threads / patch / 51267

patchOptimize run_diff_files()' rename detection

Subject: [PATCH 0/1] Optimize run_diff_files()' rename detection

## tl;dr

7 messages between Jun 8, 2019 and Jun 10, 2019. Diffs are folded; open one to read it.

replies: 6people: 4as markdown or json

Johannes Schindelin via GitGitGadget· Jun 8, 2019, 14:42 UTC · lore
Just another patch from Git for Windows' branch thicket...
Jeff Hostetler (1):
  diffcore-rename: speed up register_rename_src
 diffcore-rename.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)
base-commit: 8104ec994ea3849a968b4667d072fedd1e688642
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-142%2Fdscho%2Fregister_rename_src-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-142/dscho/register_rename_src-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/142
-- 
gitgitgadget
Jeff Hostetler via GitGitGadget· Jun 8, 2019, 14:42 UTC · re: Johannes Schindelin via GitGitGadget · lore

[PATCH 1/1] diffcore-rename: speed up register_rename_src

From: Jeff Hostetler <jeffhost@microsoft.com>

Teach register_rename_src() to see if new file pair can simply be appended to the rename_src[] array before performing the binary search to find the proper insertion point.

This is a performance optimization. This routine is called during run_diff_files in status and the caller is iterating over the sorted index, so we should expect to be able to append in the normal case. The existing insert logic is preserved so we don't have to assume that, but simply take advantage of it if possible.

Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
---
 diffcore-rename.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)
Show changes to diffcore-rename.c +13 −0
diff --git a/diffcore-rename.c b/diffcore-rename.c
index 07bd34b631..5bfc5f6c22 100644
--- a/diffcore-rename.c
+++ b/diffcore-rename.c
@@ -82,6 +82,18 @@ static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
 
 	first = 0;
 	last = rename_src_nr;
+
+	if (last > 0) {
+		struct diff_rename_src *src = &(rename_src[last-1]);
+		int cmp = strcmp(one->path, src->p->one->path);
+		if (!cmp)
+			return src;
+		if (cmp > 0) {
+			first = last;
+			goto append_it;
+		}
+	}
+
 	while (last > first) {
 		int next = (last + first) >> 1;
 		struct diff_rename_src *src = &(rename_src[next]);
@@ -95,6 +107,7 @@ static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
 		first = next+1;
 	}
 
+append_it:
 	/* insert to make it at "first" */
 	ALLOC_GROW(rename_src, rename_src_nr + 1, rename_src_alloc);
 	rename_src_nr++;
-- 
gitgitgadget
René Scharfe· Jun 8, 2019, 22:27 UTC · re: Jeff Hostetler via GitGitGadget · lore

Re: [PATCH 1/1] diffcore-rename: speed up register_rename_src

Am 08.06.19 um 16:42 schrieb Jeff Hostetler via GitGitGadget:
Show 37 quoted lines
> From: Jeff Hostetler <jeffhost@microsoft.com>
>
> Teach register_rename_src() to see if new file pair
> can simply be appended to the rename_src[] array before
> performing the binary search to find the proper insertion
> point.
>
> This is a performance optimization.  This routine is called
> during run_diff_files in status and the caller is iterating
> over the sorted index, so we should expect to be able to
> append in the normal case.  The existing insert logic is
> preserved so we don't have to assume that, but simply take
> advantage of it if possible.
>
> Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
> Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
> ---
>  diffcore-rename.c | 13 +++++++++++++
>  1 file changed, 13 insertions(+)
>
> diff --git a/diffcore-rename.c b/diffcore-rename.c
> index 07bd34b631..5bfc5f6c22 100644
> --- a/diffcore-rename.c
> +++ b/diffcore-rename.c
> @@ -82,6 +82,18 @@ static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
>
>  	first = 0;
>  	last = rename_src_nr;
> +
> +	if (last > 0) {
> +		struct diff_rename_src *src = &(rename_src[last-1]);
> +		int cmp = strcmp(one->path, src->p->one->path);
> +		if (!cmp)
> +			return src;
> +		if (cmp > 0) {
> +			first = last;
> +			goto append_it;

The goto is not necessary from a logical point of view; the binary search is skipped even without it because the loop condition below is not met anymore.

> +		}

You could decrement "last" at this point as a micro-optimization, to use all three possible outcomes of the comparison to make some progress.

Not sure if any of that would have a _measurable_ impact, though, so I don't mind the patch going in as is.

