LLVM could avoid calling memcmp when the maximum lenght is 1

[philnik777]

auto test1(const char* str, unsigned len) {
  auto cmp_len = len == 0 ? 0 : 1;
  return __builtin_memcmp("a", str, cmp_len);
}

results in a memcmp call, even though it's almost certainly faster to inline the memcmp call, especially when the place where the result is used is available.

[camc]

Hi - I would like to work on this issue.

Proposed solution:

• Detect memcmp over a constant range len, and where for each element in the range memcmp with that constant would be inlined
• In that case, replace the variable-len memcmp call with a switch over len into constant-len memcmp calls for each len in the range (in SimplifyLibCalls?)
• Then ExpandMemCmp can inline the memcmp calls

I guess this could be generalised to other lib calls.

[dtcxzyw]

It is similar to #106895.