This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH GCC]Reduce compilation time for IVOPT by skipping cost computation in use group
- From: "H.J. Lu" <hjl dot tools at gmail dot com>
- To: "Bin.Cheng" <amker dot cheng at gmail dot com>
- Cc: Richard Biener <richard dot guenther at gmail dot com>, Bin Cheng <Bin dot Cheng at arm dot com>, "gcc-patches at gcc dot gnu dot org" <gcc-patches at gcc dot gnu dot org>
- Date: Thu, 31 Mar 2016 05:27:29 -0700
- Subject: Re: [PATCH GCC]Reduce compilation time for IVOPT by skipping cost computation in use group
- Authentication-results: sourceware.org; auth=none
- References: <DB5PR08MB11443BE7AB126D01B3D51353E7820 at DB5PR08MB1144 dot eurprd08 dot prod dot outlook dot com> <CAFiYyc046x6RJugVfQpb0pUxFgOBORP9iVBaT+2d_2_jASg_2A at mail dot gmail dot com> <CAHFci2-Ltv=5Ee49f7-bGUiuCbGVR0LuxKpQd4v7Sxa5fcGgrg at mail dot gmail dot com>
On Wed, Mar 30, 2016 at 5:11 AM, Bin.Cheng <amker.cheng@gmail.com> wrote:
> On Wed, Mar 30, 2016 at 9:09 AM, Richard Biener
> <richard.guenther@gmail.com> wrote:
>> On Thu, Mar 24, 2016 at 6:26 PM, Bin Cheng <Bin.Cheng@arm.com> wrote:
>>> Hi,
>>> Quite lot of time is used when IVOPT computes cost for <use, cand> pairs. As a matter of fact, some pairs are very similar to each other, and we can abstract and compute cost only once for these pairs. This is a patch doing so, the idea is skipping cost computation for sub-uses in each group, of course it may result in different assembly code for some complicated cases because it estimates cost rather than doing real computation. I did double check one of such case that the change in generated assembly is not degeneration. For an IVOPT heavy program (spec2k/173), this patch reduces IVOPT's compilation time by 7~8%, as well as the memory consumption on my developing machine.
>>>
>>> Bootstrap & test on x86_64.
>>>
>>> For spec2k6 data on x86_64. Maybe because I ran spec2k6 compiled with patched GCC in unclean environment, some cases are regressed by small amount (< %1). I manually compared assembly code for several cases, including ones with the largest regression (still within <1%). I could confirm that generated assembly code is exact the same as unpatched GCC, except for function emit_library_call_value_1 in 403.gcc/calls.c.
>>>
>>> In this case, difference of IVOPT dumps is as below:
>>>
>>> $ diff -y trunk/calls.c.154t.ivopts patch/calls.c.154t.ivopts
>>>
>>> <bb 44>: <bb 44>:
>>> # val_21 = PHI <val_175(168), val_650(43)> # val_21 = PHI <val_175(168), val_650(43)>
>>> _811 = (void *) ivtmp.322_829; _811 = (void *) ivtmp.322_829;
>>> MEM[base: _811, offset: -48B] = val_21; | MEM[base: _811, offset: -32B] = val_21;
>>> _810 = (void *) ivtmp.322_829; _810 = (void *) ivtmp.322_829;
>>> MEM[base: _810, offset: -40B] = mode_163; | MEM[base: _810, offset: -24B] = mode_163;
>>> _182 = function_arg (&args_so_far, mode_163, 0B, 1); _182 = function_arg (&args_so_far, mode_163, 0B, 1);
>>> _809 = (void *) ivtmp.322_829; _809 = (void *) ivtmp.322_829;
>>> MEM[base: _809, offset: -32B] = _182; | MEM[base: _809, offset: -16B] = _182;
>>> _807 = (void *) ivtmp.322_829; _807 = (void *) ivtmp.322_829;
>>> MEM[base: _807, offset: -24B] = 0; | MEM[base: _807, offset: -8B] = 0;
>>> _185 = (struct args_size *) ivtmp.322_829; | _801 = ivtmp.322_829 + 16;
>>> _801 = ivtmp.322_829 + 18446744073709551600; <
>>> _800 = (struct args_size *) _801; _800 = (struct args_size *) _801;
>>> _186 = _800; | _185 = _800;
>>> > _186 = (struct args_size *) ivtmp.322_829;
>>> _187 = _182 != 0B; _187 = _182 != 0B;
>>> _188 = (int) _187; _188 = (int) _187;
>>> locate_and_pad_parm (mode_163, 0B, _188, 0B, &args_size, _1 locate_and_pad_parm (mode_163, 0B, _188, 0B, &args_size, _1
>>> _802 = (void *) ivtmp.322_829; _802 = (void *) ivtmp.322_829;
>>> _190 = MEM[base: _802, offset: 8B]; | _190 = MEM[base: _802, offset: 24B];
>>> if (_190 != 0B) if (_190 != 0B)
>>> goto <bb 45>; goto <bb 45>;
