Re: [PATCH v2 4/5] sha1_name: Parse less while finding common prefix

[Derrick Stolee]

My v3 patch is incoming, but I wanted to respond directly to this message.

On 9/25/2017 7:42 PM, Stefan Beller wrote:

On Mon, Sep 25, 2017 at 2:54 AM, Derrick Stolee  wrote:

Create get_hex_char_from_oid() to parse oids one hex character at a
time. This prevents unnecessary copying of hex characters in
extend_abbrev_len() when finding the length of a common prefix.

p0008.1: find_unique_abbrev() for existing objects
--

For 10 repeated tests, each checking 100,000 known objects, we find the
following results when running in a Linux VM:

|   | Pack  | Packed  | Loose  | Base   | New||
| Repo  | Files | Objects | Objects| Time   | Time   | Rel%   |
|---|---|-|||||
| Git   | 1 |  230078 |  0 | 0.08 s | 0.08 s |   0.0% |
| Git   | 5 |  230162 |  0 | 0.17 s | 0.16 s | - 5.9% |
| Git   | 4 |  154310 |  75852 | 0.14 s | 0.12 s | -14.3% |
| Linux | 1 | 5606645 |  0 | 0.50 s | 0.25 s | -50.0% |
| Linux |24 | 5606645 |  0 | 2.41 s | 2.08 s | -13.7% |
| Linux |23 | 5283204 | 323441 | 1.99 s | 1.69 s | -15.1% |
| VSTS  | 1 | 4355923 |  0 | 0.40 s | 0.22 s | -45.0% |
| VSTS  |32 | 4355923 |  0 | 2.09 s | 1.99 s | - 4.8% |
| VSTS  |31 | 4276829 |  79094 | 3.60 s | 3.20 s | -11.1% |

For the Windows repo running in Windows Subsystem for Linux:

 Pack Files: 50
Packed Objects: 22,385,898
  Loose Objects: 492
  Base Time: 4.61 s
   New Time: 4.61 s
  Rel %: 0.0%

p0008.2: find_unique_abbrev() for missing objects
-

For 10 repeated tests, each checking 100,000 missing objects, we find
the following results when running in a Linux VM:

|   | Pack  | Packed  | Loose  | Base   | New||
| Repo  | Files | Objects | Objects| Time   | Time   | Rel%   |
|---|---|-|||||
| Git   | 1 |  230078 |  0 | 0.06 s | 0.05 s | -16.7% |
| Git   | 5 |  230162 |  0 | 0.14 s | 0.15 s | + 7.1% |
| Git   | 4 |  154310 |  75852 | 0.12 s | 0.12 s |   0.0% |
| Linux | 1 | 5606645 |  0 | 0.40 s | 0.17 s | -57.5% |
| Linux |24 | 5606645 |  0 | 1.59 s | 1.30 s | -18.2% |
| Linux |23 | 5283204 | 323441 | 1.23 s | 1.10 s | -10.6% |
| VSTS  | 1 | 4355923 |  0 | 0.25 s | 0.12 s | -52.0% |
| VSTS  |32 | 4355923 |  0 | 1.45 s | 1.34 s | - 7.6% |
| VSTS  |31 | 4276829 |  79094 | 1.59 s | 1.34 s | -15.7% |

For the Windows repo running in Windows Subsystem for Linux:

 Pack Files: 50
Packed Objects: 22,385,898
  Loose Objects: 492
  Base Time: 3.91 s
   New Time: 3.08 s
  Rel %: -21.1%


These number look pretty cool!


Signed-off-by: Derrick Stolee 

Signed-off-by: Derrick Stolee 

double signoff?


Oops! I'll be more careful with my format-patch in the future.




---
  sha1_name.c | 13 +++--
  1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/sha1_name.c b/sha1_name.c
index f2a1ebe49..bb47b6702 100644
--- a/sha1_name.c
+++ b/sha1_name.c
@@ -480,13 +480,22 @@ struct min_abbrev_data {
 char *hex;
  };

+static inline char get_hex_char_from_oid(const struct object_id *oid, int i)

'i' is not very descriptive, maybe add a comment?
(I realize it is just walking through the char*s one by one)

I renamed 'i' to 'pos' in my v3.



Maybe this function (together with the change in the while below)
could go into hex.c as "int progressively_cmp_oids" that returns
the position at which the given oids differ?


+{
+   static const char hex[] = "0123456789abcdef";
+
+   if ((i & 1) == 0)
+   return hex[oid->hash[i >> 1] >> 4];
+   else
+   return hex[oid->hash[i >> 1] & 0xf];
+}

sha1_to_hex_r has very similar code, though it iterates less
and covers both cases in the loop body.

That is the actual reason I propose moving this function
(or a variant thereof) to hex.c as there we can share code.


You're right that sha1_to_hex_r is similar, in fact I based my work on 
it. There are a few reasons I didn't combine the two implementations:


* I wanted to be sure my patch was only touching the code for 
disambiguating short-shas. Modifying code in hex.c would touch many more 
code paths.


