[llvm] [LoopRotation] Enable LoopRotation with -Oz if header folds (PR #72842)

[via llvm-commits]

https://github.com/mohammed-nurulhoque updated https://github.com/llvm/llvm-project/pull/72842

>From e1a6994c16e0fad76e1b38eeef834bdeb4a31fce Mon Sep 17 00:00:00 2001
From: Mohammed Nurul Hoque <Mohammed.NurulHoque at imgtec.com>
Date: Thu, 1 Jul 2021 09:32:43 +0100
Subject: [PATCH] [LoopRotation] Enable LoopRotation with -Oz if header folds

LoopRotation duplicates the header and can increase codesize. If
header size is more than a threshold, rotation is disabled. But
sometimes, the duplicated header folds, thus not increasing size.
In that case, rotate the header even if > threshold header size.
---
 .../Transforms/Utils/LoopRotationUtils.cpp    | 76 +++++++++++++++++--
 .../LoopRotate/rotate-oz-if-branch-fold.ll    | 32 ++++++++
 2 files changed, 101 insertions(+), 7 deletions(-)
 create mode 100644 llvm/test/Transforms/LoopRotate/rotate-oz-if-branch-fold.ll

diff --git a/llvm/lib/Transforms/Utils/LoopRotationUtils.cpp b/llvm/lib/Transforms/Utils/LoopRotationUtils.cpp
index 012aa5dbb9ca004..75ec5e8dae7b739 100644
--- a/llvm/lib/Transforms/Utils/LoopRotationUtils.cpp
+++ b/llvm/lib/Transforms/Utils/LoopRotationUtils.cpp
@@ -365,6 +365,65 @@ static void updateBranchWeights(BranchInst &PreHeaderBI, BranchInst &LoopBI,
   }
 }
 
+// Check if duplicating the header to the preheader will fold all duplicated
+// instructions, terminating with an unconditional jump. So No code duplication
+// would be necessary.
+static bool canRotateWithoutDuplication(Loop *L, llvm::LoopInfo *LI,
+                                        llvm::SimplifyQuery SQ) {
+  BasicBlock *OrigHeader = L->getHeader();
+  BasicBlock *OrigPreheader = L->getLoopPreheader();
+
+  BranchInst *T = dyn_cast<BranchInst>(OrigHeader->getTerminator());
+  assert(T && T->isConditional() && "Header Terminator should be conditional!");
+
+  // a value map to holds the values incoming from preheader
+  ValueToValueMapTy ValueMap;
+
+  // iterate over non-debug instructions and check which ones fold
+  for (Instruction &Inst : OrigHeader->instructionsWithoutDebug()) {
+    // For PHI nodes, the values in the cloned header are the values in the
+    // preheader
+    if (PHINode *PN = dyn_cast<PHINode>(&Inst)) {
+      InsertNewValueIntoMap(ValueMap, PN,
+                            PN->getIncomingValueForBlock(OrigPreheader));
+      continue;
+    }
+    if (Inst.mayHaveSideEffects())
+      return false;
+
+    Instruction *C = Inst.clone();
+    C->insertBefore(&Inst);
+
+    // Eagerly remap the operands of the instruction.
+    RemapInstruction(C, ValueMap,
+                     RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
+
+    // if we hit the terminator, return whether it folds
+    if (static_cast<BranchInst *>(&Inst) == T) {
+      if (isa<ConstantInt>(dyn_cast<BranchInst>(C)->getCondition())) {
+        LLVM_DEBUG(dbgs() << "Loop Rotation: bypassing header threshold "
+                             "because the header folds\n");
+        C->eraseFromParent();
+        return true;
+      } else {
+        LLVM_DEBUG(
+            dbgs() << "Loop Rotation: skipping. Header threshold exceeded "
+                      "and header doesn't fold\n");
+        C->eraseFromParent();
+        return false;
+      }
+    }
+
+    // other instructions, evaluate if possible
+    Value *V = simplifyInstruction(C, SQ);
+    if (V)
+      InsertNewValueIntoMap(ValueMap, &Inst, V);
+    C->eraseFromParent();
+  }
+
+  llvm_unreachable("OrigHeader didn't contain terminal branch!");
+}
+
 /// Rotate loop LP. Return true if the loop is rotated.
 ///
 /// \param SimplifiedLatch is true if the latch was just folded into the final
@@ -438,13 +497,16 @@ bool LoopRotate::rotateLoop(Loop *L, bool SimplifiedLatch) {
         return Rotated;
       }
       if (Metrics.NumInsts > MaxHeaderSize) {
-        LLVM_DEBUG(dbgs() << "LoopRotation: NOT rotating - contains "
-                          << Metrics.NumInsts
-                          << " instructions, which is more than the threshold ("
-                          << MaxHeaderSize << " instructions): ";
-                   L->dump());
-        ++NumNotRotatedDueToHeaderSize;
-        return Rotated;
+        if (!canRotateWithoutDuplication(L, LI, SQ)) {
+          LLVM_DEBUG(dbgs()
+                         << "LoopRotation: NOT rotating - contains "
+                         << Metrics.NumInsts
+                         << " instructions, which is more than the threshold ("
+                         << MaxHeaderSize << " instructions): ";
+                     L->dump());
+          ++NumNotRotatedDueToHeaderSize;
+          return Rotated;
+        }
       }
 
