New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Proposal: Conditional Compilation #3538

Open
kitsonk opened this Issue Jun 17, 2015 · 45 comments

Comments

Projects
None yet
@kitsonk
Copy link
Contributor

kitsonk commented Jun 17, 2015

Proposal: Conditional Compilation

Problem Statement

At design time, developers often find that they need to deal with certain scenarios to make their code ubiquitous and runs in every environment and under every runtime condition. At build time however, they want to emit code that is more suited for the runtime environment that they are targetting by not emitting code that is relevant to that environment.

This is directly related to #449 but it also covers some other issues in a similar problem space.

Similar Functionality

There are several other examples of apporaches to solving this problem:

Considerations

Most of the solutions above use "magic" language features that significantly affect the AST of the code. One of the benefits of the has.js approach is that the code is transparent for runtime feature detection and build time optimisation. For example, the following would be how design time would work:

has.add('host-node', (typeof process == "object" && process.versions && process.versions.node && process.versions.v8));

if (has('host-node')) {
    /* do something node */
}
else {
    /* do something non-node */
}

If you then wanted to do a build that targeted NodeJS, then you would simply assert to the build tool (staticHasFlags) that instead of detecting that feature at runtime, host-node was in fact true. The build tool would then realise that the else branch was unreachable and remove that branch from the built code.

Because the solution sits entirely within the language syntax without any sort of "magical" directives or syntax, it does not take a lot of knowledge for a developer to leverage it.

Also by doing this, you do not have to do heavy changes to the AST as part of the complication process and it should be easy to identify branches that are "dead" and can be dropped out of the emit.

Of course this approach doesn't specifically address conditionality of other language features, like the ability to conditionally load modules or conditional classes, though there are other features being introduced in TypeScript (e.g. local types #3266) which when coupled with this would address conditionality of other language features.

Proposed Changes

In order to support conditional compile time emitting, there needs to be a language mechanic to identify blocks of code that should be emitted under certain conditions and a mechanism for determining if they are to be emitted. There also needs to be a mechanism to determine these conditions at compile time.

Defining a Conditional Identifier at Design Time

It is proposed that a new keyword is introduced to allow the introduction of a different class of identifier that is neither a variable or a constant. Introduction of a TypeScript only keyword should not be taken lightly and it is proposed that either condition or has is used to express these identifiers. When expressed at design time, the identifier will be given a value which can be evaluated at runtime, with block scope. This then can be substituted though a compile time with another value.

Of the two keywords, this proposal suggests that has is more functional in meaning, but might be less desirable because of potential for existing code breakage, but examples utlise the has keyword.

For example, in TypeScript the following would be a way of declaring a condition:

has hostNode: boolean = Boolean(typeof process == "object" && process.versions && process.versions.node && process.versions.v8);

if (hostNode) {
    console.log('You are running under node.');
}
else {
    console.log('You are not running under node.');
}

This would then emit, assuming there is no compile time substitutional value available (and targeting ES6) as:

const hostNode = Boolean(typeof process == "object" && process.versions && process.versions.node && process.versions.v8);
if (hostNode) {
    console.log('You are running under node.');
}
else {
    console.log('You are not running under node.');
}

Defining the value of a Conditional Identifier at Compile Time

In order to provide the compile time values, an augmentation of the tsconfig.json is proposed. A new attribute will be proposed that will be named in line with the keyword of either conditionValues or hasValues. Different tsconfig.json can be used for the different builds desired. Not considered in this proposal is consideration of how these values might be passed to tsc directly.

Here is an example of tsconfig.json:

{
    "version": "1.6.0",
    "compilerOptions": {
        "target": "es5",
        "module": "umd",
        "declaration": false,
        "noImplicitAny": true,
        "removeComments": true,
        "noLib": false,
        "sourceMap": true,
        "outDir": "./"
    },
    "hasValues": {
        "hostNode": true
    }
}

Compiled Code

So given the tsconfig.json above and the following TypeScript:

has hostNode: boolean = Boolean(typeof process == "object" && process.versions && process.versions.node && process.versions.v8);

if (hostNode) {
    console.log('You are running under node.');
}
else {
    console.log('You are not running under node.');
}

You would expect the following to be emitted:

console.log('You are running under node.');

As the compiler would replace the symbol of hostNode with the value provided in tsconfig.json and then substitute that value in the AST. It would then realise that the one of the branches was unreachable at compile time and then collapse the AST branch and only emit the reachable code.

