Issue #20394 has been updated by Eregon (Benoit Daloze). BTW I think the custom String#to_i makes a lot of sense, and might be faster than String#to_i because it needs to handle fewer cases and can e.g. handle reading directly from an IO vs going through a String in between. ---------------------------------------- Feature #20394: Add an offset parameter to `String#to_i` https://bugs.ruby-lang.org/issues/20394#change-107471 * Author: byroot (Jean Boussier) * Status: Open ---------------------------------------- ### Context I maintain the `redis-client` gem, and it comes with an optional swapable implementation in C that binds the `hiredis` C client, [which used to performs up to 5 times faster in some cases](https://github.com/redis-rb/redis-client/commit/9fabd57c6786a03fe0c6021eab5b...). I recently paired with @tenderlovemaking to try to close this gap, or even try to make the pure Ruby version faster, and we came up with several optimizations that now almost make both version on par (assuming YJIT is enabled). An important source of performance loss, is that the Redis protocol is line based and to parse it in Ruby requires to slice a lot of small strings from the buffer. To give an example, here's how an Array with two String (`["foo", "plop"]`) is serialized in RESP3 (Redis protocol): ``` *2\r\n $3\r\n foo\r\n $4\r\n plop\r\n ``` From this you can understand that a big hotspot in the parser is essentially `Integer(gets)`. With @tenderlovemaking we managed to get [a fairly significant perf boost](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...) by avoiding these string allocation using `String#getbyte` and [basically implementing a rudimentary `String#to_i(offset: )` in Ruby](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...). But while the gains are huge with YJIT enabled, they are much more tame with the interpreter. And it feels a bit wrong to have to implement this sorts of things for performance reasons. ### `String#to_i(offset: )` Similar to `String#unpack(offset:)` ([Feature #18254]), I believe `String#to_i(offset: )` would be useful. ### Alternative new `String#unpack` format Another possibility would be to add a new format to `String#pack` `String#unpack` for decimal numbers. It sounds a bit weird at first, but given it supports things like Base64 and hexadecimal, perhaps it's not that much of a stretch? -- https://bugs.ruby-lang.org/

Issue #20394 has been updated by Dan0042 (Daniel DeLorme). byroot (Jean Boussier) wrote in #note-3:

We tried this, but it was way slower (https://github.com/redis-rb/redis-client/pull/150), we haven't dug much as to why. But generally I'd love if I could just rely on the internal IO buffer instead of having to maintain a buffer string an an offset, but we're far from there I'm afraid.

The idea of an offset in a string is perfectly represented via the cursor in IO/StringIO, so it would be ideal to just use that instead of adding `offset` kwargs to various core methods. If the only issue is performance I tried to get an idea of the problem with some cursory benchmarking for #getbyte and... ```ruby require "stringio" fd = File.open("1.2MB-file") str = fd.read io = StringIO.new(str) bm{ fd.rewind; nil while fd.getbyte } #=> 0.0270s bm{ io.rewind; nil while io.getbyte } #=> 0.0195s bm{ i=0; i+=1 while str.getbyte(i) } #=> 0.0237s ``` It doesn't seem like String#getbyte is much faster than File#getbyte, and StringIO#getbyte is fastest of all. Maybe the result would be different if using sockets? In that case it might be worth buffering to a StringIO. Of course even with that it's not ideal to call #getbyte multiple times to build an integer, so maybe we could have something like `IO#get_i(base=10)` which reads as many numeric characters as possible and return them as an integer. Returns `nil` if no numeric characters at cursor. WDYT? ---------------------------------------- Feature #20394: Add an offset parameter to `String#to_i` https://bugs.ruby-lang.org/issues/20394#change-107475 * Author: byroot (Jean Boussier) * Status: Open ---------------------------------------- ### Context I maintain the `redis-client` gem, and it comes with an optional swapable implementation in C that binds the `hiredis` C client, [which used to performs up to 5 times faster in some cases](https://github.com/redis-rb/redis-client/commit/9fabd57c6786a03fe0c6021eab5b...). I recently paired with @tenderlovemaking to try to close this gap, or even try to make the pure Ruby version faster, and we came up with several optimizations that now almost make both version on par (assuming YJIT is enabled). An important source of performance loss, is that the Redis protocol is line based and to parse it in Ruby requires to slice a lot of small strings from the buffer. To give an example, here's how an Array with two String (`["foo", "plop"]`) is serialized in RESP3 (Redis protocol): ``` *2\r\n $3\r\n foo\r\n $4\r\n plop\r\n ``` From this you can understand that a big hotspot in the parser is essentially `Integer(gets)`. With @tenderlovemaking we managed to get [a fairly significant perf boost](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...) by avoiding these string allocation using `String#getbyte` and [basically implementing a rudimentary `String#to_i(offset: )` in Ruby](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...). But while the gains are huge with YJIT enabled, they are much more tame with the interpreter. And it feels a bit wrong to have to implement this sorts of things for performance reasons. ### `String#to_i(offset: )` Similar to `String#unpack(offset:)` ([Feature #18254]), I believe `String#to_i(offset: )` would be useful. ### Alternative new `String#unpack` format Another possibility would be to add a new format to `String#pack` `String#unpack` for decimal numbers. It sounds a bit weird at first, but given it supports things like Base64 and hexadecimal, perhaps it's not that much of a stretch? -- https://bugs.ruby-lang.org/

