Cannot constant fold functions from <cwchar>

[philnik777]

Clang has builtins for a lot of the <cwchar> functions and can constant evaluate them when using the __builtin_ variants, but the backend can't constant fold them. There doesn't seem to be much of a technical reason, since their <cstring> counterparts can often be constant folded, and some even have passes dedicated to them.

For example,

#include <cstddef>

auto test(wchar_t *ptr, size_t count) {
  wchar_t buffer[] = L"Banane";
  return __builtin_wmemchr(buffer, L'a', __builtin_wcslen(buffer)) != nullptr;
}

static_assert([] {
  wchar_t buffer[] = L"Banane";
  return __builtin_wmemchr(buffer, L'a', __builtin_wcslen(buffer)) != nullptr;
}());

compiles just fine, but test generates horrible code compared to what should be possible: https://godbolt.org/z/xP3vM9jPn

[AaronBallman]

Given how much code on Windows uses wchar_t, improving this would probably have some nice performance wins in practice.

[nikic]

From a quick look, the only wchar_t function which is optimized right now is wcslen:

llvm-project/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

Lines 1072 to 1080 in beb12f9

	Value *LibCallSimplifier::optimizeWcslen(CallInst *CI, IRBuilderBase &B) {
	Module &M = *CI->getModule();
	unsigned WCharSize = TLI->getWCharSize(M) * 8;
	// We cannot perform this optimization without wchar_size metadata.
	if (WCharSize == 0)
	return nullptr;
	return optimizeStringLength(CI, B, WCharSize);
	}

[nikic]

@s-barannikov It occurs to me that this may also be a good way to phase in support for non-8-bit bytes in SLC. If we support wchar functions, which take 16 or 32 bit characters, then supporting 16/32-bit char would be a trivial change on top of that.

[s-barannikov]

The main issue with constant folding wcs* functions is that it is often required to examine contents of the string and we don't have a data structure for this. For char we use StringRef assuming that host and target chars have the same bitwidth, see getConstantStringInfo.

To support folding wcs* functions we need this function to return a view with arbitrary-sized elements (APStringRef / TargetStringRef / TStringRef?). This data structure can be used for supporting folding the ordinary str* functions in the case of CHAR_BIT != 8. I implemented it downstream but it wasn't good enough to put it for review. In #106541 / #106542 I just bail out when it comes to examining the contents of a string.

... then supporting 16/32-bit char would be a trivial change on top of that.

Supporting wchar_t needs more work in SimplifyLibCalls.cpp than supporting non-8-bit chars: we need to add wide char functions to TLI so that we can fold e.g., wcschr -> wmemchr. Look for calls to various emit* functions in this file.