@tinganho

This comment has been minimized.

Copy link
Contributor

tinganho commented Jun 18, 2015

I don't quite like that the proposal isn't addressing conditional imports. I think that is the core issue of having conditionals in the first place.

const hostNode = Boolean(typeof process == "object" && process.versions && process.versions.node && process.versions.v8);
if (hostNode) {
console.log('You are running under node.');
}
else {
console.log('You are not running under node.');
}

I'm not sure if I like the runtime property check. We are all developers and we know TypeScript is a compiled language. Why do we need to have a solution when we don't define any flag values? I would rather see it compile as if all conditionals where true then.

I also like the design of conditional flags in C# and C++ because they look like "commented code". Which kind of infer that they only interfere on compile time and not during runtime. They are also simple to understand just like a simple if statement. I guess they are also simpler for the compiler, just let the scanner scan or skip characters depending on if the conditional are true or false. Instead of having a complex tree shaking that removes dead code.

With your solution you are also using runtime statements if-else statements that are being shaked off at compile time.

@kitsonk

This comment has been minimized.

Copy link
Contributor Author

kitsonk commented Jun 18, 2015

I don't quite like that the proposal isn't addressing conditional imports. I think that is the core issue of having conditionals in the first place.

I think that would be a different solution. That is mainly because it is not likely the same pattern can be used with imports without breaking the functionality. The way we addressed this in Dojo, which then worked for both runtime and build time, was to utilise the AMD loader plugin mechanism to evaluate the "magical" MID string expressed as a ternary expression to determine if a certain has flag value and rewrite the MID appropriately. Because ES6 Modules does not currently solve conditional loading, I assumed (hopefully correctly) that TypeScript would want to wait until that solution is evident before solving that themselves.

I'm not sure if I like the runtime property check. We are all developers and we know TypeScript is a compiled language. Why do we need to have a solution when we don't define any flag values? I would rather see it compile as if all conditionals where true then.

Sometimes a developer will want their code to be emitted as isomorphic, especially if they want to distribute it as a library without the end user having to be aware of TypeScript. This is aligned to design goal "4. Emit clean, idiomatic, recognizable JavaScript code." as well as "7. Preserve runtime behavior of all JavaScript code." and "10. Be a cross-platform development tool." What the runtime code is, is up to the developer and whether it gets emitted or not is up to the developer.

I also like the design of conditional flags in C# and C++ because they look like "commented code".

Maybe you should come up with an alternative proposal. Also, it would seem that that is something that TypeScript has largely avoided, "compiler hints". I dislike "auto-magic" comments/compiler hints personally and seeing it avoided in TypeScript made me happy. You often end up with surprises and the TypeScript Design Goals also state that TypeScript should not "introduce behaviour that is likely to surprise users".

I guess they are also simpler for the compiler, just let the scanner scan or skip characters depending on if the conditional are true or false.

Not necessarily, you will still be modifying the AST if you expect things like intellisense to continue to work. Ignoring written code under certain conditions is never straight forward. I am not sure why you feel comments make this process any easier for the compiler.

With your solution you are also using runtime statements if-else statements that are being shaked off at compile time.

Exactly, but only when supplied with compile time values. Why do you feel that is a bad thing?

@stephanedr

This comment has been minimized.

Copy link

stephanedr commented Jun 18, 2015

I would also like that TS supports conditional compilation, but more in the C++/C# way too.
My main use is for assertions.
With your proposal @kitsonk, I would be able to empty a function, but not remove its call, e.g.:

has debug = true;
var assert = debugMode ? function (cond: boolean) { if (cond) throw new AssertionError(); }
                       : function (cond: boolean) { };

With a C++/C# implementation:

#if DEBUG
function _assert(cond: boolean) { if (cond) throw new AssertionError(); }
#define assert(cond) _assert(cond)
#else
#define assert(cond)
#endif

Additionally, if TS would provide "macros" for the current filename and line number, I would be able to retrieve them relative to the TS source code, whereas today a stack trace reports line numbers relative to the generated JS file...

