[Bug 293] New: Utilize assert (ptr == (ptr & -align)) to assume alignment

gdc-bugzilla at gdcproject.org gdc-bugzilla at gdcproject.org
Thu Jun 21 15:35:33 UTC 2018


https://bugzilla.gdcproject.org/show_bug.cgi?id=293

            Bug ID: 293
           Summary: Utilize assert (ptr == (ptr & -align)) to assume
                    alignment
           Product: GDC
           Version: 8,x
          Hardware: x86_64
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: Normal
         Component: gdc
          Assignee: ibuclaw at gdcproject.org
          Reporter: araspik at protonmail.com

This is a feature request.
GCC provides __builtin_assume_aligned to let the programmer inform the compiler
that a pointer is aligned. Some pointer alignment code I am using is not making
sense to the compiler, and so it is not using aligned SSE movement instructions
when possible. Here is my alignment (pseudo)code (note that I have excluded
casting pointers to arithmetic types and back):

len = -dst & 0xf; // Gets amount needed to align to 16 bytes
if (len) {
  if (len & 1) {
    *cast(ubyte*)dst = *cast(ubyte*)src;
    dst++;
    src++;
    assert ((dst & 1) == 0); // This should work to assume X-byte alignment
  }
  ...
}

// This should work, and the compiler should assume it's true even when asserts
are not being compiled.
assert ((dst & 0xf) == 0); // Now aligned.
assert ((dst & -0x10) == dst); // Same thing.
// I have to do this instead:
dst &= 0xf;

struct void16 {void[16] data;} // Example SSE register optimization
*cast(void16*)dst = *cast(void16*)src; // Should use movdqa and move into dst
aligned
// Does not work via asserts.

This could be used as a global assumption optimization.

GDC arguments (note that no AVX, but SSE 4.2):
gdc -march=native -mtune=native -frelease -Ofast
Additionally, I have excluded Phobos, enabled static compilation, and disabled
PIC and moduleinfo.

-- 
You are receiving this mail because:
You are watching all bug changes.


More information about the D.gnu mailing list