       // When preparing for LTO, avoid rotating loops with calls that could be
diff --git a/llvm/test/Transforms/LoopRotate/rotate-oz-if-branch-fold.ll b/llvm/test/Transforms/LoopRotate/rotate-oz-if-branch-fold.ll
new file mode 100644
index 000000000000000..cca386ebc495b23
--- /dev/null
+++ b/llvm/test/Transforms/LoopRotate/rotate-oz-if-branch-fold.ll
@@ -0,0 +1,32 @@
+; RUN: opt < %s -S -Oz -debug -debug-only=loop-rotate 2>&1 | FileCheck %s -check-prefix=OZ
+
+; Loop should be rotated for -Oz if the duplicated branch in the header
+; can be folded.
+; Test adapted from gcc-c-torture/execute/pr90949.c
+
+; OZ: bypassing header threshold
+; OZ: rotating Loop at depth 1 containing: %tailrecurse.i<header><exiting>,%tailrecurse.backedge.i<latch>
+; OZ: into Loop at depth 1 containing: %tailrecurse.backedge.i<header><latch><exiting>
+
+%struct.Node = type { %struct.Node* }
+
+ at space = dso_local global [2 x %struct.Node] zeroinitializer, align 8
+
+define dso_local signext i32 @main() local_unnamed_addr #2 {
+entry:
+  store %struct.Node* null, %struct.Node** getelementptr inbounds ([2 x %struct.Node], [2 x %struct.Node]* @space, i64 0, i64 0, i32 0), align 8
+  br label %tailrecurse.i
+
+tailrecurse.i:                                    ; preds = %tailrecurse.backedge.i, %entry
+  %module.tr.i = phi %struct.Node* [ getelementptr inbounds ([2 x %struct.Node], [2 x %struct.Node]* @space, i64 0, i64 0), %entry ], [ %module.tr.be.i, %tailrecurse.backedge.i ]
+  %cmp.i = icmp eq %struct.Node* %module.tr.i, null
+  br i1 %cmp.i, label %walk.exit, label %tailrecurse.backedge.i
+
+tailrecurse.backedge.i:                           ; preds = %tailrecurse.i
+  %module.tr.be.in.i = getelementptr inbounds %struct.Node, %struct.Node* %module.tr.i, i64 0, i32 0
+  %module.tr.be.i = load %struct.Node*, %struct.Node** %module.tr.be.in.i, align 8
+  br label %tailrecurse.i
+
+walk.exit:                                        ; preds = %tailrecurse.i
+  ret i32 0
+}