[Bug rtl-optimization/87608] Very slow swap operations

[amonakov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87608

Alexander Monakov <amonakov at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |amonakov at gcc dot gnu.org

--- Comment #1 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
Note the compiler can evaluate the initialization loop and then also evaluate
the effect of static_sort1 call, so the testcase might give misleading results.
To avoid that, pass the address of 'a' to rdtsc, or introduce a compiler
barrier with an asm:

  asm volatile ("" :: "r"(a) : "memory");

Furthermore, note that the CPU executes the rdtsc instruction without waiting
for all preceding computations to complete. Using lfence just before rdtsc will
ensure that rdtsc reads the cycle counter only after all preceding computations
are done.

On this testcase I think LLVM introduces ternary select operations in the IR
fairly early and then works with straight-line code; in contrast, in GCC we
scalarize the array and have a soup of BBs and phi nodes throughout gimple
passes, which would be very hard to properly clean up on rtl.