rtld: actually resolve memcpy plt

[rlibby]

Summary

The call to memcpy() meant to cause plt resolution in _thr_rtld_init()
was getting optimized by the compiler. Tell the compiler not to use its
memcpy() builtin in thr_rtld.c.

[kib]

Should we disable all built-ins for the file, for the same reasoning?

[rlibby]

In D45891#1046235, @kib wrote:

Should we disable all built-ins for the file, for the same reasoning?

That makes sense to me, for future insurance. I checked just now that -fno-builtin didn't reveal any additional cases. Do you prefer to make that change?

[rlibby]

kib: just disable all builtins

[kib]

lib/libthr/Makefile

35

I would add a comment there, noting that this flag is there for functional purpose, instead of suppressing some warning.

[rlibby]

kib: comment the cflag

[kib]

Thank you

[arichardson]

No sure how much the builtins optimization matters here but if it does the other workaround would be:

void* memcpy_for_plt(void*, const void*, size_t) asm("memcpy")

And then call that.

[rlibby]

In D45891#1046815, @arichardson wrote:

No sure how much the builtins optimization matters here but if it does the other workaround would be:

void* memcpy_for_plt(void*, const void*, size_t) asm("memcpy")

And then call that.

Thanks for the info. I think @kib could speak better to whether any code in this file could be important in terms of performance. In any case, I believe the one memcpy was the only site actually affected when I examined the codegen for amd64.

[kib]

In D45891#1046821, @rlibby wrote:

In D45891#1046815, @arichardson wrote:

No sure how much the builtins optimization matters here but if it does the other workaround would be:

void* memcpy_for_plt(void*, const void*, size_t) asm("memcpy")

And then call that.

Thanks for the info. I think @kib could speak better to whether any code in this file could be important in terms of performance. In any case, I believe the one memcpy was the only site actually affected when I examined the codegen for amd64.

lock_acquire/release are the only functions which performance is critical for runtime.

That said, I do believe that correctness trumps speed.