[llvm-bugs] [Bug 28126] New: SCoP detection fails on Julia version of PolyBench's `trmm` benchmark
via llvm-bugs
llvm-bugs at lists.llvm.org
Tue Jun 14 08:08:37 PDT 2016
https://llvm.org/bugs/show_bug.cgi?id=28126
Bug ID: 28126
Summary: SCoP detection fails on Julia version of PolyBench's
`trmm` benchmark
Product: Polly
Version: unspecified
Hardware: PC
OS: All
Status: NEW
Severity: normal
Priority: P
Component: Optimizer
Assignee: polly-dev at googlegroups.com
Reporter: matthias.j.reisinger at gmail.com
CC: llvm-bugs at lists.llvm.org
Classification: Unclassified
Created attachment 16540
--> https://llvm.org/bugs/attachment.cgi?id=16540&action=edit
Julia's LLVM IR for `trmm`
I ported PolyBench's `trmm` benchmark to Julia to test Polly on the LLVM IR
that Julia produces when compiling it. The according Julia function that I used
for this purpose looks as follows:
@polly function kernel_trmm(alpha, A, B)
m,n = size(B)
for i = 1:m, j = 1:n
for k = (i+1):m
B[i,j] += A[k,i] * B[k,j]
end
B[i,j] = alpha * B[i,j]
end
end
The LLVM code that Julia emits for this function (when disabling bounds-checks
via `--check-bounds=no`) is attached as `kernel_trmm.ll`.
I invoked Polly the following way on this LLVM code:
$ opt -debug-only=polly-detect -polly-codegen -S llvm/kernel_trmm.ll >
/dev/null
This produces the following output:
Checking region: top => <Function Return>
Top level region is invalid
Checking region: top.split => L10
Non affine loop bound '({-2,+,-1}<%if> + %37)' in loop: if12
Checking region: if => L.L10_crit_edge
Non affine loop bound '({-2,+,-1}<%if> + %37)' in loop: if12
Checking region: if => L.loopexit
Non affine loop bound '({-2,+,-1}<%if> + %37)' in loop: if12
Checking region: if11 => L4.L.loopexit_crit_edge
Non affine loop bound '({-2,+,-1}<%if> + %37)' in loop: if12
Checking region: if11 => L7
Non affine loop bound '({-2,+,-1}<%if> + %37)' in loop: if12
Checking region: if12 => L5.L7_crit_edge
OK
Expanding if12 => L5.L7_crit_edge
Trying if12 => L7
to if12 => L7
Region can not profitably be optimized!
For computing loop bounds Julia emits `select` instructions and it seems that
they are the source of the problem. This issue has also been subject to
discussion on the polly-dev mailing list:
https://groups.google.com/forum/#!topic/polly-dev/P2RZUlKRo0I
The original C version of `trmm` however, when compiled via clang, can be
optimized by Polly without problems.
I am currently working on a solution and hope to come up with new findings as
soon as possible.
--
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20160614/1d0bf55f/attachment.html>
More information about the llvm-bugs
mailing list