* I realize that the extra branch in my version is slower than the 
branchless loop body in sha1_to_hex_r, so either I would slow that 
method or make the method call more complicated by returning two chars 
at a time.


* I wanted to strongly hint that the method should be inlined, but I'm 
not sure how to guarantee that happens across a linker boundary without 
doing strange things in header files.


I'm happy to revisit this after my patch is complete, since I think 
there are interesting trade-offs to consider here. I'd prefer to keep 
this discussion separate from the focus on 

Re: [PATCH v2 4/5] sha1_name: Parse less while finding common prefix

[Stefan Beller]

On Mon, Sep 25, 2017 at 2:54 AM, Derrick Stolee  wrote:
> Create get_hex_char_from_oid() to parse oids one hex character at a
> time. This prevents unnecessary copying of hex characters in
> extend_abbrev_len() when finding the length of a common prefix.
>
> p0008.1: find_unique_abbrev() for existing objects
> --
>
> For 10 repeated tests, each checking 100,000 known objects, we find the
> following results when running in a Linux VM:
>
> |   | Pack  | Packed  | Loose  | Base   | New||
> | Repo  | Files | Objects | Objects| Time   | Time   | Rel%   |
> |---|---|-|||||
> | Git   | 1 |  230078 |  0 | 0.08 s | 0.08 s |   0.0% |
> | Git   | 5 |  230162 |  0 | 0.17 s | 0.16 s | - 5.9% |
> | Git   | 4 |  154310 |  75852 | 0.14 s | 0.12 s | -14.3% |
> | Linux | 1 | 5606645 |  0 | 0.50 s | 0.25 s | -50.0% |
> | Linux |24 | 5606645 |  0 | 2.41 s | 2.08 s | -13.7% |
> | Linux |23 | 5283204 | 323441 | 1.99 s | 1.69 s | -15.1% |
> | VSTS  | 1 | 4355923 |  0 | 0.40 s | 0.22 s | -45.0% |
> | VSTS  |32 | 4355923 |  0 | 2.09 s | 1.99 s | - 4.8% |
> | VSTS  |31 | 4276829 |  79094 | 3.60 s | 3.20 s | -11.1% |
>
> For the Windows repo running in Windows Subsystem for Linux:
>
> Pack Files: 50
> Packed Objects: 22,385,898
>  Loose Objects: 492
>  Base Time: 4.61 s
>   New Time: 4.61 s
>  Rel %: 0.0%
>
> p0008.2: find_unique_abbrev() for missing objects
> -
>
> For 10 repeated tests, each checking 100,000 missing objects, we find
> the following results when running in a Linux VM:
>
> |   | Pack  | Packed  | Loose  | Base   | New||
> | Repo  | Files | Objects | Objects| Time   | Time   | Rel%   |
> |---|---|-|||||
> | Git   | 1 |  230078 |  0 | 0.06 s | 0.05 s | -16.7% |
> | Git   | 5 |  230162 |  0 | 0.14 s | 0.15 s | + 7.1% |
> | Git   | 4 |  154310 |  75852 | 0.12 s | 0.12 s |   0.0% |
> | Linux | 1 | 5606645 |  0 | 0.40 s | 0.17 s | -57.5% |
> | Linux |24 | 5606645 |  0 | 1.59 s | 1.30 s | -18.2% |
> | Linux |23 | 5283204 | 323441 | 1.23 s | 1.10 s | -10.6% |
> | VSTS  | 1 | 4355923 |  0 | 0.25 s | 0.12 s | -52.0% |
> | VSTS  |32 | 4355923 |  0 | 1.45 s | 1.34 s | - 7.6% |
> | VSTS  |31 | 4276829 |  79094 | 1.59 s | 1.34 s | -15.7% |
>
> For the Windows repo running in Windows Subsystem for Linux:
>
> Pack Files: 50
> Packed Objects: 22,385,898
>  Loose Objects: 492
>  Base Time: 3.91 s
>   New Time: 3.08 s
>  Rel %: -21.1%
>

These number look pretty cool!

> Signed-off-by: Derrick Stolee 
>
> Signed-off-by: Derrick Stolee 

double signoff?

> ---
>  sha1_name.c | 13 +++--
>  1 file changed, 11 insertions(+), 2 deletions(-)
>
> diff --git a/sha1_name.c b/sha1_name.c
> index f2a1ebe49..bb47b6702 100644
> --- a/sha1_name.c
> +++ b/sha1_name.c
> @@ -480,13 +480,22 @@ struct min_abbrev_data {
> char *hex;
>  };
>
> +static inline char get_hex_char_from_oid(const struct object_id *oid, int i)

'i' is not very descriptive, maybe add a comment?
(I realize it is just walking through the char*s one by one)

Maybe this function (together with the change in the while below)
could go into hex.c as "int progressively_cmp_oids" that returns
the position at which the given oids differ?

