Reland "[Transforms] LoopIdiomRecognize recognize strlen and wcslen #108985"

[mustartt]

Reland #108985

In this patch, also add strlen and wcslen to the common idiom names list so that functions named strlen or wcslen does not get transformed.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Henry Jiang (mustartt)

Changes

Reland #108985

In this patch, also add strlen and wcslen to the common idiom names list so that functions named strlen or wcslen does not get transformed.

---

Patch is 61.34 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/132572.diff

9 Files Affected:

• (modified) llvm/include/llvm/Transforms/Scalar/LoopIdiomRecognize.h (+6)
• (modified) llvm/include/llvm/Transforms/Utils/BuildLibCalls.h (+6)
• (modified) llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp (+293-5)
• (modified) llvm/lib/Transforms/Utils/BuildLibCalls.cpp (+9)
• (added) llvm/test/Transforms/LoopIdiom/strlen-noidiom.ll (+64)
• (added) llvm/test/Transforms/LoopIdiom/strlen-not-emittable.ll (+66)
• (added) llvm/test/Transforms/LoopIdiom/strlen.ll (+614)
• (added) llvm/test/Transforms/LoopIdiom/wcslen16.ll (+126)
• (added) llvm/test/Transforms/LoopIdiom/wcslen32.ll (+134)

diff --git a/llvm/include/llvm/Transforms/Scalar/LoopIdiomRecognize.h b/llvm/include/llvm/Transforms/Scalar/LoopIdiomRecognize.h
index 0c6406d861851..241a3fc109360 100644
--- a/llvm/include/llvm/Transforms/Scalar/LoopIdiomRecognize.h
+++ b/llvm/include/llvm/Transforms/Scalar/LoopIdiomRecognize.h
@@ -34,6 +34,12 @@ struct DisableLIRP {
 
   /// When true, Memcpy is disabled.
   static bool Memcpy;
+
+  /// When true, Strlen is disabled.
+  static bool Strlen;
+
+  /// When true, Wcslen is disabled.
+  static bool Wcslen;
 };
 
 /// Performs Loop Idiom Recognize Pass.
diff --git a/llvm/include/llvm/Transforms/Utils/BuildLibCalls.h b/llvm/include/llvm/Transforms/Utils/BuildLibCalls.h
index a8fb38e726004..50f695dbe6c07 100644
--- a/llvm/include/llvm/Transforms/Utils/BuildLibCalls.h
+++ b/llvm/include/llvm/Transforms/Utils/BuildLibCalls.h
@@ -93,6 +93,12 @@ namespace llvm {
   Value *emitStrLen(Value *Ptr, IRBuilderBase &B, const DataLayout &DL,
                     const TargetLibraryInfo *TLI);
 
+  /// Emit a call to the wcslen function to the builder, for the specified
+  /// pointer. Ptr is required to be some pointer type, and the return value has
+  /// 'size_t' type.
+  Value *emitWcsLen(Value *Ptr, IRBuilderBase &B, const DataLayout &DL,
+                    const TargetLibraryInfo *TLI);
+
   /// Emit a call to the strdup function to the builder, for the specified
   /// pointer. Ptr is required to be some pointer type, and the return value has
   /// 'i8*' type.
diff --git a/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp b/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
index 8dc7f325f6262..4940dd8c47dde 100644
--- a/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
+++ b/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
@@ -20,8 +20,7 @@
 //
 // TODO List:
 //
-// Future loop memory idioms to recognize:
-//   memcmp, strlen, etc.
+// Future loop memory idioms to recognize: memcmp, etc.
 //
 // This could recognize common matrix multiplies and dot product idioms and
 // replace them with calls to BLAS (if linked in??).
@@ -33,6 +32,7 @@
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/MapVector.h"
+#include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SetVector.h"
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallVector.h"
@@ -97,6 +97,7 @@ using namespace llvm;
 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
 STATISTIC(NumMemMove, "Number of memmove's formed from loop load+stores");
+STATISTIC(NumStrLen, "Number of strlen's and wcslen's formed from loop loads");
 STATISTIC(
     NumShiftUntilBitTest,
     "Number of uncountable loops recognized as 'shift until bitttest' idiom");
@@ -126,6 +127,22 @@ static cl::opt<bool, true>
                       cl::location(DisableLIRP::Memcpy), cl::init(false),
                       cl::ReallyHidden);
 
