, we will not set branch weights
+  // here at all.
+  if (!EstimatedloopInvocationWeight)
+    return true;
+  if (!LatchBranch)
+    return false;
----------------
jdenny-ornl wrote:

Oops, this is redundant now.  Will remove.

https://github.com/llvm/llvm-project/pull/152775