[clang] [Clang] Allow the value of unroll count to be zero in `#pragma GCC unroll` and `#pragma unroll` (PR #88666)

[via cfe-commits]

yronglin wrote:

> There's one more questionable effect of this commit (together with the fix in #89567). Some of our code started triggering "loop not unrolled" warnings. A reduced example is at https://godbolt.org/z/jG4xsGY7s
> 
> ```
> #include <cstddef>
> #include <bit>
> 
> template <bool Flag>
> int FailToBuild(int n) {
>   constexpr int N = 100;
>   auto init = [=]() { return Flag ? n : 0UL; };
>   auto cond = [=](size_t ix) { return Flag ? ix != 0 : ix < 10; };
>   auto iter = [=](size_t ix) {
>     return Flag ? ix & ~(1ULL << std::countr_zero(ix)) : ix + 1;
>   };
> #pragma unroll Flag ? 1 : N
>   for (size_t ix = init(); cond(ix); ix = iter(ix)) {
>     n *= n;
>   }
>   return n;
> }
> 
> int foo(int n) {
>     return FailToBuild<true>(n);
> }
> 
> int bar(int n) {
>     return FailToBuild<false>(n);
> }
> ```
> 
> `clang -std=c++20 -O3` started generating the following warning on this code after the commit:
> 
> ```
> <source>:13:3: warning: loop not unrolled: the optimizer was unable to perform the requested transformation; the transformation might be disabled or specified as part of an unsupported transformation ordering [-Wpass-failed=transform-warning]
>    13 |   for (size_t ix = init(); cond(ix); ix = iter(ix)) {
> ```
> 
> It doesn't seem the generated code has actually changed, thus, the question is whether these are legit warnings that used to be missed, or vice versa?



> There's one more questionable effect of this commit (together with the fix in #89567). Some of our code started triggering "loop not unrolled" warnings. A reduced example is at https://godbolt.org/z/jG4xsGY7s
> 
> ```
> #include <cstddef>
> #include <bit>
> 
> template <bool Flag>
> int FailToBuild(int n) {
>   constexpr int N = 100;
>   auto init = [=]() { return Flag ? n : 0UL; };
>   auto cond = [=](size_t ix) { return Flag ? ix != 0 : ix < 10; };
>   auto iter = [=](size_t ix) {
>     return Flag ? ix & ~(1ULL << std::countr_zero(ix)) : ix + 1;
>   };
> #pragma unroll Flag ? 1 : N
>   for (size_t ix = init(); cond(ix); ix = iter(ix)) {
>     n *= n;
>   }
>   return n;
> }
> 
> int foo(int n) {
>     return FailToBuild<true>(n);
> }
> 
> int bar(int n) {
>     return FailToBuild<false>(n);
> }
> ```
> 
> `clang -std=c++20 -O3` started generating the following warning on this code after the commit:
> 
> ```
> <source>:13:3: warning: loop not unrolled: the optimizer was unable to perform the requested transformation; the transformation might be disabled or specified as part of an unsupported transformation ordering [-Wpass-failed=transform-warning]
>    13 |   for (size_t ix = init(); cond(ix); ix = iter(ix)) {
> ```
> 
> It doesn't seem the generated code has actually changed, thus, the question is whether these are legit warnings that used to be missed, or vice versa?

Thank you for your report, I will find out the cause as soon as possible.

https://github.com/llvm/llvm-project/pull/88666