> +{
> +   static const char hex[] = "0123456789abcdef";
> +
> +   if ((i & 1) == 0)
> +   return hex[oid->hash[i >> 1] >> 4];
> +   else
> +   return hex[oid->hash[i >> 1] & 0xf];
> +}

sha1_to_hex_r has very similar code, though it iterates less
and covers both cases in the loop body.

That is the actual reason I propose moving this function
(or a variant thereof) to hex.c as there we can share code.

[PATCH v2 4/5] sha1_name: Parse less while finding common prefix

[Derrick Stolee]

Create get_hex_char_from_oid() to parse oids one hex character at a
time. This prevents unnecessary copying of hex characters in
extend_abbrev_len() when finding the length of a common prefix.

p0008.1: find_unique_abbrev() for existing objects
--

For 10 repeated tests, each checking 100,000 known objects, we find the
following results when running in a Linux VM:

|   | Pack  | Packed  | Loose  | Base   | New||
| Repo  | Files | Objects | Objects| Time   | Time   | Rel%   |
|---|---|-|||||
| Git   | 1 |  230078 |  0 | 0.08 s | 0.08 s |   0.0% |
| Git   | 5 |  230162 |  0 | 0.17 s | 0.16 s | - 5.9% |
| Git   | 4 |  154310 |  75852 | 0.14 s | 0.12 s | -14.3% |
| Linux | 1 | 5606645 |  0 | 0.50 s | 0.25 s | -50.0% |
| Linux |24 | 5606645 |  0 | 2.41 s | 2.08 s | -13.7% |
| Linux |23 | 5283204 | 323441 | 1.99 s | 1.69 s | -15.1% |
| VSTS  | 1 | 4355923 |  0 | 0.40 s | 0.22 s | -45.0% |
| VSTS  |32 | 4355923 |  0 | 2.09 s | 1.99 s | - 4.8% |
| VSTS  |31 | 4276829 |  79094 | 3.60 s | 3.20 s | -11.1% |

For the Windows repo running in Windows Subsystem for Linux:

Pack Files: 50
Packed Objects: 22,385,898
 Loose Objects: 492
 Base Time: 4.61 s
  New Time: 4.61 s
 Rel %: 0.0%

p0008.2: find_unique_abbrev() for missing objects
-

For 10 repeated tests, each checking 100,000 missing objects, we find
the following results when running in a Linux VM:

|   | Pack  | Packed  | Loose  | Base   | New||
| Repo  | Files | Objects | Objects| Time   | Time   | Rel%   |
|---|---|-|||||
| Git   | 1 |  230078 |  0 | 0.06 s | 0.05 s | -16.7% |
| Git   | 5 |  230162 |  0 | 0.14 s | 0.15 s | + 7.1% |
| Git   | 4 |  154310 |  75852 | 0.12 s | 0.12 s |   0.0% |
| Linux | 1 | 5606645 |  0 | 0.40 s | 0.17 s | -57.5% |
| Linux |24 | 5606645 |  0 | 1.59 s | 1.30 s | -18.2% |
| Linux |23 | 5283204 | 323441 | 1.23 s | 1.10 s | -10.6% |
| VSTS  | 1 | 4355923 |  0 | 0.25 s | 0.12 s | -52.0% |
| VSTS  |32 | 4355923 |  0 | 1.45 s | 1.34 s | - 7.6% |
| VSTS  |31 | 4276829 |  79094 | 1.59 s | 1.34 s | -15.7% |

For the Windows repo running in Windows Subsystem for Linux:

Pack Files: 50
Packed Objects: 22,385,898
 Loose Objects: 492
 Base Time: 3.91 s
  New Time: 3.08 s
 Rel %: -21.1%

Signed-off-by: Derrick Stolee 

Signed-off-by: Derrick Stolee 
---
 sha1_name.c | 13 +++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/sha1_name.c b/sha1_name.c
index f2a1ebe49..bb47b6702 100644
--- a/sha1_name.c
+++ b/sha1_name.c
@@ -480,13 +480,22 @@ struct min_abbrev_data {
char *hex;
 };
 
+static inline char get_hex_char_from_oid(const struct object_id *oid, int i)
+{
+   static const char hex[] = "0123456789abcdef";
+
+   if ((i & 1) == 0)
+   return hex[oid->hash[i >> 1] >> 4];
+   else
+   return hex[oid->hash[i >> 1] & 0xf];
+}
+
 static int extend_abbrev_len(const struct object_id *oid, void *cb_data)
 {
struct min_abbrev_data *mad = cb_data;
 
-   char *hex = oid_to_hex(oid);
unsigned int i = mad->init_len;
-   while (mad->hex[i] && mad->hex[i] == hex[i])
+   while (mad->hex[i] && mad->hex[i] == get_hex_char_from_oid(oid, i))
i++;
 
if (i < GIT_MAX_RAWSZ && i >= mad->cur_len)
-- 
2.14.1.538.g56ec8fc98.dirty