https://bugs.llvm.org/show_bug.cgi?id=48429
Bug ID: 48429
Summary: Generated scatter instructions are slower than scalar
version
Product: libraries
Version: trunk
Hardware: PC
OS: Linux
Status: NEW
Severity: enhancement
Priority: P
Component: Backend: X86
Assignee: unassignedb...@nondot.org
Reporter: car...@google.com
CC: craig.top...@gmail.com, llvm-bugs@lists.llvm.org,
llvm-...@redking.me.uk, pengfei.w...@intel.com,
spatel+l...@rotateright.com
Compile the following code with command line
clang '--target=x86_64-grtev4-linux-gnu' -maes -m64 -mcx16 -msse4.2 -mpclmul
'-mprefer-vector-width=128' -fexperimental-new-pass-manager
-fsized-deallocation -O3 '-std=gnu++17' -c scatter.cc -save-temps
__attribute((target("avx,avx2,fma,avx512f,avx512dq,avx512bw"))) void
foo(int d, const float* ptr, float* dest)
{
const float* ptr_end = ptr + d;
for (; ptr != ptr_end; ++ptr, dest += 16) {
dest[0] = ptr[-1 * d];
dest[1] = ptr[0 * d];
dest[2] = ptr[1 * d];
dest[3] = ptr[2 * d];
}
}
llvm generates 4 element scatters, which is more than 50% slower than scalar
version on my skylake desktop.
The problem is in function int X86TTIImpl::getGatherScatterOpCost(), it has
already found scatter is not profitable if avx512vl is not enabled, so it
should be scalarized, and return a scalarized cost. But the caller
LoopVectorize doesn't know it's a scalarized cost, it thinks it's a scatter
cost, and compares it with a different scalar cost computed by
getMemInstScalarizationCost, and unfortunately X86 backend computed scalar cost
is smaller than LoopVectorize computed scalar cost, so LoopVectorize thinks
scatter is cheaper than scalarize, and generates the slow scatter version.
--
You are receiving this mail because:
You are on the CC list for the bug._______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