Some points from your proposal:

  • You seem to consider such a flag as constant, so what is the interest in keeping the 2 alternatives (if/else) when no substitution is given (its value cannot be changed at runtime)? For me the dead code elimination should apply as well.
  • Instead of introducing a new keyword has, I think we could stay with const and allow all constants to be overridable at compile time (and apply dead code elimination with any constant expressions).
@kitsonk

This comment has been minimized.

Copy link
Contributor Author

kitsonk commented Jun 19, 2015

I would also like that TS supports conditional compilation, but more in the C++/C# way too.
My main use is for assertions.

@mhegazy said "Pre-processor directives (i.e.#ifdefs) are not desirable." I took him at his word.

As far as your example, I am a bit lost on how the following would eliminate the function call? Doesn't it generate an assert as a noop anyways, just like you did in TypeScript?

Additionally, if TS would provide "macros" for the current filename and line number, I would be able to retrieve them relative to the TS source code, whereas today a stack trace reports line numbers relative to the generated JS file...

Totally different topic... You do know about the sourceMap compiler option?

You seem to consider such a flag as constant, so what is the interest in keeping the 2 alternatives (if/else) when no substitution is given (its value cannot be changed at runtime)? For me the dead code elimination should apply as well.

There is a big difference between build/compile time value and a runtime value. The intent of my proposal is to allow both, without changing the source code. You can provide all your "feature" logic in your code and then you can choose to have it compiled out, or resolved runtime. In theory, you would always want to have some runtime value. Other examples would be like if you where trying to shim things like Promises or Object.observe. You would write the code to handle both cases, with a clear feature/condition and then you can choose to have that resolved at runtime (and the whole set of code is emitted, including the resolution logic) or you could choose to have two builds, both optimised for the features being there or not.

Instead of introducing a new keyword has, I think we could stay with const and allow all constants to be overridable at compile time (and apply dead code elimination with any constant expressions).

Why I proposed it was because I felt it would lead to less surprises. Again, according to the TypeScript Design Goals, TypeScript should not "introduce behaviour that is likely to surprise users". By leveraging const you would have to go through all your code and make sure you didn't have name collisions before you were sure that the compiler wouldn't surprise you. While I suspect introducing new keywords isn't straight forward (as has or condition) would have to be excised from everyone's code, it would more likely throw syntax errors than actually surprise. I maybe wrong on that aspect though, as I don't know how complex what I am proposing would take to implement.

@stephanedr

This comment has been minimized.

Copy link

stephanedr commented Jun 19, 2015

@mhegazy said "Pre-processor directives (i.e.#ifdefs) are not desirable." I took him at his word.

I just wanted to report my need for conditional compilation, which doesn't seem supported by your proposal. At the end I don't really care of the exact mechanism which might be implemented.
For example, a decorator like C#'s Conditional[DEBUG] could be fine for me, but this would also require the TS compiler to remove calls to the excluded method.

As far as your example, I am a bit lost on how the following would eliminate the function call? Doesn't it generate an assert as a noop anyways, just like you did in TypeScript?

When DEBUG is 0 (or not defined), assert(...); is replaced by ;, so completely removed (consider text replacement). Compared to a solution which would call an empty function, this also removes its argument(s) (e.g. with assert(a() == b()), a() and b() will be executed, their results compared, then the comparison result passed to the empty function).

Totally different topic... You do know about the sourceMap compiler option?

My goal is not to debug the code under a browser, but to create a mail with the exception stack trace when an uncaught exception occurs at client side.
By having those "macros" I could embed them to exception messages.

There is a big difference between build/compile time value and a runtime value. The intent of my proposal is to allow both, without changing the source code. You can provide all your "feature" logic in your code and then you can choose to have it compiled out, or resolved runtime. In theory, you would always want to have some runtime value. Other examples would be like if you where trying to shim things like Promises or Object.observe. You would write the code to handle both cases, with a clear feature/condition and then you can choose to have that resolved at runtime (and the whole set of code is emitted, including the resolution logic) or you could choose to have two builds, both optimised for the features being there or not.