Issue #20394 has been updated by mame (Yusuke Endoh). As @eregon said, `String#to_i(offset:)` with byte offset looks strange to me. I like `IO#get_i(base=10)` and `StringIO#get_i` (despite the name). I was a little concerned that it returns infinitely large Bignum when the IO reads "99999999..." in terms of Denial of Service. But this is essentially the same for `IO#gets`, so it may be acceptable to prohibit using `get_i` for unreliable IO. About the method name, I'd propose `IO#get_integer` or `IO#parse_integer` or `IO#scan_integer`. I don't think it will be used so frequently enough to give an abbreviated name like `get_i`. BTW, I think other reader than Integer will be wanted. Recently I saw [The One Billion Row Challenge](https://github.com/gunnarmorling/1brc) and thought that I want `IO#parse_float`. Generalizing, we may want `IO#scanf`, but that's probably overkill? (This comment includes ideas from chatting with @osyoyu) ---------------------------------------- Feature #20394: Add an offset parameter to `String#to_i` https://bugs.ruby-lang.org/issues/20394#change-107498 * Author: byroot (Jean Boussier) * Status: Open ---------------------------------------- ### Context I maintain the `redis-client` gem, and it comes with an optional swapable implementation in C that binds the `hiredis` C client, [which used to performs up to 5 times faster in some cases](https://github.com/redis-rb/redis-client/commit/9fabd57c6786a03fe0c6021eab5b...). I recently paired with @tenderlovemaking to try to close this gap, or even try to make the pure Ruby version faster, and we came up with several optimizations that now almost make both version on par (assuming YJIT is enabled). An important source of performance loss, is that the Redis protocol is line based and to parse it in Ruby requires to slice a lot of small strings from the buffer. To give an example, here's how an Array with two String (`["foo", "plop"]`) is serialized in RESP3 (Redis protocol): ``` *2\r\n $3\r\n foo\r\n $4\r\n plop\r\n ``` From this you can understand that a big hotspot in the parser is essentially `Integer(gets)`. With @tenderlovemaking we managed to get [a fairly significant perf boost](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...) by avoiding these string allocation using `String#getbyte` and [basically implementing a rudimentary `String#to_i(offset: )` in Ruby](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...). But while the gains are huge with YJIT enabled, they are much more tame with the interpreter. And it feels a bit wrong to have to implement this sorts of things for performance reasons. ### `String#to_i(offset: )` Similar to `String#unpack(offset:)` ([Feature #18254]), I believe `String#to_i(offset: )` would be useful. ### Alternative new `String#unpack` format Another possibility would be to add a new format to `String#pack` `String#unpack` for decimal numbers. It sounds a bit weird at first, but given it supports things like Base64 and hexadecimal, perhaps it's not that much of a stretch? -- https://bugs.ruby-lang.org/

Issue #20394 has been updated by Dan0042 (Daniel DeLorme). mame (Yusuke Endoh) wrote in #note-6:

Generalizing, we may want `IO#scanf`, but that's probably overkill?