+bool DisableLIRP::Strlen;
+static cl::opt<bool, true>
+    DisableLIRPStrlen("disable-loop-idiom-strlen",
+                      cl::desc("Proceed with loop idiom recognize pass, but do "
+                               "not convert loop(s) to strlen."),
+                      cl::location(DisableLIRP::Strlen), cl::init(false),
+                      cl::ReallyHidden);
+
+bool DisableLIRP::Wcslen;
+static cl::opt<bool, true>
+    EnableLIRPWcslen("disable-loop-idiom-wcslen",
+                     cl::desc("Proceed with loop idiom recognize pass, "
+                              "enable conversion of loop(s) to wcslen."),
+                     cl::location(DisableLIRP::Wcslen), cl::init(false),
+                     cl::ReallyHidden);
+
 static cl::opt<bool> UseLIRCodeSizeHeurs(
     "use-lir-code-size-heurs",
     cl::desc("Use loop idiom recognition code size heuristics when compiling "
@@ -246,6 +263,7 @@ class LoopIdiomRecognize {
 
   bool recognizeShiftUntilBitTest();
   bool recognizeShiftUntilZero();
+  bool recognizeAndInsertStrLen();
 
   /// @}
 };
@@ -295,7 +313,8 @@ bool LoopIdiomRecognize::runOnLoop(Loop *L) {
 
   // Disable loop idiom recognition if the function's name is a common idiom.
   StringRef Name = L->getHeader()->getParent()->getName();
-  if (Name == "memset" || Name == "memcpy")
+  if (Name == "memset" || Name == "memcpy" || Name == "strlen" ||
+      Name == "wcslen")
     return false;
 
   // Determine if code size heuristics need to be applied.
@@ -1494,7 +1513,7 @@ bool LoopIdiomRecognize::runOnNoncountableLoop() {
 
   return recognizePopcount() || recognizeAndInsertFFS() ||
          recognizeShiftUntilBitTest() || recognizeShiftUntilZero() ||
-         recognizeShiftUntilLessThan();
+         recognizeShiftUntilLessThan() || recognizeAndInsertStrLen();
 }
 
 /// Check if the given conditional branch is based on the comparison between
@@ -1512,7 +1531,7 @@ static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry,
   if (!Cond)
     return nullptr;
 
-  ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
+  auto *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
   if (!CmpZero || !CmpZero->isZero())
     return nullptr;
 
@@ -1529,6 +1548,275 @@ static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry,
   return nullptr;
 }
 