OK, I misunderstood your example. In has hostNode: boolean = ... I considered that the expression was constant so simplified by TS as either true or false. So as it was generated in JS as const, there was no more way to dynamically modify it. If the expression is indeed not constant, you indeed need to keep the 2 alternatives.

@RichiCoder1

This comment has been minimized.

Copy link

RichiCoder1 commented Jun 19, 2015

@stephanedr Just jumping in, the idea of giving the compiler hints using design-time decorators like @conditional(DEBUG) has been tossed around more then once and would partial meet the goal of this.

@kitsonk

This comment has been minimized.

Copy link
Contributor Author

kitsonk commented Jun 20, 2015

My goal is not to debug the code under a browser, but to create a mail with the exception stack trace when an uncaught exception occurs at client side.

Again, side topic... Mozilla's Source Map would allow you to determine the TypeScript original positions for the source and actually rewrite the stack trace. There are a few other more complete tools out there that leverage source-map and would do the heavy lifting for you. I suspect even with your macros, it would be hard to cover all the transforms that occur during transpilation, where this is more certain.

@stephanedr

This comment has been minimized.

Copy link

stephanedr commented Jun 22, 2015

@RichiCoder1, correct me if I'm wrong, but using decorators will only allow to empty the assert function. What I would like is to remove all the assert calls.

@kitsonk, providing file name / line number should be quite easy for a compiler, as it already maintains them to report compilation errors.

@am11

This comment has been minimized.

Copy link

am11 commented Nov 14, 2015

Regarding passing the flags to process, I have a suggestion inspired by some of the node.js based conventions (among which JSCS has aced in this aspect by incorporating every approach):

Note: (in case of redefinitions / clashes) precedence order: descending

  • arguments passed to tsc: tsc --cond:FOO1=BAR1 --cond:FOO2=BAR2

  • definition in package.json (in case of node.js):

    "cond": {
      "FOO1": "BAR1",
      "FOO2": "BAR2"
    }
  • definition in tsconfig.json or .tscrc (same json as defined above).

    • The mechanism of discovering configuration is usually starting from cwd till the root of drive and if it is nowhere to found then lastly look into the home directory before warning user and fallback to default settings (or throw "unable to locate configuration").
  • environment variables: FOO1=BAR1.

@cevek

This comment has been minimized.

Copy link

cevek commented Feb 16, 2016

Webpack has this feature in the box https://webpack.github.io/docs/list-of-plugins.html#defineplugin

new webpack.DefinePlugin({
    VERSION: JSON.stringify("5fa3b9"),
    BROWSER_SUPPORTS_HTML5: true,
    TWO: "1+1",
    "typeof window": JSON.stringify("object")
})
console.log("Running App version " + VERSION);
if(!BROWSER_SUPPORTS_HTML5) require("html5shiv");
@SomaticIT

This comment has been minimized.

Copy link

SomaticIT commented Mar 17, 2016

Hi everyone,

Conditional compilation is a must-have feature in Typescript. The idea of both runtime et precompilation time constants is also a very good idea.

But I think we should use a C#/C++ syntax but adapted to JavaScript :

Sample source

#define HOST_NODE (typeof process == "object" && process.versions && process.versions.node && process.versions.v8))
#define COMPILE_OPTIONS ["some", "default", "options"]

#if HOST_NODE
console.log("I'm running in Node.JS");
#else
console.log("I'm running in browser");
#endif

#if COMPILE_OPTIONS.indexOf("default") !== -1
console.log("Default option is configured");
#endif

Edit: Remove equals signs to keep C-style syntax as suggested by @stephanedr .

