[PGO] Fix zeroed estimated trip count

llvm/docs/LangRef.rst

@@ -8063,6 +8063,21 @@ pass should record the new estimates by calling
 loop, ``llvm::getLoopEstimatedTripCount`` returns its value instead of
 estimating the trip count from the loop's ``branch_weights`` metadata.
 
+Zero
+""""
+
+Some passes set ``llvm.loop.estimated_trip_count`` to 0.  For example, after
+peeling 10 or more iterations from a loop with an estimated trip count of 10,
+``llvm.loop.estimated_trip_count`` becomes 0 on the remaining loop.  It
+indicates that, each time execution reaches the peeled iterations, execution is
+estimated to exit them without reaching the remaining loop's header.
+
+Even if the probability of reaching a loop's header is low, if it is reached, it
+is the start of an iteration.  Consequently, some passes historically assume
+that ``llvm::getLoopEstimatedTripCount`` always returns a positive count or
+``std::nullopt``.  Thus, it returns ``std::nullopt`` when
+``llvm.loop.estimated_trip_count`` is 0.
+
 '``llvm.licm.disable``' Metadata
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 

llvm/include/llvm/Transforms/Utils/LoopUtils.h

@@ -326,7 +326,10 @@ LLVM_ABI void addStringMetadataToLoop(Loop *TheLoop, const char *MDString,
 /// - \c std::nullopt, if the implementation is unable to handle the loop form
 ///   of \p L (e.g., \p L must have a latch block that controls the loop exit).
 /// - The value of \c llvm.loop.estimated_trip_count from the loop metadata of
-///   \p L, if that metadata is present.
+///   \p L, if that metadata is present.  In the special case that the value is
+///   zero, return \c std::nullopt instead as that is historically what callers
+///   expect when a loop is estimated to execute no iterations (i.e., its header
+///   is not reached).
 /// - Else, a new estimate of the trip count from the latch branch weights of
 ///   \p L.
 ///
@@ -353,10 +356,11 @@ getLoopEstimatedTripCount(Loop *L,
 /// to handle the loop form of \p L (e.g., \p L must have a latch block that
 /// controls the loop exit).  Otherwise, return true.
 ///
-/// In addition, if \p EstimatedLoopInvocationWeight, set the branch weight
-/// metadata of \p L to reflect that \p L has an estimated
-/// \p EstimatedTripCount iterations and has \c *EstimatedLoopInvocationWeight
-/// exit weight through the loop's latch.
+/// In addition, if \p EstimatedLoopInvocationWeight:
+/// - Set the branch weight metadata of \p L to reflect that \p L has an
+///   estimated \p EstimatedTripCount iterations and has
+///   \c *EstimatedLoopInvocationWeight exit weight through the loop's latch.
+/// - If \p EstimatedTripCount is zero, zero the branch weights.
 ///
 /// TODO: Eventually, once all passes have migrated away from setting branch
 /// weights to indicate estimated trip counts, this function will drop the

llvm/lib/Transforms/Utils/LoopUtils.cpp

@@ -912,12 +912,14 @@ llvm::getLoopEstimatedTripCount(Loop *L,
   }
 
   // Return the estimated trip count from metadata unless the metadata is
-  // missing or has no value.
+  // missing or has no value.  Return std::nullopt if it's zero.
   if (auto TC = getOptionalIntLoopAttribute(L, LLVMLoopEstimatedTripCount)) {
     LLVM_DEBUG(dbgs() << "getLoopEstimatedTripCount: "
                       << LLVMLoopEstimatedTripCount << " metadata has trip "
-                      << "count of " << *TC << " for " << DbgLoop(L) << "\n");
-    return TC;
+                      << "count of " << *TC
+                      << (*TC == 0 ? " (returning std::nullopt)" : "")
+                      << " for " << DbgLoop(L) << "\n");
+    return *TC == 0 ? std::nullopt : std::optional(*TC);
   }
 
   // Estimate the trip count from latch branch weights.

llvm/test/Transforms/LoopVectorize/vectorize-zero-estimated-trip-count.ll

@@ -0,0 +1,34 @@
+; Check that an estimated trip count of zero does not crash or otherwise break
+; LoopVectorize behavior while it tries to create runtime memory checks inside
+; an outer loop.
+
+; RUN: opt -passes=loop-vectorize -S %s | FileCheck %s
+
+target triple = "x86_64-unknown-linux-gnu"
+
+; Look for basic signs that vectorization ran and produced memory checks.
+; CHECK: @test(
+; CHECK: vector.memcheck:
+; CHECK: vector.body:
+; CHECK: inner:
+
+define void @test(ptr addrspace(1) %p, i32 %n) {
+entry:
+  br label %outer
+outer:
+  br label %inner
+inner:
+  %i = phi i32 [ %inc, %inner ], [ 0, %outer ]
+  store i32 0, ptr addrspace(1) %p
+  %load = load i32, ptr addrspace(1) null
+  %inc = add i32 %i, 1
+  %cmp = icmp slt i32 %i, %n
+  br i1 %cmp, label %inner, label %outer.latch
+outer.latch:
+  br i1 %cmp, label %outer, label %exit, !llvm.loop !0
+exit:
+  ret void
+}
+
+!0 = distinct !{!0, !1}
+!1 = !{!"llvm.loop.estimated_trip_count", i32 0}

llvm/test/Verifier/llvm.loop.estimated_trip_count.ll

@@ -36,12 +36,24 @@ exit:
 ; RUN: echo '!1 = !{!"llvm.loop.estimated_trip_count", i16 5}' >> %t
 ; RUN: %{RUN} GOOD
 
-; i32 value.
+; i32 arbitrary value.
 ; RUN: cp %s %t
 ; RUN: chmod u+w %t
 ; RUN: echo '!1 = !{!"llvm.loop.estimated_trip_count", i32 5}' >> %t
 ; RUN: %{RUN} GOOD
 