> +	}
> +
>  	while (last > first) {
>  		int next = (last + first) >> 1;

Hmm, "last" and "first" are ints as well, so this will give weird results when "last" > INT_MAX / 2. I thought 19716b21a4 ("cleanup: fix possible overflow errors in binary search", 2017-10-08) got us rid of those, but git grep -n '(.*+.*) >> 1' actually finds some more cases:

   builtin/ls-files.c:376:         int next = (last + first) >> 1;
   diffcore-rename.c:26:           int next = (last + first) >> 1;
   diffcore-rename.c:86:           int next = (last + first) >> 1;
   dir.c:704:              int cmp, next = (last + first) >> 1;
   name-hash.c:349:                        int mid = (begin + end) >> 1;
   read-cache.c:552:               int next = (last + first) >> 1;
   sh-i18n--envsubst.c:252:          size_t j = (j1 + j2) >> 1;
(Not caused by this patch, of course, just noticed it in the context.)
Show 10 quoted lines
>  		struct diff_rename_src *src = &(rename_src[next]);
> @@ -95,6 +107,7 @@ static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
>  		first = next+1;
>  	}
>
> +append_it:
>  	/* insert to make it at "first" */
>  	ALLOC_GROW(rename_src, rename_src_nr + 1, rename_src_alloc);
>  	rename_src_nr++;
>

Anyway, this here should work as well (and fix the possible overflow), but may be too terse and quirky:

Show changes to diffcore-rename.c +2 −3
diff --git a/diffcore-rename.c b/diffcore-rename.c
index 07bd34b631..f2a9007e08 100644
--- a/diffcore-rename.c
+++ b/diffcore-rename.c
@@ -76,14 +76,13 @@ static int rename_src_nr, rename_src_alloc;

 static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
 {
-	int first, last;
+	int first, last, next;
 	struct diff_filespec *one = p->one;
 	unsigned short score = p->score;

 	first = 0;
 	last = rename_src_nr;
-	while (last > first) {
-		int next = (last + first) >> 1;
+	for (next = last - 1; last > first; next = first + (last - first) / 2) {
 		struct diff_rename_src *src = &(rename_src[next]);
 		int cmp = strcmp(one->path, src->p->one->path);
 		if (!cmp)
Jeff Hostetler· Jun 10, 2019, 14:23 UTC · re: René Scharfe · lore

Re: [PATCH 1/1] diffcore-rename: speed up register_rename_src

On 6/8/2019 6:27 PM, René Scharfe wrote:
Show 42 quoted lines
> Am 08.06.19 um 16:42 schrieb Jeff Hostetler via GitGitGadget:
>> From: Jeff Hostetler <jeffhost@microsoft.com>
>>
>> Teach register_rename_src() to see if new file pair
>> can simply be appended to the rename_src[] array before
>> performing the binary search to find the proper insertion
>> point.
>>
>> This is a performance optimization.  This routine is called
>> during run_diff_files in status and the caller is iterating
>> over the sorted index, so we should expect to be able to
>> append in the normal case.  The existing insert logic is
>> preserved so we don't have to assume that, but simply take
>> advantage of it if possible.
>>
>> Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
>> Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
>> ---
>>   diffcore-rename.c | 13 +++++++++++++
>>   1 file changed, 13 insertions(+)
>>
>> diff --git a/diffcore-rename.c b/diffcore-rename.c
>> index 07bd34b631..5bfc5f6c22 100644
>> --- a/diffcore-rename.c
>> +++ b/diffcore-rename.c
>> @@ -82,6 +82,18 @@ static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
>>
>>   	first = 0;
>>   	last = rename_src_nr;
>> +
>> +	if (last > 0) {
>> +		struct diff_rename_src *src = &(rename_src[last-1]);
>> +		int cmp = strcmp(one->path, src->p->one->path);
>> +		if (!cmp)
>> +			return src;
>> +		if (cmp > 0) {
>> +			first = last;
>> +			goto append_it;
> 
> The goto is not necessary from a logical point of view; the binary
> search is skipped even without it because the loop condition below is
> not met anymore.

Perhaps this is just a personal style thing, but I prefer the goto because it makes it clear that we know the answer and are done searching. I know it's a small thing and I know we do it all the time, but setting state here in just the right way as to defeat the loop would work, but is a little less clear. But again, that is a personal style thing I guess.

Show 7 quoted lines
>> +		}
> 
> You could decrement "last" at this point as a micro-optimization, to use
> all three possible outcomes of the comparison to make some progress.
> 
> Not sure if any of that would have a _measurable_ impact, though, so I
> don't mind the patch going in as is.