+namespace {
+
+class StrlenVerifier {
+public:
+  explicit StrlenVerifier(const Loop *CurLoop, ScalarEvolution *SE,
+                          const TargetLibraryInfo *TLI)
+      : CurLoop(CurLoop), SE(SE), TLI(TLI) {}
+
+  bool isValidStrlenIdiom() {
+    // Give up if the loop has multiple blocks, multiple backedges, or
+    // multiple exit blocks
+    if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1 ||
+        !CurLoop->getUniqueExitBlock())
+      return false;
+
+    // It should have a preheader and a branch instruction.
+    BasicBlock *Preheader = CurLoop->getLoopPreheader();
+    if (!Preheader)
+      return false;
+
+    BranchInst *EntryBI = dyn_cast<BranchInst>(Preheader->getTerminator());
+    if (!EntryBI)
+      return false;
+
+    // The loop exit must be conditioned on an icmp with 0 the null terminator.
+    // The icmp operand has to be a load on some SSA reg that increments
+    // by 1 in the loop.
+    BasicBlock *LoopBody = *CurLoop->block_begin();
+
+    // Skip if the body is too big as it most likely is not a strlen idiom.
+    if (!LoopBody || LoopBody->size() >= 15)
+      return false;
+
+    BranchInst *LoopTerm = dyn_cast<BranchInst>(LoopBody->getTerminator());
+    Value *LoopCond = matchCondition(LoopTerm, LoopBody);
+    if (!LoopCond)
+      return false;
+
+    LoadInst *LoopLoad = dyn_cast<LoadInst>(LoopCond);
+    if (!LoopLoad || LoopLoad->getPointerAddressSpace() != 0)
+      return false;
+
+    OperandType = LoopLoad->getType();
+    if (!OperandType || !OperandType->isIntegerTy())
+      return false;
+
+    // See if the pointer expression is an AddRec with constant step a of form
+    // ({n,+,a}) where a is the width of the char type.
+    Value *IncPtr = LoopLoad->getPointerOperand();
+    const SCEVAddRecExpr *LoadEv =
+        dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IncPtr));
+    if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
+      return false;
+    LoadBaseEv = LoadEv->getStart();
+
+    LLVM_DEBUG(dbgs() << "pointer load scev: " << *LoadEv << "\n");
+
+    const SCEVConstant *Step =
+        dyn_cast<SCEVConstant>(LoadEv->getStepRecurrence(*SE));
+    if (!Step)
+      return false;
+
+    unsigned StepSize = 0;
+    StepSizeCI = dyn_cast<ConstantInt>(Step->getValue());
+    if (!StepSizeCI)
+      return false;
+    StepSize = StepSizeCI->getZExtValue();
+
+    // Verify that StepSize is consistent with platform char width.
+    OpWidth = OperandType->getIntegerBitWidth();
+    unsigned WcharSize = TLI->getWCharSize(*LoopLoad->getModule());
+    if (OpWidth != StepSize * 8)
+      return false;
+    if (OpWidth != 8 && OpWidth != 16 && OpWidth != 32)
+      return false;
+    if (OpWidth >= 16)
+      if (OpWidth != WcharSize * 8)
+        return false;
+
+    // Scan every instruction in the loop to ensure there are no side effects.
+    for (Instruction &I : *LoopBody)
+      if (I.mayHaveSideEffects())
+        return false;
+
+    BasicBlock *LoopExitBB = CurLoop->getExitBlock();
+    if (!LoopExitBB)
+      return false;
+
+    for (PHINode &PN : LoopExitBB->phis()) {
+      if (!SE->isSCEVable(PN.getType()))
+        return false;
+
+      const SCEV *Ev = SE->getSCEV(&PN);
+      if (!Ev)
+        return false;
+
+      LLVM_DEBUG(dbgs() << "loop exit phi scev: " << *Ev << "\n");
+
+      // Since we verified that the loop trip count will be a valid strlen
+      // idiom, we can expand all lcssa phi with {n,+,1} as (n + strlen) and use
+      // SCEVExpander materialize the loop output.
+      const SCEVAddRecExpr *AddRecEv = dyn_cast<SCEVAddRecExpr>(Ev);
+      if (!AddRecEv || !AddRecEv->isAffine())
+        return false;
+
+      // We only want RecAddExpr with recurrence step that is constant. This
+      // is good enough for all the idioms we want to recognize. Later we expand
+      // and materialize the recurrence as {base,+,a} -> (base + a * strlen)
+      if (!dyn_cast<SCEVConstant>(AddRecEv->getStepRecurrence(*SE)))
+        return false;
+    }
+
+    return true;
+  }
+
+public:
+  const Loop *CurLoop;
+  ScalarEvolution *SE;
+  const TargetLibraryInfo *TLI;
+
+  unsigned OpWidth;
+  ConstantInt *StepSizeCI;
+  const SCEV *LoadBaseEv;
+  Type *OperandType;
+};
+
+} // namespace
+
+/// The Strlen Idiom we are trying to detect has the following structure
+///
+/// preheader:
+///   ...
+///   br label %body, ...
+///
+/// body:
+///   ... ; %0 is incremented by a gep
+///   %1 = load i8, ptr %0, align 1
+///   %2 = icmp eq i8 %1, 0
+///   br i1 %2, label %exit, label %body
+///
+/// exit:
+///   %lcssa = phi [%0, %body], ...
+///
+/// We expect the strlen idiom to have a load of a character type that
+/// is compared against '\0', and such load pointer operand must have scev
+/// expression of the form {%str,+,c} where c is a ConstantInt of the
+/// appropiate character width for the idiom, and %str is the base of the string
