[llvm] [Transforms] LoopIdiomRecognize recognize strlen and wcslen (PR #108985)

[via llvm-commits]

================
@@ -1524,6 +1545,232 @@ static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry,
   return nullptr;
 }
 
+/// Recognizes a strlen idiom by checking for loops that increment
+/// a char pointer and then subtract with the base pointer.
+///
+/// If detected, transforms the relevant code to a strlen function
+/// call, and returns true; otherwise, returns false.
+///
+/// The core idiom we are trying to detect is:
+/// \code
+///     start = str;
+///     do {
+///       str++;
+///     } while(*str != '\0');
+/// \endcode
+///
+/// The transformed output is similar to below c-code:
+/// \code
+///     str = start + strlen(start)
+///     len = str - start
+/// \endcode
+///
+/// Later the pointer subtraction will be folded by InstCombine
+bool LoopIdiomRecognize::recognizeAndInsertStrLen() {
+  if (DisableLIRPStrlen)
+    return false;
+
+  // Give up if the loop has multiple blocks or multiple backedges.
+  if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
+    return false;
+
+  // It should have a preheader containing nothing but an unconditional branch.
+  auto *Preheader = CurLoop->getLoopPreheader();
+  if (!Preheader || &Preheader->front() != Preheader->getTerminator())
+    return false;
----------------
RolandF77 wrote:

This condition seems unnecessarily restrictive. For example, it gets the idiom in check1 but not check2:

#include <stdio.h>
#include <string.h>

void check1(char *s) {
  char *p = s;
  do {
    p++;
  } while (*p);
  printf("%s %ld\n", s, strlen(s));
  printf ("%ld\n", p - s);
}

void check2(char *s) {
  printf("%s %ld\n", s, strlen(s));
  char *p = s;
  do {
    p++;
  } while (*p);
  printf ("%ld\n", p - s);
}

int main(int argc, char **argv) {
  check1(argv[0]);
  check2(argv[0]);
  return 0;
}


https://github.com/llvm/llvm-project/pull/108985