Runtime compilation

This would then emit, assuming there is no compile time substitutional value available (and targeting ES6) as:

const __tsc__HOST_NODE = (typeof process == "object" && process.versions && process.versions.node && process.versions.v8));
const __tsc__COMPILE_OPTIONS = ["some", "default", "options"];

if (__tsc__HOST_NODE) {
    console.log("I'm running in Node.JS");
} else {
    console.log("I'm running in browser");
}

if (__tsc__COMPILE_OPTIONS.indexOf("default") !== -1) {
    console.log("Default option is configured");
}

tsconfig.json configuration

Assuming you define some constants in your tsconfig.json :

{
    "defines": {
        "HOST_NODE": false,
        "COMPILE_OPTIONS": ["some", "other", "options"]
    }
}

This would then emit as :

console.log("I'm running in browser");

CLI configuration

Or if you define contants in an other way by using CLI :

$ tsc --define=HOST_NODE:true --define=COMPILE_OPTIONS:["some", "default", "options"]

This would then emit as :

console.log("I'm running in NodeJS");
console.log("Default option is configured");

Typings emitting and interpretation

Assuming you have a module designed like this :

#define HOST_NODE (typeof process == "object" && process.versions && process.versions.node && process.versions.v8))
#define COMPILE_OPTIONS ["some", "default", "options"]

export function commonFunction() { }

#if HOST_NODE
export function nodeSpecificFunction() { }
#endif

#if COMPILE_OPTIONS.indexOf("default") !== -1
export function dynamicOptionFunction() { }
#endif

export class MyClass {
    common() { }
#if HOST_NODE
    nodeSpecific() { }
#endif
}

Edit: Remove equals signs to keep C-style syntax as suggested by @stephanedr .
Edit: Added Class case as suggested by @stephanedr .

If no definitions are configured, it should be interpreted like this :

export function commonFunction(): void;
export function nodeSpecificFunction?(): void;
export function dynamicOptionFunction?(): void;
export class MyClass {
    common(): void;
    nodeSpecific?(): void;
}

Edit: Added Class case as suggested by @stephanedr .

If you define some constants in your tsconfig.json :

{
    "defines": {
        "HOST_NODE": false,
        "COMPILE_OPTIONS": ["some", "default", "options"]
    }
}

Then, it should be interpreted like this :

export function commonFunction(): void;
export function dynamicOptionFunction(): void;
export class MyClass {
    common(): void;
}

Edit: Added Class case as suggested by @stephanedr .

It allows compiler and EDIs to ignore some parts of the code based on compiler configuration.

Function-like syntax

Based on @stephanedr comments.

Assuming following sample source

#define DEBUG !!process.env.DEBUG
#if DEBUG
function _assert(cond: boolean): void {
    if (!cond)
        throw new AssertionError();
}
#define assert(cond: boolean): void _assert(cond)
#endif

type BasicConstructor = { new (...args: Object[]) => T };
#if DEBUG
function _cast<T>(type: BasicConstructor, object: Object): T {
    #assert(object instanceof type);
    return <T>object;
}
#define cast<T>(type: BasicConstructor, object: T) _cast(type, object)
#else
#define cast<T>(type: BasicConstructor, object: T) <T>object
#endif

class C {
    f(a: number) {
        #assert(a >= 0 && a <= 10);
        let div = #cast(HTMLDivElement, document.getElementById(...));
    }
}

This would then emit, assuming there is no compile time substitutional value available (and targeting ES6) as:

const __tsc__EMPTY = function () { return; };
const __tsc__DEBUG= !!process.env.DEBUG;

let __tsc__assert = __tsc__EMPTY;
if (__tsc__DEBUG) {
    function _assert(cond) {
         if (!cond)
             throw new AssertionError();
    }
    __tsc__assert = function (cond) { return _assert(cond); };
}

let __tsc__cast = __tsc__EMPTY;
if (__tsc__DEBUG) {
    function _cast(type, object) {
        __tsc__assert(object instanceof type);
        return object;
    }
    __tsc__cast = function (type, object) { return _cast(type, object); };
}
else {
    __tsc__cast = function (type, object) { return object; };
}

class C {
    f(a) {
        __tsc__assert(a >= 0 && a <= 10);
        let div = __tsc__cast(HTMLDivElement, document.getElementById(...));
    }
}

Assuming you define DEBUG constant with value true using CLI or typings.json , this would then emit as :

function _assert(cond) {
    if (!cond)
       throw new AssertionError();
}

function _cast(type, object) {
    _assert(object instanceof type);
    return object;
}

class C {
    f(a) {
        _assert(a >= 0 && a <= 10);
        let div = _cast(HTMLDivElement, document.getElementById(...));
    }
}

Now, assuming you define DEBUG constant with value false using CLI or typings.json , this would then emit as :

class C {
    f(a) {
        let div = document.getElementById(...);
    }
}

Conclusion

I think it allows a more granular conditional compilation by using the power of JavaScript.
Compiler can evaluate expressions passed by compiler directives #if #elseif...

