Julian Brown,
<http://mid.mail-archive.com/
20210920134603.
16459021@squid.athome>:
| [...] the randomness shouldn't be necessary for the
| correctness of the patch (i.e. it could just be "base = bounds_lo", or
| indeed folded into the line after).
|
| The "ar.invalid ()" case happens when we fail to allocate a block of
| memory in LDS space for broadcasting a particular set of variables,
| and trigger a fall-back path in the broadcasting code that adds extra
| barriers around the broadcast in question. I imagine I was thinking
| that adding randomness could mean we can "get lucky" sometimes and
| avoid needing those barriers in some cases, but in fact I don't think
| that was implemented, so the randomness is useless. (Or it could just
| have been leftover debug code... oops).
gcc/
PR other/102408
* omp-oacc-neuter-broadcast.cc (oacc_do_neutering): Evaluate
'random ()' to '0'.
if (ar.invalid ())
{
- unsigned HOST_WIDE_INT base;
- base = bounds_lo + random () % 512;
- base = (base + align - 1) & ~(align - 1);
+ unsigned HOST_WIDE_INT base
+ = (bounds_lo + align - 1) & ~(align - 1);
if (base + size > bounds_hi)
error_at (UNKNOWN_LOCATION, "shared-memory region overflow");
std::pair<unsigned HOST_WIDE_INT, bool> base_inrng