+/// And, that all lcssa phis have the form {...,+,n} where n is a constant,
+///
+/// When transforming the output of the strlen idiom, the lccsa phi are
+/// expanded using SCEVExpander as {base scev,+,a} -> (base scev + a * strlen)
+/// and all subsequent uses are replaced. For example,
+///
+/// \code{.c}
+///     const char* base = str;
+///     while (*str != '\0')
+///         ++str;
+///     size_t result = str - base;
+/// \endcode
+///
+/// will be transformed as follows: The idiom will be replaced by a strlen
+/// computation to compute the address of the null terminator of the string.
+///
+/// \code{.c}
+///     const char* base = str;
+///     const char* end = base + strlen(str);
+///     size_t result = end - base;
+/// \endcode
+///
+/// In the case we index by an induction variable, as long as the induction
+/// variable has a constant int increment, we can replace all such indvars
+/// with the closed form computation of strlen
+///
+/// \code{.c}
+///     size_t i = 0;
+///     while (str[i] != '\0')
+///         ++i;
+///     size_t result = i;
+/// \endcode
+///
+/// Will be replaced by
+///
+/// \code{.c}
+///     size_t i = 0 + strlen(str);
+///     size_t result = i;
+/// \endcode
+///
+bool LoopIdiomRecognize::recognizeAndInsertStrLen() {
+  if (DisableLIRP::All)
+    return false;
+
+  StrlenVerifier Verifier(CurLoop, SE, TLI);
+
+  if (!Verifier.isValidStrlenIdiom())
+    return false;
+
+  BasicBlock *Preheader = CurLoop->getLoopPreheader();
+  BasicBlock *LoopExitBB = CurLoop->getExitBlock();
+
+  if (Verifier.OpWidth == 8) {
+    if (DisableLIRP::Strlen)
+      return false;
+    if (!isLibFuncEmittable(Preheader->getModule(), TLI, LibFunc_strlen))
+      return false;
+  } else {
+    if (DisableLIRP::Wcslen)
+      return false;
+    if (!isLibFuncEmittable(Preheader->getModule(), TLI, LibFunc_wcslen))
+      return false;
+  }
+
+  IRBuilder<> Builder(Preheader->getTerminator());
+  SCEVExpander Expander(*SE, Preheader->getModule()->getDataLayout(),
+                        "strlen_idiom");
+  Value *MaterialzedBase = Expander.expandCodeFor(
+      Verifier.LoadBaseEv, Verifier.LoadBaseEv->getType(),
+      Builder.GetInsertPoint());
+
+  Value *StrLenFunc = nullptr;
+  if (Verifier.OpWidth == 8) {
+    StrLenFunc = emitStrLen(MaterialzedBase, Builder, *DL, TLI);
+  } else {
+    StrLenFunc = emitWcsLen(MaterialzedBase, Builder, *DL, TLI);
+  }
+  assert(StrLenFunc && "Failed to emit strlen function.");
+
+  const SCEV *StrlenEv = SE->getSCEV(StrLenFunc);
+  SmallVector<PHINode *, 4> Cleanup;
+  for (PHINode &PN : LoopExitBB->phis()) {
+    // We can now materialize the loop output as all phi have scev {base,+,a}.
+    // We expand the phi as:
+    //   %strlen = call i64 @strlen(%str)
+    //   %phi.new = base expression + step * %strlen
+    const SCEV *Ev = SE->getSCEV(&PN);
+    const SCEVAddRecExpr *AddRecEv = dyn_cast<SCEVAddRecExpr>(Ev);
+    const SCEVConstant *Step =
+        dyn_cast<SCEVConstant>(AddRecEv->getStepRecurrence(*SE));
+    const SCEV *Base = AddRecEv->getStart();
+
+    // It is safe to truncate to base since if base is narrower than size_t
+    // the equivalent user code will have to truncate anyways.
+    const SCEV *NewEv = SE->getAddExpr(
+        Base, SE->getMulExpr(Step, SE->getTruncateOrSignExtend(
+                                       StrlenEv, Base->getType())));
+
+    Value *MaterializedPHI = Expander.expandCodeFor(NewEv, NewEv->getType(),
+                                                    Builder.GetInsertPoint());
+    Expander.clear();
+    PN.replaceAllUsesWith(MaterializedPHI);
+    Cleanup.push_back(&PN);
+  }
+
+  // All LCSSA Loop Phi are dead, the left over dead loop body can be cleaned
+  // up by later passes
+  for (PHINode *PN : Cleanup)
+    RecursivelyDeleteDeadPHINode(PN);
+  SE->forgetLoop(CurLoop);
+
+  ++NumStrLen;
+  LLVM_DEBUG(dbgs() << "  Formed strlen idiom: " << *StrLenFunc << "\n");
+  ORE.emit([&]() {
+    return OptimizationRemark(DEBUG_TYPE, "recognizeAndInsertStrLen",
+                              CurLoop->getStartLoc(), Preheader)
+           << "Transformed " << StrLenFunc->getName() << " loop idiom";
+  });
+
+  return true;
+}
+
 /// Check if the given conditional branch is based on an unsigned less-than
 /// comparison between a variable and a constant, and if the comparison is false
 /// the control yields to the loop entry. If the branch matches the behaviour,
diff --git a/llvm/lib/Transforms/Utils/BuildLibCalls.cpp b/llvm/lib/Transforms/Utils/BuildLibCalls.cpp
index 2301be6977cef..24eefc91117b4 100644
--- a/llvm/lib/Transforms/Utils/BuildLibCalls.cpp
+++ b/llvm/lib/Transforms/Utils/BuildLibCalls.cpp
@@ -1582,6 +1582,15 @@ Value *llvm::emitStrLen(Value *Ptr, IRBuilderBase &B, const DataLayout &DL,
   return emitLibCall(LibFunc_strlen, SizeTTy, CharPtrTy, Ptr, B, TLI);
 }
 