Moreover it clearly separates (in both code-style and evaluation) the compiler directives from your code, like it used to be on C-style compiled languages.

What do you think ?
Should I start a new issue to avoid confusion ?

@stephanedr

This comment has been minimized.

Copy link

stephanedr commented Mar 18, 2016

@SomaticIT A few remarks:

1/ For people who know C/C# syntax, the "=" sign between the name and the value may be a bit disturbing. Why not keeping the C/C# syntax?

2/ It should allow something like:

class C {
    #if DEBUG
    f() {}
    #endif
}

(for sure, here DEBUG needs to evaluate to a constant to generate valid ES6 code).

3/ It should also support function-like syntax, e.g.:

#if DEBUG
function _assert(cond: boolean): void {
    if (!cond)
        throw new AssertionError();
}
#define assert(cond) _assert(cond)
#else
#define assert(cond)
#endif

#if DEBUG
function _cast<T>(type: { new (...args: Object[]) => T }, object: Object): T {
    assert(object instanceof type);
    return <T>object;
}
#define cast(type, object) _cast(type, object)
#else
#define cast(type, object) <type>object
#endif

class C {
    f(a: number) {
        assert(a >= 0 && a <= 10);
        let div = cast(HTMLDivElement, document.getElementById(...));
    }
}

The simplest to get assert() and cast() managed by Intellisense is to declare them (before their corresponding macro implementations):

/** ... */
declare function assert(cond: boolean): void;
/** ... */
declare function cast<T>(type: { new (...args: Object[]) => T }, object: Object): T;
@SomaticIT

This comment has been minimized.

Copy link

SomaticIT commented Mar 18, 2016

@stephanedr

1/ I agree with you, I edited my comment to remove = in define.

2/ I also agree with you, I edited my comment to add this exemple in Typings emitting and interpretation part.

3/ I think this case is really interesting but I think we should improve compilation emitting and typings interpretation in this particular case.
I added a Function-like syntax part. What do you think ?

This was referenced Mar 21, 2016

@stephanedr

This comment has been minimized.

Copy link

stephanedr commented Mar 21, 2016

@SomaticIT
3a/ When DEBUG=0, the call is replaced by the given text, substituting the arguments. Fine.
But when it's a non-constant expression, you call functions. Which is not the same because you still evaluate the arguments.

assert(anotherModule.aDebugFunction());
  • anotherModule may be not imported in DEBUG mode,
  • anotherModule may be compiled with DEBUG=0 at compile time, so without aDebugFunction().
  • Even if both exist, you may want to not call it in non-DEBUG mode.

3b/ I'm not sure this can always be parsed, particularly with return union types (where the type ends? where the "body" starts?).
I think we'll need a separator (=> ?).

3c/ A same function-like define can be implemented several times (here 2, but might be more). Where should we put the JSDoc (avoiding duplication)?

An alternate syntax might be:

/** ... */
#macro assert(cond: boolean): void {
    #if DEBUG
    if (!cond)
        throw new AssertionError();
    #endif
}
/** ... */
#macro cast<T>(type: BasicConstructor, object: Object): T {
    assert(object instanceof type);  // or #assert if we want to distinguish macros from normal functions.
    return <T>object;
}

Note that I'm not especially attached to a "#" syntax, so might also be:

macro assert(cond: boolean): void { ... }

[macro]
function assert(cond: boolean): void { ... }
@stephanedr

This comment has been minimized.

Copy link

stephanedr commented Mar 21, 2016

Sorry I've used the C# decorator syntax. Let's use the TypeScript's one.

@macro
function assert(cond: boolean): void { ... }
@Knagis

This comment has been minimized.

Copy link
Contributor

Knagis commented Apr 20, 2016

My project needs conditional compilation in order to support partial builds - when the user chooses that he wants to have a version of the library that only has components 1, 3, X... in it. While most of that can be supported by splitting the code into separate files, there are some cases when I need to define that certain class members are for one component only. For this I currently have a script that uses the compiler API and uses a regex to apply simple conditional text replacements in the source code.