This previously existed in the stdlib but was removed: #16170 byroot (Jean Boussier) wrote in #note-7:

Indeed, but the problem is that you then have very few methods to parse values or peak in the buffer to find elements. Lots of methods needed for parsing various protocols are present on `String`, but not `IO`.

With StringIO you have the best of both worlds! You can use the methods with built-in #pos, like #getbyte, or drop down to the underlying #string if you want to use specialized methods like #unpack or #scan ! We could even add #unpack and #scan to StringIO and they would start the unpack/scan operation at the current #pos. As well as any other methods needed for parsing various protocols. It comes down to: we can either make String a better buffer (adding `#to_i(offset:)`), or make StringIO a better buffer (adding `#get_integer`). IMHO the latter seems a better direction in the long run.

Also until recently, methods like `IO#gets` didn't have a way to be timedout, so they weren't safe to use in such context. In 3.2+ there's now `IO#timeout=`, but as I said parsing directly from the IO turned out much slower in my attempts, but perhaps I didn't do it well.

I think the timeout stuff is orthogonal to parsing with offsets. Let's compare String vs StringIO so we can ignore timeouts and focus on the API design of #to_i vs #get_integer ```ruby # @str is String @str << io.gets case @str[@offset] #@offset is character position, so this may be O(n) operation when "*" @offset += 1 nb_elements = @str.to_i(offset: @offset) @offset += ??? end # @io is StringIO @io.string << io.gets case @io.getc when "*" nb_elements = @io.get_integer end ``` It seems to me that the Redis protocol, like most wire protocols, is designed to be parsed in a streaming fashion, so a stream-oriented API like StringIO works pretty well. But I recognize the speed aspect is important, with String#getbyte much faster than StringIO#getbyte when YJIT is enabled... which I find very strange. Unlike IO, StringIO is just a wrapper over a String object, so in theory there's no reason why they should be so different. ---------------------------------------- Feature #20394: Add an offset parameter to `String#to_i` https://bugs.ruby-lang.org/issues/20394#change-107509 * Author: byroot (Jean Boussier) * Status: Open ---------------------------------------- ### Context I maintain the `redis-client` gem, and it comes with an optional swapable implementation in C that binds the `hiredis` C client, [which used to performs up to 5 times faster in some cases](https://github.com/redis-rb/redis-client/commit/9fabd57c6786a03fe0c6021eab5b...). I recently paired with @tenderlovemaking to try to close this gap, or even try to make the pure Ruby version faster, and we came up with several optimizations that now almost make both version on par (assuming YJIT is enabled). An important source of performance loss, is that the Redis protocol is line based and to parse it in Ruby requires to slice a lot of small strings from the buffer. To give an example, here's how an Array with two String (`["foo", "plop"]`) is serialized in RESP3 (Redis protocol): ``` *2\r\n $3\r\n foo\r\n $4\r\n plop\r\n ``` From this you can understand that a big hotspot in the parser is essentially `Integer(gets)`. With @tenderlovemaking we managed to get [a fairly significant perf boost](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...) by avoiding these string allocation using `String#getbyte` and [basically implementing a rudimentary `String#to_i(offset: )` in Ruby](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...). But while the gains are huge with YJIT enabled, they are much more tame with the interpreter. And it feels a bit wrong to have to implement this sorts of things for performance reasons. ### `String#to_i(offset: )` Similar to `String#unpack(offset:)` ([Feature #18254]), I believe `String#to_i(offset: )` would be useful. ### Alternative new `String#unpack` format Another possibility would be to add a new format to `String#pack` `String#unpack` for decimal numbers. It sounds a bit weird at first, but given it supports things like Base64 and hexadecimal, perhaps it's not that much of a stretch? -- https://bugs.ruby-lang.org/

Issue #20394 has been updated by byroot (Jean Boussier). Status changed from Open to Closed

This interpretation mismatch could be a source of vulnerability.

Good catch @mame, that does indeed make the `to_i` proposal much more problematic. I guess I don't really have much of a proposal anymore.

With StringIO you have the best of both worlds!

Not today no. You can look at gems that implement network protocols, I'm yet to find one that uses `StringIO` as a buffer, `StringIO` isn't as convenient as you make it out to be. Maybe it could become that, but it isn't today.