+; i32 boundary value of 1.
+; RUN: cp %s %t
+; RUN: chmod u+w %t
+; RUN: echo '!1 = !{!"llvm.loop.estimated_trip_count", i32 1}' >> %t
+; RUN: %{RUN} GOOD
+
+; i32 boundary value of 0.
+; RUN: cp %s %t
+; RUN: chmod u+w %t
+; RUN: echo '!1 = !{!"llvm.loop.estimated_trip_count", i32 0}' >> %t
+; RUN: %{RUN} GOOD
+
 ; i64 value.
 ; RUN: cp %s %t
 ; RUN: chmod u+w %t

llvm/unittests/Transforms/Utils/LoopUtilsTest.cpp

@@ -14,6 +14,7 @@
 #include "llvm/AsmParser/Parser.h"
 #include "llvm/IR/Dominators.h"
 #include "llvm/IR/Module.h"
+#include "llvm/IR/ProfDataUtils.h"
 #include "llvm/Support/SourceMgr.h"
 #include "gtest/gtest.h"
 
@@ -195,3 +196,83 @@ TEST(LoopUtils, nestedLoopSharedLatchEstimatedTripCount) {
         EXPECT_EQ(getLoopEstimatedTripCount(Outer), std::nullopt);
       });
 }
+
+// {get,set}LoopEstimatedTripCount implement special handling of zero.
+TEST(LoopUtils, zeroEstimatedTripCount) {
+  LLVMContext C;
+  const char *IR =
+      "define void @foo(i1 %c) {\n"
+      "entry:\n"
+      "  br label %loop0\n"
+      "loop0:\n"
+      "  br i1 %c, label %loop0, label %loop1\n"
+      "loop1:\n"
+      "  br i1 %c, label %loop1, label %loop2, !llvm.loop !1\n"
+      "loop2:\n"
+      "  br i1 %c, label %loop2, label %exit, !prof !5, !llvm.loop !2\n"
+      "exit:\n"
+      "  ret void\n"
+      "}\n"
+      "!1 = distinct !{!1, !3}\n"
+      "!2 = distinct !{!2, !3, !4}\n"
+      "!3 = !{!\"foo\", i32 5}\n"
+      "!4 = !{!\"llvm.loop.estimated_trip_count\", i32 10}\n"
+      "!5 = !{!\"branch_weights\", i32 1, i32 9}\n"
+      "\n";
+
+  // With EstimatedLoopInvocationWeight, setLoopEstimatedTripCount sets branch
+  // weights and llvm.loop.estimated_trip_count all to 0, so
+  // getLoopEstimatedTripCount returns std::nullopt.  It does not touch other
+  // loop metadata, if any.
+  std::unique_ptr<Module> M = parseIR(C, IR);
+  run(*M, "foo",
+      [&](Function &F, DominatorTree &DT, ScalarEvolution &SE, LoopInfo &LI) {
+        assert(LI.end() - LI.begin() == 3 && "Expected three loops");
+        for (Loop *L : LI) {
+          Instruction &LatchBranch = *L->getLoopLatch()->getTerminator();
+          std::optional<int> Foo = getOptionalIntLoopAttribute(L, "foo");
+
+          EXPECT_EQ(setLoopEstimatedTripCount(
+                        L, 0, /*EstimatedLoopInvocationWeight=*/1),
+                    true);
+
+          SmallVector<uint32_t, 2> Weights;
+          EXPECT_EQ(extractBranchWeights(LatchBranch, Weights), true);
+          EXPECT_EQ(Weights[0], 0u);
+          EXPECT_EQ(Weights[1], 0u);
+          EXPECT_EQ(getOptionalIntLoopAttribute(L, "foo"), Foo);
+          EXPECT_EQ(getOptionalIntLoopAttribute(L, LLVMLoopEstimatedTripCount),
+                    0);
+          EXPECT_EQ(getLoopEstimatedTripCount(L), std::nullopt);
+        }
+      });
+
+  // Without EstimatedLoopInvocationWeight, setLoopEstimatedTripCount sets
+  // llvm.loop.estimated_trip_count to 0, so getLoopEstimatedTripCount returns
+  // std::nullopt.  It does not touch branch weights or other loop metadata, if
+  // any.
+  M = parseIR(C, IR);
+  run(*M, "foo",
+      [&](Function &F, DominatorTree &DT, ScalarEvolution &SE, LoopInfo &LI) {
+        assert(LI.end() - LI.begin() == 3 && "Expected three loops");
+        for (Loop *L : LI) {
+          Instruction &LatchBranch = *L->getLoopLatch()->getTerminator();
+          std::optional<int> Foo = getOptionalIntLoopAttribute(L, "foo");
+          SmallVector<uint32_t, 2> WeightsOld;
+          bool HasWeights = extractBranchWeights(LatchBranch, WeightsOld);
+
+          EXPECT_EQ(setLoopEstimatedTripCount(L, 0), true);
+
+          SmallVector<uint32_t, 2> WeightsNew;
+          EXPECT_EQ(extractBranchWeights(LatchBranch, WeightsNew), HasWeights);
+          if (HasWeights) {
+            EXPECT_EQ(WeightsNew[0], WeightsOld[0]);
+            EXPECT_EQ(WeightsNew[1], WeightsOld[1]);
+          }
+          EXPECT_EQ(getOptionalIntLoopAttribute(L, "foo"), Foo);
+          EXPECT_EQ(getOptionalIntLoopAttribute(L, LLVMLoopEstimatedTripCount),
+                    0);
+          EXPECT_EQ(getLoopEstimatedTripCount(L), std::nullopt);
+        }
+      });
+}