+Value *llvm::emitWcsLen(Value *Ptr, IRBuilderBase &B, const DataLayout &DL,
+                        const TargetLibraryInfo *TLI) {
+  assert(Ptr && Ptr->getType()->isPointerTy() &&
+         "Argument to wcslen intrinsic must be a pointer.");
+  Type *PtrTy = B.getPtrTy();
+  Type *SizeTTy = getSizeTTy(B, TLI);
+  return emitLibCall(LibFunc_wcslen, SizeTTy, PtrTy, Ptr, B, TLI);
+}
+
 Value *llvm::emitStrDup(Value *Ptr, IRBuilderBase &B,
                         const TargetLibraryInfo *TLI) {
   Type *CharPtrTy = B.getPtrTy();
diff --git a/llvm/test/Transforms/LoopIdiom/strlen-noidiom.ll b/llvm/test/Transforms/LoopIdiom/strlen-noidiom.ll
new file mode 100644
index 0000000000000..498ec52ed26ee
--- /dev/null
+++ b/llvm/test/Transforms/LoopIdiom/strlen-noidiom.ll
@@ -0,0 +1,64 @@
+; RUN: opt -passes='loop(loop-idiom)' < %s -S | FileCheck %s
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; CHECK-LABEL: @strlen(
+; CHECK-NOT: call{{.*}} strlen
+; CHECK-LABEL: @wcslen(
+; CHECK-NOT: call{{.*}} wcslen
+
+define i64 @strlen(ptr %str) {
+entry:
+  br label %while.cond
+
+while.cond:
+  %str.addr.0 = phi ptr [ %str, %entry ], [ %incdec.ptr, %while.cond ]
+  %0 = load i8, ptr %str.addr.0, align 1
+  %cmp.not = icmp eq i8 %0, 0
+  %incdec.ptr = getelementptr i8, ptr %str.addr.0, i64 1
+  br i1 %cmp.not, label %while.end, label %while.cond
+
+while.end:
+  %sub.ptr.lhs.cast = ptrtoint ptr %str.addr.0 to i64
+  %sub.ptr.rhs.cast = ptrtoint ptr %str to i64
+  %sub.ptr.sub = sub i64 %sub.ptr.lhs.cast, %sub.ptr.rhs.cast
+  ret i64 %sub.ptr.sub
+}
+
+define i64 @wcslen(ptr %src) {
+entry:
+  %cmp = icmp eq ptr %src, null
+  br i1 %cmp, label %return, label %lor.lhs.false
+
+lor.lhs.false:                                    ; preds = %entry
+  %0 = load i32, ptr %src, align 4
+  %cmp1 = icmp eq i32 %0, 0
+  br i1 %cmp1, label %return, label %while.cond.preheader
+
+while.cond.preheader:                             ; preds = %lor.lhs.false
+  br label %while.cond
+
+while.cond:                                       ; preds = %while.cond.preheader, %while.cond
+  %src.pn = phi ptr [ %curr.0, %while.cond ], [ %src, %while.cond.preheader ]
+  %curr.0 = getelementptr inbounds i8, ptr %src.pn, i64 4
+  %1 = load i32, ptr %curr.0, align 4
+  %tobool.not = icmp eq i32 %1, 0
+  br i1 %tobool.not, label %while.end, label %while.cond
+
+while.end:                                        ; preds = %while.cond
+  %curr.0.lcssa = phi ptr [ %curr.0, %while.cond ]
+  %sub.ptr.lhs.cast = ptrtoint ptr %curr.0.lcssa to i64
+  %sub.ptr.rhs.cast = ptrtoint ptr %src to i64
+  %sub.ptr.sub = sub i64 %sub.ptr.lhs.cast, %sub.ptr.rhs.cast
+  %sub.ptr.div = ashr exact i64 %sub.ptr.sub, 2
+  br label %return
+
+return:                                           ; preds = %entry, %lor.lhs.false, %while.end
+  %retval.0 = phi i64 [ %sub.ptr.div, %while.end ], [ 0, %lor.lhs.false ], [ 0, %entry ]
