[LoopPeel] Fix BFI when peeling last iteration without guard

llvm/include/llvm/Support/BranchProbability.h

@@ -57,6 +57,8 @@ class BranchProbability {
   // Create a BranchProbability object from 64-bit integers.
   LLVM_ABI static BranchProbability getBranchProbability(uint64_t Numerator,
                                                          uint64_t Denominator);
+  // Create a BranchProbability from a double, which must be from 0 to 1.
+  LLVM_ABI static BranchProbability getBranchProbability(double Prob);
 
   // Normalize given probabilties so that the sum of them becomes approximate
   // one.
@@ -71,6 +73,7 @@ class BranchProbability {
 
   uint32_t getNumerator() const { return N; }
   static uint32_t getDenominator() { return D; }
+  double toDouble() const { return double(N) / D; }
 
   // Return (1 - Probability).
   BranchProbability getCompl() const { return BranchProbability(D - N); }

llvm/lib/Support/BranchProbability.cpp

@@ -60,6 +60,11 @@ BranchProbability::getBranchProbability(uint64_t Numerator,
   return BranchProbability(Numerator >> Scale, Denominator);
 }
 
+BranchProbability BranchProbability::getBranchProbability(double Prob) {
+  assert(0 <= Prob && Prob <= 1 && "Probability must be between 0 and 1!");
+  return BranchProbability(std::round(Prob * D), D);
+}
+
 // If ConstD is not zero, then replace D by ConstD so that division and modulo
 // operations by D can be optimized, in case this function is not inlined by the
 // compiler.

llvm/lib/Transforms/Utils/LoopPeel.cpp

@@ -1183,8 +1183,33 @@ bool llvm::peelLoop(Loop *L, unsigned PeelCount, bool PeelLast, LoopInfo *LI,
 
     // If the original loop may only execute a single iteration we need to
     // insert a trip count check and skip the original loop with the last
-    // iteration peeled off if necessary.
-    if (!SE->isKnownNonZero(BTC)) {
+    // iteration peeled off if necessary.  Either way, we must update branch
+    // weights to maintain the loop body frequency.
+    if (SE->isKnownNonZero(BTC)) {
+      // We have just proven that, when reached, the original loop always
+      // executes at least two iterations.  Thus, we unconditionally execute
+      // both the remaining loop's initial iteration and the peeled iteration.
+      // But that increases the latter's frequency above its frequency in the
+      // original loop.  To maintain the total frequency, we compensate by
+      // decreasing the remaining loop body's frequency to indicate one less
+      // iteration.
+      //
+      // We use this formula to convert probability to/from frequency:
+      // Sum(i=0..inf)(P^i) = 1/(1-P) = Freq.
+      if (BranchProbability P = getLoopProbability(L); !P.isUnknown()) {
+        // Trying to subtract one from an infinite loop is pointless, and our
+        // formulas then produce division by zero, so skip that case.
+        if (BranchProbability ExitP = P.getCompl(); !ExitP.isZero()) {
+          double Freq = 1 / ExitP.toDouble();
+          // No branch weights can produce a frequency of less than one given
+          // the initial iteration, and our formulas produce a negative
+          // probability if we try.
+          double NewFreq = std::max(Freq - 1, 1.0);
+          setLoopProbability(
+              L, BranchProbability::getBranchProbability(1 - 1 / NewFreq));
+        }
+      }
+    } else {
       NewPreHeader = SplitEdge(PreHeader, Header, &DT, LI);
       SCEVExpander Expander(*SE, Latch->getDataLayout(), "loop-peel");
 

llvm/lib/Transforms/Utils/LoopUnrollRuntime.cpp

@@ -495,16 +495,13 @@ static Loop *CloneLoopBlocks(Loop *L, Value *NewIter,
           BranchProbability ProbReaching = BranchProbability::getOne();
           for (unsigned N = Count - 2; N >= 1; --N) {
             ProbReaching *= probOfNextInRemainder(OriginalLoopProb, N);
-            FreqRemIters += double(ProbReaching.getNumerator()) /
-                            ProbReaching.getDenominator();
+            FreqRemIters += ProbReaching.toDouble();
           }
         }
         // Solve for the loop probability that would produce that frequency.
         // Sum(i=0..inf)(Prob^i) = 1/(1-Prob) = FreqRemIters.
-        double ProbDouble = 1 - 1 / FreqRemIters;
-        BranchProbability Prob = BranchProbability::getBranchProbability(
-            std::round(ProbDouble * BranchProbability::getDenominator()),
-            BranchProbability::getDenominator());
+        BranchProbability Prob =
+            BranchProbability::getBranchProbability(1 - 1 / FreqRemIters);
         setBranchProbability(RemainderLoopLatch, Prob, /*ForFirstTarget=*/true);
       }
       NewIdx->addIncoming(Zero, InsertTop);

llvm/test/Transforms/LoopUnroll/branch-weights-freq/peel-last-iteration-no-guard.ll

