[
https://issues.apache.org/jira/browse/ARROW-13382?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17383520#comment-17383520
]
David Li commented on ARROW-13382:
----------------------------------
I actually have a guess: the name of the mangled SumArray symbol is the same
between the non-SIMD and the various SIMD implementations. So what may be
happening is we're generating the same SumArray in aggregate_basic.cc.o,
aggregate_basic_avx2.cc.o, and aggregate_basic_avx512.cc.o. They clash, and
previously the linker picked the non-SIMD one and now it picks a SIMD one. I'm
guessing this because in the 'good' run, the "AVX512" version of the kernel has
no AVX512 in it.
> [C++] Aggregation over scalars fails autobrew R job
> ---------------------------------------------------
>
> Key: ARROW-13382
> URL: https://issues.apache.org/jira/browse/ARROW-13382
> Project: Apache Arrow
> Issue Type: Bug
> Components: C++
> Reporter: David Li
> Assignee: David Li
> Priority: Blocker
> Fix For: 5.0.0
>
>
> [https://github.com/ursacomputing/crossbow/runs/3091873413#step:7:488]
> {noformat}
> *** caught illegal operation ***
> address 0x109dc30cc, cause 'illegal opcode'Traceback:
> 1: compute__CallFunction(function_name, args, options)
> 2: call_function(FUN, a, options = list(na.rm = na.rm, na.min_count =
> na.min_count))
> 3: scalar_aggregate("sum", ..., na.rm = na.rm)
> 4: sum.ArrowDatum(<environment>, na.rm = FALSE)
> 5: eval_bare(expr, quo_get_env(quo))
> 6: quasi_label(enquo(object), arg = "object") {noformat}
> I would guess at first glance the compiler is autovectorizing something more
> than necessary?
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