Yes, it looks like we could decrement "last" here. And yes, I suspect it would have minimal impact. I'll pass on that in this series to keep it focused on the advertised goal of simply appending if possible.

Show 57 quoted lines
> 
>> +	}
>> +
>>   	while (last > first) {
>>   		int next = (last + first) >> 1;
> 
> Hmm, "last" and "first" are ints as well, so this will give weird results
> when "last" > INT_MAX / 2.  I thought 19716b21a4 ("cleanup: fix possible
> overflow errors in binary search", 2017-10-08) got us rid of those, but
> git grep -n '(.*+.*) >> 1' actually finds some more cases:
> 
>     builtin/ls-files.c:376:         int next = (last + first) >> 1;
>     diffcore-rename.c:26:           int next = (last + first) >> 1;
>     diffcore-rename.c:86:           int next = (last + first) >> 1;
>     dir.c:704:              int cmp, next = (last + first) >> 1;
>     name-hash.c:349:                        int mid = (begin + end) >> 1;
>     read-cache.c:552:               int next = (last + first) >> 1;
>     sh-i18n--envsubst.c:252:          size_t j = (j1 + j2) >> 1;
> 
> (Not caused by this patch, of course, just noticed it in the context.)
> 
>>   		struct diff_rename_src *src = &(rename_src[next]);
>> @@ -95,6 +107,7 @@ static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
>>   		first = next+1;
>>   	}
>>
>> +append_it:
>>   	/* insert to make it at "first" */
>>   	ALLOC_GROW(rename_src, rename_src_nr + 1, rename_src_alloc);
>>   	rename_src_nr++;
>>
> 
> Anyway, this here should work as well (and fix the possible overflow),
> but may be too terse and quirky:
> 
> diff --git a/diffcore-rename.c b/diffcore-rename.c
> index 07bd34b631..f2a9007e08 100644
> --- a/diffcore-rename.c
> +++ b/diffcore-rename.c
> @@ -76,14 +76,13 @@ static int rename_src_nr, rename_src_alloc;
> 
>   static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
>   {
> -	int first, last;
> +	int first, last, next;
>   	struct diff_filespec *one = p->one;
>   	unsigned short score = p->score;
> 
>   	first = 0;
>   	last = rename_src_nr;
> -	while (last > first) {
> -		int next = (last + first) >> 1;
> +	for (next = last - 1; last > first; next = first + (last - first) / 2) {
>   		struct diff_rename_src *src = &(rename_src[next]);
>   		int cmp = strcmp(one->path, src->p->one->path);
>   		if (!cmp)
> 

I'd like to limit the scope of this patch series to just the advertised topic and save the arithmetic fixes to another patch series like 19716b21a4.

Jeff
SZEDER Gábor· Jun 10, 2019, 12:26 UTC · re: Jeff Hostetler via GitGitGadget · lore

Re: [PATCH 1/1] diffcore-rename: speed up register_rename_src

On Sat, Jun 08, 2019 at 07:42:42AM -0700, Jeff Hostetler via GitGitGadget wrote:
Show 13 quoted lines
> From: Jeff Hostetler <jeffhost@microsoft.com>
> 
> Teach register_rename_src() to see if new file pair
> can simply be appended to the rename_src[] array before
> performing the binary search to find the proper insertion
> point.
> 
> This is a performance optimization.  This routine is called
> during run_diff_files in status and the caller is iterating
> over the sorted index, so we should expect to be able to
> append in the normal case.  The existing insert logic is
> preserved so we don't have to assume that, but simply take
> advantage of it if possible.

Could you add a command and performance figures to the commit message to show off the benefits of this patch?

From the description it's clear that it's a performance optimization, but it's unclear whether it has a noticeable, or at least measurable effect, or it's "only" a micro-optimization.

I tried something more substantial than a simple 'git status':
  # without this patch
  $ time perf record -g ./git log --oneline -m --name-only v2.20.0 >/dev/null
  [ ... ]
  
  real    2m4.320s
  user    2m0.913s
  sys     0m2.284s
  $ perf report -g none |grep -E '(diffcore_rename|register_rename_src)'
      52.40%     0.79%  git      git                 [.] diffcore_rename
       0.01%     0.01%  git      git                 [.] register_rename_src

but it looks like that while more than half of the considerable runtime is spent detecting renames, the time spent in register_rename_src(), the function optimized in this patch, is negligible.