@@ -0,0 +1,87 @@
+; Check that the loop body frequency is maintained when LoopPeel both:
+; - Peels one unconditional iteration after the loop.
+; - Does not add a guard to sometimes skip the remaining loop because it has
+;   proven the original loop always executes at least two iterations, which
+;   become the initial iteration and the peeled iteration.
+
+; DEFINE: %{exitWeight} =
+; DEFINE: %{loopWeight} =
+; DEFINE: %{loopFreqOld} =
+; DEFINE: %{loopFreqNew} =
+
+; DEFINE: %{run} = \
+; DEFINE:   cp %s %t.ll && chmod +w %t.ll && \
+; DEFINE:   echo '!0 = !{!"branch_weights", i32 %{exitWeight}, ' \
+; DEFINE:       'i32 %{loopWeight}}' >> %t.ll && \
+; DEFINE:   opt -p "print<block-freq>,loop-unroll,print<block-freq>" \
+; DEFINE:       -unroll-full-max-count=0 -S %t.ll 2>&1 | \
+; DEFINE:     FileCheck -DLOOP_FREQ_OLD='%{loopFreqOld}' \
+; DEFINE:         -DLOOP_FREQ_NEW='%{loopFreqNew}' %s
+
+; Branch weights give the original loop 10 iterations.  We expect that
+; loopFreqOld = loopFreqNew + 1.
+; REDEFINE: %{exitWeight} = 1
+; REDEFINE: %{loopWeight} = 9
+; REDEFINE: %{loopFreqOld} = 10.0
+; REDEFINE: %{loopFreqNew} = 9.0
+; RUN: %{run}
+
+; Branch weights give the original loop 2 iterations.  We expect that
+; loopFreqOld = loopFreqNew + 1.
+; REDEFINE: %{exitWeight} = 1
+; REDEFINE: %{loopWeight} = 1
+; REDEFINE: %{loopFreqOld} = 2.0
+; REDEFINE: %{loopFreqNew} = 1.0
+; RUN: %{run}
+
+; Branch weights give the original loop 1 iteration, but LoopPeel proved it has
+; at least 2.  There is no loop probability that produces a frequency below 1,
+; so the original total frequency cannot be maintained.
+; REDEFINE: %{exitWeight} = 1
+; REDEFINE: %{loopWeight} = 0
+; REDEFINE: %{loopFreqOld} = 1.0
+; REDEFINE: %{loopFreqNew} = 1.0
+; RUN: %{run}
+
+; Branch weights say the original loop is infinite, maximizing the frequency,
+; so LoopPeel does not try to decrement it.
+; REDEFINE: %{exitWeight} = 0
+; REDEFINE: %{loopWeight} = 1
+; REDEFINE: %{loopFreqOld} = 2147483647.8
+; REDEFINE: %{loopFreqNew} = 2147483647.8
+; RUN: %{run}
+
+; Everything other than loop should be 1.0 because it is reached once.
+;
+;      CHECK: block-frequency-info: test
+; CHECK-NEXT: - entry: float = 1.0,
+; CHECK-NEXT: - loop: float = [[LOOP_FREQ_OLD]],
+; CHECK-NEXT: - exit: float = 1.0,
+;
+;      CHECK: block-frequency-info: test
+; CHECK-NEXT: - entry: float = 1.0,
+; CHECK-NEXT: - loop: float = [[LOOP_FREQ_NEW]],
+; CHECK-NEXT: - exit.peel.begin: float = 1.0,
+; CHECK-NEXT: - loop.peel: float = 1.0,
+; CHECK-NEXT: - exit.peel.next: float = 1.0,
+; CHECK-NEXT: - loop.peel.next: float = 1.0,
+; CHECK-NEXT: - exit: float = 1.0,
+
+declare void @f(i32)
+
+define void @test() {
+entry:
+  br label %loop
+
+loop:
+  %i = phi i32 [ 0, %entry ], [ %inc, %loop ]
+  %isLast = icmp eq i32 %i, 20
+  %sel = select i1 %isLast, i32 1, i32 0
+  call void @f(i32 %sel)
+  %inc = add i32 %i, 1
+  %isLast1 = icmp eq i32 %i, 20
+  br i1 %isLast1, label %exit, label %loop, !prof !0
+
+exit:
+  ret void
+}