We have a restriction in x86_64 targets on the no of memory references inside a loop that can be held in a loop buffer and take advantage of it. The patch is trying to find the unroll count, that would still satisfy number of memory references after unrolling and can be held in a loop buffer Regards, Venkat.
-----Original Message----- From: linaro-toolchain [mailto:linaro-toolchain-bounces@lists.linaro.org] On Behalf Of Kugan Sent: Thursday, February 18, 2016 3:15 AM To: Virendra Kumar Pathak; Linaro Toolchain Mailman List Subject: Re: setting loop buffer size in the gcc (aarch64)
Hi,
On 18/02/16 01:51, Virendra Kumar Pathak wrote:
Hi Toolchain Group,
I am trying to study the effect of loop buffer size on loop unrolling & the way gcc (aarch64) handles this.
It depends on the micro-architecture. Usually, loop buffer helps to hold the loop completely and supplies the instruction fetch unit from there. The main benefit used to be the dynamic energy reduction. i.e., you don't access the main (L1) cache for the loop iterations.
Loop unrolling on the other hand can remove the control instructions and allow the compiler to optimize across loop iterations.
To my understanding, Loop Buffer is like i-cache which contains pre-decoded instruction that can be re-used if branch instruction loopbacks to an instruction which is still present in the buffer. For example, in Intel’s Nehalem loop buffer size is 28 u-ops. In LLVM compiler, it seems LoopMicroOpBufferSize is for the same purpose. However, I could not find any parameter/variable inside config/aarch64 representing loop buffer size. I am using Linaro gcc 5.2.1
[Question]
- Is there any example inside aarch64 (or in general) which uses the
loop buffer size in loop unrolling decision? If yes, could you please mention the relevant files or code section?
Look at this patch for x86: https://gcc.gnu.org/ml/gcc-patches/2013-11/msg02567.html
This is implemented using TARGET_LOOP_UNROLL_ADJUST as you have found out.
Thanks, Kugan
- Otherwise any guidance/input on adding this support in aarch64
backend assuming architecture has the loop buffer support.
[My Experiments/Code Browsing] I have collected following information from code browsing. Please correct if I missed or misunderstood something.
TARGET_LOOP_UNROLL_ADJUST - This target hook return the number of times a loop can be unrolled. This can be used to handle the architecture constraint such number of memory references inside a loop e.g. ix86_loop_unroll_adjust() & s390_loop_unroll_adjust(). On the same note, can this be used to handle loop buffer size too?
Without above hook, in loop-unroll.c parameters like PARAM_MAX_UNROLLED_INSNS (default 200), PARAM_MAX_AVERAGE_UNROLLED_INSNS (default 80) decides the
unrolling
factor. e.g. nunroll = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / loop->ninsns;
In config/aarch64.c, I found align_loops variable in aarch64_override_options_after_change() function. I guess this an alignment done before starting the loop header in the executable. This should not play any role in loop unrolling. Right?
So any guidance on how we can instruct aarch64 backend to utilize loop buffer size in deciding the loop unrolling factor?
Thanks in advance for your time.
-- with regards, Virendra Kumar Pathak
linaro-toolchain mailing list linaro-toolchain@lists.linaro.org https://lists.linaro.org/mailman/listinfo/linaro-toolchain
linaro-toolchain mailing list linaro-toolchain@lists.linaro.org https://lists.linaro.org/mailman/listinfo/linaro-toolchain