[llvm-bugs] [Bug 26450] ARM code runs 2x slower compared to gcc

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=26450

James Molloy <james.molloy at arm.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |INVALID

--- Comment #5 from James Molloy <james.molloy at arm.com> ---
Hi,

OK, there's two things here:

Firstly, it seems __umodsi3 and friends are significantly slower than
__aeabi_idivmod. GCC is generating __aeabi_idivmod - perhaps we should? We
select __modsi3 unless the target is EABI or Android - I suspect that should be
EABI, Android or GNUEABI.

GCC 4.9:        1.24s
Clang 3.7:      3.48s
Clang 3.7 (using __aeabi_idivmod): 1.15s

Secondly, you're not specifying a CPU. That's why your division is going out to
the library. Unless you're on a Cortex-A9, you'll have hardware division. Use
-mcpu to enable it.

GCC 4.9 with -mcpu=cortex-a15: 276ms
Clang 3.7 with -mcpu=cortex-a15: 258ms

(I had to switch to using perf stat's task-clock metric because time elapsed
was getting too noisy)

By the way: "I've just discovered how immature LLVM/Clang was on ARM." (from
https://users.rust-lang.org/t/executable-size-and-performance-vs-c/4496/34)

That's a little over the top - the ARM backend is around 10 years old now, it's
fairly mature.

James

-- 
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/20160203/159f1783/attachment.html>