[llvm] r342578 - [X86] Handle COPYs of physregs better (regalloc hints)
Jonas Paulsson via llvm-commits
llvm-commits at lists.llvm.org
Sat Sep 29 11:14:32 PDT 2018
Hi Jordan,
thanks for the test case. I think I found the reason for the compile
time regression - please try https://reviews.llvm.org/D52316.
/Jonas
On 2018-09-27 21:52, Jordan Rupprecht wrote:
> Sure,
> I wasn't able to reduce the test case much -- the long compilation
> time seems to be a factor of the length of the method, so a small
> method wouldn't reproduce it.
> $ time bin/llc /tmp/repro.ll -o /tmp/repro-r342577.o
> bin/llc /tmp/repro.ll -o /tmp/repro-r342577.o 2.87s user 0.07s system
> 99% cpu 2.942 total
> $ time bin/llc /tmp/repro.ll -o /tmp/repro-r342578.o
> bin/llc /tmp/repro.ll -o /tmp/repro-r342578.o 7.71s user 0.08s system
> 99% cpu 7.795 total
> I've attached repro.ll, but I generated it from the C++ snippet above,
> with a copy of Eigen pulled from a few minutes ago. It wasn't slow
> enough with the version of eigen at head, so I added a 35x35 dummy
> matrix to the example.
>
> I've noticed a small sampling of other strange failures caused by this
> commit as well, one of which just revealed an actual bug in a test,
> but another which I haven't quite figured out yet -- I'm still trying
> to see if it's a bug with this commit.
>
> On Thu, Sep 27, 2018 at 3:45 AM Jonas Paulsson
> <paulsson at linux.vnet.ibm.com> wrote:
>> Hi,
>>
>> I would be happy to take a look. Could you please post a preprocessed /
>> bitcode (reduced if possible) test case, preferably llc input, with a
>> command line?
>>
>> /Jonas
>>
>>
>> On 2018-09-27 10:14, Simon Pilgrim wrote:
>>> This is a question for Jonas - I just kept the x86 tests up to date
>>> until the patch was accepted.
>>>
>>> On 27/09/2018 00:08, Jordan Rupprecht wrote:
>>>> Hi Simon,
>>>>
>>>> We seem to have some extreme compile slowdowns in sanitizer (asan,
>>>> msan, and tsan) modes after r342578. In one case, we've seen something
>>>> built with tsan and -O2 go from 20 seconds to 12 minutes. It doesn't
>>>> seem to be a general slowdown, just certain "interesting" pieces of
>>>> code (heavy template users, generated code that has thousands of
>>>> expressions, etc.)
>>>>
>>>> We've seen it in all sanitizer modes, but the tsan example mentioned
>>>> can be distilled down to something like this:
>>>>
>>>> $ time bin/clang -c repro.cpp -I</path/to/eigen> -O2
>>>> -fno-strict-aliasing -fsanitize=thread -o repro.o
>>>> @r342577: 5.25s user 0.68s system 99% cpu 5.984 total
>>>> @r342578: 9.50s user 0.26s system 99% cpu 9.811 total
>>>>
>>>> $ cat repro.cpp
>>>> #include <Eigen/Core>
>>>> Eigen::MatrixXi foo(int x) {
>>>> Eigen::MatrixXi m;
>>>> switch (x) {
>>>> case 3:
>>>> m.resize(3, 3);
>>>> m << 1, 2, 3, //
>>>> 1, 2, 3, //
>>>> 1, 2, 3;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> break;
>>>> case 5:
>>>> m.resize(5, 5);
>>>> m << 1, 2, 3, 4, 5, //
>>>> 1, 2, 3, 4, 5, //
>>>> 1, 2, 3, 4, 5, //
>>>> 1, 2, 3, 4, 5, //
>>>> 1, 2, 3, 4, 5;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> break;
>>>> case 7:
>>>> m.resize(7, 7);
>>>> m << 1, 2, 3, 4, 5, 6, 7, //
>>>> 1, 2, 3, 4, 5, 6, 7, //
>>>> 1, 2, 3, 4, 5, 6, 7, //
>>>> 1, 2, 3, 4, 5, 6, 7, //
>>>> 1, 2, 3, 4, 5, 6, 7, //
>>>> 1, 2, 3, 4, 5, 6, 7, //
>>>> 1, 2, 3, 4, 5, 6, 7;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> break;
>>>> case 25:
>>>> m.resize(25, 25);
>>>> m << 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
>>>> 17, 18, 19,
>>>> 20, 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> break;
>>>> case 30:
>>>> m.resize(30, 30);
>>>> m << 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
>>>> 17, 18, 19,
>>>> 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, //
>>>> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
>>>> 18, 19, 20,
>>>> 21, 22, 23, 24, 25, 26, 27, 28, 29, 30;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> m.array() -= 1;
>>>> break;
>>>> default:
>>>> break;
>>>> }
>>>> return m;
>>>> }
>>>>
>>>> In some cases, we can workaround this by rewriting the methods to be
>>>> simpler. However, it'd be nice to understand what's going on -- is
>>>> there a O(n^2) algorithm that enableMultipleCopyHints() is enabling
>>>> that would cause large methods to be slow to compile?
>>>>
>>>> Thanks -- Jordan
More information about the llvm-commits
mailing list