Show 39 quoted lines
> Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
> Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
> ---
>  diffcore-rename.c | 13 +++++++++++++
>  1 file changed, 13 insertions(+)
> 
> diff --git a/diffcore-rename.c b/diffcore-rename.c
> index 07bd34b631..5bfc5f6c22 100644
> --- a/diffcore-rename.c
> +++ b/diffcore-rename.c
> @@ -82,6 +82,18 @@ static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
>  
>  	first = 0;
>  	last = rename_src_nr;
> +
> +	if (last > 0) {
> +		struct diff_rename_src *src = &(rename_src[last-1]);
> +		int cmp = strcmp(one->path, src->p->one->path);
> +		if (!cmp)
> +			return src;
> +		if (cmp > 0) {
> +			first = last;
> +			goto append_it;
> +		}
> +	}
> +
>  	while (last > first) {
>  		int next = (last + first) >> 1;
>  		struct diff_rename_src *src = &(rename_src[next]);
> @@ -95,6 +107,7 @@ static struct diff_rename_src *register_rename_src(struct diff_filepair *p)
>  		first = next+1;
>  	}
>  
> +append_it:
>  	/* insert to make it at "first" */
>  	ALLOC_GROW(rename_src, rename_src_nr + 1, rename_src_alloc);
>  	rename_src_nr++;
> -- 
> gitgitgadget
Jeff Hostetler· Jun 10, 2019, 14:55 UTC · re: SZEDER Gábor · lore

Re: [PATCH 1/1] diffcore-rename: speed up register_rename_src

On 6/10/2019 8:26 AM, SZEDER Gábor wrote:
Show 40 quoted lines
> On Sat, Jun 08, 2019 at 07:42:42AM -0700, Jeff Hostetler via GitGitGadget wrote:
>> From: Jeff Hostetler <jeffhost@microsoft.com>
>>
>> Teach register_rename_src() to see if new file pair
>> can simply be appended to the rename_src[] array before
>> performing the binary search to find the proper insertion
>> point.
>>
>> This is a performance optimization.  This routine is called
>> during run_diff_files in status and the caller is iterating
>> over the sorted index, so we should expect to be able to
>> append in the normal case.  The existing insert logic is
>> preserved so we don't have to assume that, but simply take
>> advantage of it if possible.
> 
> Could you add a command and performance figures to the commit message
> to show off the benefits of this patch?
> 
>  From the description it's clear that it's a performance optimization,
> but it's unclear whether it has a noticeable, or at least measurable
> effect, or it's "only" a micro-optimization.
> 
> I tried something more substantial than a simple 'git status':
> 
>    # without this patch
>    $ time perf record -g ./git log --oneline -m --name-only v2.20.0 >/dev/null
>    [ ... ]
>    
>    real    2m4.320s
>    user    2m0.913s
>    sys     0m2.284s
>    $ perf report -g none |grep -E '(diffcore_rename|register_rename_src)'
>        52.40%     0.79%  git      git                 [.] diffcore_rename
>         0.01%     0.01%  git      git                 [.] register_rename_src
> 
> but it looks like that while more than half of the considerable
> runtime is spent detecting renames, the time spent in
> register_rename_src(), the function optimized in this patch, is
> negligible.
> 

I just wanted to send an ACK. I'll include perf numbers and more clarity after I dig up my notes on this.

Jeff
Jeff Hostetler· Jun 10, 2019, 14:43 UTC · re: Johannes Schindelin via GitGitGadget · lore

Re: [PATCH 0/1] Optimize run_diff_files()' rename detection

On 6/8/2019 10:42 AM, Johannes Schindelin via GitGitGadget wrote:
Show 14 quoted lines
> Just another patch from Git for Windows' branch thicket...
> 
> Jeff Hostetler (1):
>    diffcore-rename: speed up register_rename_src
> 
>   diffcore-rename.c | 13 +++++++++++++
>   1 file changed, 13 insertions(+)
> 
> 
> base-commit: 8104ec994ea3849a968b4667d072fedd1e688642
> Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-142%2Fdscho%2Fregister_rename_src-v1
> Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-142/dscho/register_rename_src-v1
> Pull-Request: https://github.com/gitgitgadget/git/pull/142
> 

For the sake of full disclosure, we added this patch to Git for Windows in December 2016.

It was discussed on the mailing list in April 2017 [1] but it was shelved for various reasons.

Let me put this one on hold while I dig up my notes from 2016 and re-review of the original mailing list suggestions and see where I want to take this patch going forward.

Jeff

[1] https://public-inbox.org/git/20170418194421.22453-1-git@jeffhostetler.com/

← back to recent threads