It comes down to: we can either make String a better buffer [...] or make StringIO a better buffer [...]. IMHO the latter seems a better direction in the long run.

The thing is, was StringIO even thought as a buffer in the first place? My understanding is that it's just meant as a facade to pass strings to APIs that expect an IO, nothing more. I don't see any extra methods on it that suggest using it as a buffer. Now for better or for worse, since Ruby strings are mutable, and historically they didn't have an encoding, they're used as buffers everywhere. i.e. look at methods like `IO#read_nonblock`, they take a string parameter called `outbuf`, look at core gems like `net-http`, they use that same "read into a string, then parse the string" pattern, etc. I'd be all for a dedicated Buffer class that allow to efficiently parse text protocols like HTTP and RESP3, but right now all we got is `String`.

I think the timeout stuff is orthogonal to parsing with offsets.

It's not. If you are parsing a stream directly from an IO using blocking methods, you must be able to timeout if the character or pattern you are waiting for never comes. In your example, that `io.gets` could potentially block forever if the IO is a socket and the server is unresponsive or malicious. Hence why most if not all protocol clients use `read_nonblock` into a String and then parse the string.

with String#getbyte much faster than StringIO#getbyte when YJIT is enabled... which I find very strange. Unlike IO, StringIO is just a wrapper over a String object, so in theory there's no reason why they should be so different.

It's not that strange. YJIT has special optimization for `String#getbyte` and some other String methods because they are such hotspots. Additionally, `StringIO` is implemented in C so YJIT can't optimize it, and even if it was re-written in Ruby, it would be some extra method calls that YJIT isn't yet capable of inlining. So in the end I think I'll just implement some small gems that provide the capabilities I need. ---------------------------------------- Feature #20394: Add an offset parameter to `String#to_i` https://bugs.ruby-lang.org/issues/20394#change-107517 * Author: byroot (Jean Boussier) * Status: Closed ---------------------------------------- ### Context I maintain the `redis-client` gem, and it comes with an optional swapable implementation in C that binds the `hiredis` C client, [which used to performs up to 5 times faster in some cases](https://github.com/redis-rb/redis-client/commit/9fabd57c6786a03fe0c6021eab5b...). I recently paired with @tenderlovemaking to try to close this gap, or even try to make the pure Ruby version faster, and we came up with several optimizations that now almost make both version on par (assuming YJIT is enabled). An important source of performance loss, is that the Redis protocol is line based and to parse it in Ruby requires to slice a lot of small strings from the buffer. To give an example, here's how an Array with two String (`["foo", "plop"]`) is serialized in RESP3 (Redis protocol): ``` *2\r\n $3\r\n foo\r\n $4\r\n plop\r\n ``` From this you can understand that a big hotspot in the parser is essentially `Integer(gets)`. With @tenderlovemaking we managed to get [a fairly significant perf boost](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...) by avoiding these string allocation using `String#getbyte` and [basically implementing a rudimentary `String#to_i(offset: )` in Ruby](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...). But while the gains are huge with YJIT enabled, they are much more tame with the interpreter. And it feels a bit wrong to have to implement this sorts of things for performance reasons. ### `String#to_i(offset: )` Similar to `String#unpack(offset:)` ([Feature #18254]), I believe `String#to_i(offset: )` would be useful. ### Alternative new `String#unpack` format Another possibility would be to add a new format to `String#pack` `String#unpack` for decimal numbers. It sounds a bit weird at first, but given it supports things like Base64 and hexadecimal, perhaps it's not that much of a stretch? -- https://bugs.ruby-lang.org/

Issue #20394 has been updated by byroot (Jean Boussier).

What about IO::Buffer? It was introduced somewhat stealthily and its API is somewhat unique, and yet by its idea it seems to be exactly that.

