Mirror of forum.nim-lang.org

310 :: About LOC default initialization

• stringer (orginal) [2013-12-03T12:44:21+01:00] view original

  I've noticed that Nimrod generates this kind of pattern for temporaries inside a proc call. For ex:

  
  proc add (a: int, b: int): int {.noinit, inline.} =
    result = a + b
  let x = add (1, add (2, 3))
  Run

  
  LOC1 = 0;
  LOC1 = add_121227(2, 3);
  x = add_121227(1, LOC1);
  Run

  While the first zero initialization of LOC1 will be most likely optimized away by a C compiler, my concern is when using compound types or simd types in which case a memset occurs:

  
  let af_f = mm256_add_ps (cf, mm256_add_ps (af, bf))
  Run

  
  memset((void*)&LOC1, 0, sizeof(LOC1));
  LOC1 = _mm256_add_ps(af, bf);
  aff = _mm256_add_ps(cf, LOC1);
  Run

  In this case, the memset is not optimized away. Is there some way to bypass this behavior?

  Note that when using explicitly a let myself, this is of course not happening. For ex:

  
  let tmp0 = mm256_add_ps (af, bf)
  let af_f = mm256_add_ps (cf, tmp0)
  Run

  ... results in the best possible x86 code.

  Thanks a lot !

---

• Araq (orginal) [2013-12-03T13:32:09+01:00] view original

  Should be fixed now.

---

• stringer (orginal) [2013-12-03T21:38:23+01:00] view original

  Thanks !!!

---