[libc++] Remove temporary stack-allocated array pointer

[ddcc]

Summary

Manual inspection of generated IR shows temporary stack allocation

[curdeius]

Do you have an example where this occurs (e.g. godbolt link)?

[ddcc]

Hmm, I thought mem2reg was supposed to run even at O0. Turns out that's not the case, which explains why the alloca [3 x i64]* was not removed.

[ldionne]

libcxx/include/string

1561

Actually, that's a reference to an array. So I don't think we should be creating any temporary here.

[ddcc]

Here is the IR output for a small example, in the _ZNSt3__112basic_stringIcNS_11char_traitsIcEENS_9allocatorIcEEE6__zeroEv function: https://godbolt.org/z/174Pav

At O0, the reference in __a is lowered as a stack-allocated temporary array pointer, which stores the value of __r_.first().__r.__words. Then, it is loaded back out from __a in every iteration of the loop body and combined with __i in a GEP expression.

This change does remove the temporary from being emitted at O0, although the mem2reg pass does eliminate it at higher optimization levels.

[ldionne]

In D96966#2572873, @ddcc wrote:

Here is the IR output for a small example, in the _ZNSt3__112basic_stringIcNS_11char_traitsIcEENS_9allocatorIcEEE6__zeroEv function: https://godbolt.org/z/174Pav

At O0, the reference in __a is lowered as a stack-allocated temporary array pointer, which stores the value of __r_.first().__r.__words. Then, it is loaded back out from __a in every iteration of the loop body and combined with __i in a GEP expression.

This change does remove the temporary from being emitted at O0, although the mem2reg pass does eliminate it at higher optimization levels.

Oh, thanks for the clarification. So the temporary was the pointer to the array. I thought we were trying to eliminate a stack allocated array. Anyway, it looks like this is not relevant anyway. Thank you!

libcxx/include/string

pointer __get_pointer() _NOEXCEPT
         {return __is_long() ? __get_long_pointer() : __get_short_pointer();}
     _LIBCPP_INLINE_VISIBILITY
     const_pointer __get_pointer() const _NOEXCEPT
         {return __is_long() ? __get_long_pointer() : __get_short_pointer();}
 
     _LIBCPP_INLINE_VISIBILITY
     void __zero() _NOEXCEPT
         {
-            size_type (&__a)[__n_words] = __r_.first().__r.__words;
             for (unsigned __i = 0; __i < __n_words; ++__i)
-                __a[__i] = 0;
+                __r_.first().__r.__words[__i] = 0;
         }
 
     template <size_type __a> static
         _LIBCPP_INLINE_VISIBILITY
         size_type __align_it(size_type __s) _NOEXCEPT
             {return (__s + (__a-1)) & ~(__a-1);}
     enum {__alignment = 16};
     static _LIBCPP_INLINE_VISIBILITY