I'm aware of it, but it only offer binary oriented methods. Perhaps it could be extended to offer text protocols oriented methods, but it's really starting from zero right now. If you look at @ioquatix 's own HTTP gem, it uses its own `Async::IO::Stream` class which funnily enough is a `String` subclass... https://github.com/socketry/async-io/blob/085b506b9bc5b4669f1d65b6fe1b972f09..., which kinda confirms my point that in Ruby today, the best buffer is `String`. ---------------------------------------- Feature #20394: Add an offset parameter to `String#to_i` https://bugs.ruby-lang.org/issues/20394#change-107519 * Author: byroot (Jean Boussier) * Status: Closed ---------------------------------------- ### Context I maintain the `redis-client` gem, and it comes with an optional swapable implementation in C that binds the `hiredis` C client, [which used to performs up to 5 times faster in some cases](https://github.com/redis-rb/redis-client/commit/9fabd57c6786a03fe0c6021eab5b...). I recently paired with @tenderlovemaking to try to close this gap, or even try to make the pure Ruby version faster, and we came up with several optimizations that now almost make both version on par (assuming YJIT is enabled). An important source of performance loss, is that the Redis protocol is line based and to parse it in Ruby requires to slice a lot of small strings from the buffer. To give an example, here's how an Array with two String (`["foo", "plop"]`) is serialized in RESP3 (Redis protocol): ``` *2\r\n $3\r\n foo\r\n $4\r\n plop\r\n ``` From this you can understand that a big hotspot in the parser is essentially `Integer(gets)`. With @tenderlovemaking we managed to get [a fairly significant perf boost](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...) by avoiding these string allocation using `String#getbyte` and [basically implementing a rudimentary `String#to_i(offset: )` in Ruby](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...). But while the gains are huge with YJIT enabled, they are much more tame with the interpreter. And it feels a bit wrong to have to implement this sorts of things for performance reasons. ### `String#to_i(offset: )` Similar to `String#unpack(offset:)` ([Feature #18254]), I believe `String#to_i(offset: )` would be useful. ### Alternative new `String#unpack` format Another possibility would be to add a new format to `String#pack` `String#unpack` for decimal numbers. It sounds a bit weird at first, but given it supports things like Base64 and hexadecimal, perhaps it's not that much of a stretch? -- https://bugs.ruby-lang.org/

Issue #20394 has been updated by byroot (Jean Boussier).

but rest assured once we can start using IO::Buffer exclusively, I will do so.

I'd be curious to see an HTTP1 parser using IO::Buffer. I don't see how you'd do it today. ---------------------------------------- Feature #20394: Add an offset parameter to `String#to_i` https://bugs.ruby-lang.org/issues/20394#change-107522 * Author: byroot (Jean Boussier) * Status: Closed ---------------------------------------- ### Context I maintain the `redis-client` gem, and it comes with an optional swapable implementation in C that binds the `hiredis` C client, [which used to performs up to 5 times faster in some cases](https://github.com/redis-rb/redis-client/commit/9fabd57c6786a03fe0c6021eab5b...). I recently paired with @tenderlovemaking to try to close this gap, or even try to make the pure Ruby version faster, and we came up with several optimizations that now almost make both version on par (assuming YJIT is enabled). An important source of performance loss, is that the Redis protocol is line based and to parse it in Ruby requires to slice a lot of small strings from the buffer. To give an example, here's how an Array with two String (`["foo", "plop"]`) is serialized in RESP3 (Redis protocol): ``` *2\r\n $3\r\n foo\r\n $4\r\n plop\r\n ``` From this you can understand that a big hotspot in the parser is essentially `Integer(gets)`. With @tenderlovemaking we managed to get [a fairly significant perf boost](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...) by avoiding these string allocation using `String#getbyte` and [basically implementing a rudimentary `String#to_i(offset: )` in Ruby](https://github.com/redis-rb/redis-client/commit/41b3abe94243d2598211d448c4e4...). But while the gains are huge with YJIT enabled, they are much more tame with the interpreter. And it feels a bit wrong to have to implement this sorts of things for performance reasons. ### `String#to_i(offset: )` Similar to `String#unpack(offset:)` ([Feature #18254]), I believe `String#to_i(offset: )` would be useful. ### Alternative new `String#unpack` format Another possibility would be to add a new format to `String#pack` `String#unpack` for decimal numbers. It sounds a bit weird at first, but given it supports things like Base64 and hexadecimal, perhaps it's not that much of a stretch? -- https://bugs.ruby-lang.org/