>>> else else
>>> goto <bb 46>; goto <bb 46>;
>>>
>>> <bb 45>: <bb 45>:
>>> fancy_abort ("calls.c", 3724, &__FUNCTION__); fancy_abort ("calls.c", 3724, &__FUNCTION__);
>>>
>>> It's only an offset difference in IV. And below is difference of generated assembly:
>>> $ diff -y trunk/calls.S patch/calls.S
>>> .L489: .L489:
>>> leaq -80(%rbp), %rdi leaq -80(%rbp), %rdi
>>> xorl %edx, %edx xorl %edx, %edx
>>> movl $1, %ecx movl $1, %ecx
>>> movl %r13d, %esi movl %r13d, %esi
>>> movq %rax, -48(%r15) <
>>> movl %r13d, -40(%r15) <
>>> call function_arg <
>>> movl $0, -24(%r15) <
>>> movq %rax, -32(%r15) movq %rax, -32(%r15)
>>> > movl %r13d, -24(%r15)
>>> > call function_arg
>>> xorl %edx, %edx xorl %edx, %edx
>>> pushq %r12 | movq %rax, -16(%r15)
>>> testq %rax, %rax testq %rax, %rax
>>> pushq %r15 | leaq 16(%r15), %rax <--I1
>>> leaq -16(%r15), %r9 | movl $0, -8(%r15)
>>> leaq -112(%rbp), %r8 leaq -112(%rbp), %r8
>>> > pushq %r12
>>> setne %dl setne %dl
>>> movl %r13d, %edi | movq %r15, %r9 <--I2
>>> > pushq %rax <--I3
>>> xorl %ecx, %ecx xorl %ecx, %ecx
>>> xorl %esi, %esi xorl %esi, %esi
>>> > movl %r13d, %edi
>>> call locate_and_pad_parm call locate_and_pad_parm
>>> cmpq $0, 8(%r15) | cmpq $0, 24(%r15)
>>> popq %rax popq %rax
>>> popq %rdx popq %rdx
>>> jne .L602 jne .L602
>>>
>>> There is one additional move instruction (I2) after patching, but I believe it's a choice of RA. If we switch %rax/%r9 in instructions I1/I2 as below:
>>> ...
>>> leaq 16(%r15), %r9
>>> ...
>>> movq %r15, %rax
>>> pushq %r15
>>>
>>> Then I2 becomes redundant and can be removed.
>>>
>>> I will collect performance data on AArch64 to make sure there is no breakage either. So is it OK?
>>
>> I think the patch is ok - note that I have a hard time following the
>> code, esp. the 'first' flag.
>>
>> + /* Add cost for sub uses in group. */
>> + do
>> + {
>> + /* Compute cost for the first sub_use with different offset to
>> + the first one and use it afterwards, because the cost could
>> + be very different if the offset is different. */
>> + if (first && use->addr_offset != sub_use->addr_offset)
>> + {
>> + first = false;
>> + sub_cost = get_computation_cost (data, sub_use, cand, true,
>> + NULL, &can_autoinc, NULL);
>> + if (infinite_cost_p (sub_cost))
>> + {
>> + cost = infinite_cost;
>> + break;
>> + }
>> + }
>> +
>> + cost = add_costs (cost, sub_cost);
>> + sub_use = sub_use->next;
>> + }
>> + while (sub_use);
>>
>> we start this loop with first = true, so for each sub-use we compute
>> no new sub-cost until
>> use->addr_offset changes for the first time after which we will never
>> compute sub-cost
>> again. So we call get_computation_cost at most twice for al sub-uses.
>>
>> I suppose all sub-uses have equal ->addr_base. Suppose sub-uses are then sorted
>> after ->addr_offset what keeps that list from having three different
>> addr_offset but
>> with "very different cost"? There seems to be group_address_uses but
>> that suggests
>> the cost might be actually the same for all sub-uses.
>>
>> So adding a little explaining before the loop over sub-uses would be
>> appreciated.
>
> Thanks for reviewing. Here is the patch with trivially revised comments.
> I also collected benchmark data for spec2k6 on AArch64, there is no
> surprise except for case 456.hmmer. I double checked generated
> assembly and can confirm it's not real.
> Will apply the patch later.
>
> Thanks,
> bin
>>
>> Thanks,
>> Richard.
>>
>>> Thanks,
>>> bin
>>>
>>> 2016-03-23 Bin Cheng <bin.cheng@arm.com>
>>>
>>> * tree-ssa-loop-ivopts.c (struct comp_cost): New scrach field.
>>> (no_cost, infinite_cost): Initialize the new field.
>>> (get_computation_cost_at): Record setup cost.
>>> (determine_use_iv_cost_address): Skip cost computation for sub
>>> uses if we can estimate it without losing accuracy.
>>>
This may have caused bootstrap failure on ia32:
https://gcc.gnu.org/ml/gcc-regression/2016-03/msg00347.html
--
H.J.