+  ret i64 %retval.0
+}
+
+
+!llvm.module.flags = !{!0}
+!0 = !{i32 1, !"wchar_size", i32 4}
diff --git a/llvm/test/Transforms/LoopIdiom/strlen-not-emittable.ll b/llvm/test/Transforms/LoopIdiom/strlen-not-emittable.ll
new file mode 100644
index 0000000000000..00fbe89846a30
--- /dev/null
+++ b/llvm/test/Transforms/LoopIdiom/strlen-not-emittable.ll
@@ -0,0 +1,66 @@
+; RUN: opt -passes='loop(loop-idiom)' < %s -S | FileCheck %s
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; invalid libcall prototype
+declare void @strlen(i32)
+declare void @wcslen(i32)
+
+define i64 @valid_wcslen32(ptr %src) {
+; CHECK-LABEL: valid_wcslen32
+; CHECK-NOT: call {{.*}} @wcslen
+entry:
+  %cmp = icmp eq ptr %src, null
+  br i1 %cmp, label %return, label %lor.lhs.false
+
+lor.lhs.false:                                    ; preds = %entry
+  %0 = load i32, ptr %src, align 4
+  %cmp1 = icmp eq i32 %0, 0
+  br i1 %cmp1, label %return, label %while.cond.preheader
+
+while.cond.preheader:                             ; preds = %lor.lhs.false
+  br label %while.cond
+
+while.cond:                                       ; preds = %while.cond.preheader, %while.cond
+  %src.pn = phi ptr [ %curr.0, %while.cond ], [ %src, %while.cond.preheader ]
+  %curr.0 = getelementptr inbounds i8, ptr %src.pn, i64 4
+  %1 = load i32, ptr %curr.0, align 4
+  %tobool.not = icmp eq i32 %1, 0
+  br i1 %tobool.not, label %while.end, label %while.cond
+
+while.end:                                        ; preds = %while.cond
+  %curr.0.lcssa = phi ptr [ %curr.0, %while.cond ]
+  %sub.ptr.lhs.cast = ptrtoint ptr %curr.0.lcssa to i64
+  %sub.ptr.rhs.cast = ptrtoint ptr %src to i64
+  %sub.ptr.sub = sub i64 %sub.ptr.lhs.cast, %sub.ptr.rhs.cast
+  %sub.ptr.div = ashr exact i64 %sub.ptr.sub, 2
+  br label %return
+
+return:                                           ; preds = %entry, %lor.lhs.false, %while.end
+  %retval.0 = phi i64 [ %sub.ptr.div, %while.end ], [ 0, %lor.lhs.false ], [ 0, %entry ]
+  ret i64 %retval.0
+}
+
+define i64 @valid_strlen(ptr %str) {
+; CHECK-LABEL: valid_strlen
+; CHECK-NOT: call {{.*}} @strlen
+entry:
+  br label %while.cond
+
+while.cond:
+  %str.addr.0 = phi ptr [ %str, %entry ], [ %incdec.ptr, %while.cond ]
+  %0 = load i8, ptr %str.addr.0, align 1
+  %cmp.not = icmp ...
[truncated]

[mstorsjo]

LGTM

[alexfh]

This doesn't work, if the corresponding function is in a namespace. But I get it's a rather niche case, which is better addressed by -fno-builtin-strlen or -ffreestanding.

[mustartt]

Yeah it does not work with mangled symbols. The original goal of this was to prevent the optimizer from calling the idiom function it self. If the idiom is in a namespace, it will still be transformed to

foo::strlen(char const*):
        jmp     strlen@PLT

which is arguably a lot better and more correct than accidentally creating an infinite tail call recursion:

strlen(char const*):
        jmp     strlen