@igabesz

This comment has been minimized.

Copy link

igabesz commented Jun 12, 2016

Some background: C/C++ and C# support preprocessors, Java doesn't. However, one can use the C++ preprocessor system in Java, see this post.

I was somewhat curious about this feature in TS and I think there are valid cases where preprocessors could be used without abusing the TS/JS dev and build process. But I'm not convinced that this should be done during the TS compilation. Gulp offers some preprocessing right now, see preprocess or gulp-preprocess or Webpack's similar feature mentioned by @cevek.

Here is a simple experimenting of mine to check out gulp-preprocess with TS. (Don't expect much.) This preprocessor seems to be kinda useful, not with all of the C++ preprocessor features though.

Currently I see these major weaknesses of external preprocessors.

  • They can lead to invalid TS code before preprocessing. This is a real pain, however, in most cases this can be solved, avoided or workarounded.
  • The TS compiler won't see the difference caused by them. Changing the interface based on build options seems to indicate bad architecture. (Or at least something being very far from the JS/TS world.)
  • Need to use an external build system (e.g. Gulp). But if preprocessors are required then there already should be a fairly complex build system.

So I think that using preprocessors in a TS environment is a very special requirement (i.e. not general) that can be solved by using already existing tools. I recommend reconsidering this feature after at least a half or one year. There are more important and way more general feature requests now.

@stephanedr

This comment has been minimized.

Copy link

stephanedr commented Aug 25, 2016

Is webpack able to remove function calls, e.g. assert() (for sure without surrounding each call with an if)?

Support of "macros" would also solve requests like #8655.

@GabrielDelepine

This comment has been minimized.

Copy link

GabrielDelepine commented Sep 9, 2016

@stephanedr
@JohnWhiteTB

Let me give you an example of the DefinePlugin of webpack (sorry if it's strictly out of the initial scare of the issue)

1/ Defined your constant in your webpack.config.js

new webpack.DefinePlugin({
    'MY_CONFIG': {
        'ENVIRONMENT': 'development'
    },
}),

2/ Make Typescript "happy" with this global constant with the custom-typings.d.ts (or any name *.d.ts)

declare class myConfig {
    ENVIRONMENT: string;
}

declare const MY_CONFIG: myConfig;

3/ use it anywhere you want

if ('development' === MY_CONFIG.ENVIRONMENT) {
    console.log('development environment');
} else {
    console.log('NOT the development environment');
}

in the JS transpiled code, it will be transformed to :

if (true) {
    console.log('development environment');
} else {
    console.log('NOT the development environment');
}

and after the minification step, it will become :

    console.log('development environment');

"et voilà !"

@stephanedr

This comment has been minimized.

Copy link

stephanedr commented Sep 9, 2016

@GabrielDelepine
What about support of "macros"?
The typical example is assert() and cast(), where as TS code like this:

function f(a: number) {
    assert(a >= 0 && a <= 10);
    let div = cast(HTMLDivElement, document.getElementById(...));
}

would be finally generated as follows in release mode:

function f(a) {
    let div = document.getElementById(...);
}

The point here is that if (MY_CONFIG.ENVIRONMENT ...) are embedded into assert() and cast().

@domchen

This comment has been minimized.

Copy link
Contributor

domchen commented Oct 13, 2016

Hi guys, check out this project : https://github.com/domchen/typescript-plus . It is an enhanced version of the original typescript compiler, which provides conditional compilation.

You can use the defines option to declare global variables that the compiler will assume to be constants (unless defined in scope). Then all the defined global variables will be replaced with the corresponding constants. For example:

tsconfig.json:

{
    "compilerOptions": {
        "defines": {
            "DEBUG": false,
            "LANGUAGE": "en_US"
        }
    }
}

TypeScript:

declare var DEBUG:boolean;
declare var LANGUAGE:string;

if (DEBUG) {
    console.log("DEBUG is true");
}

console.log("The language is : " + LANGUAGE);

function someFunction():void {
    let DEBUG = true;
    if (DEBUG) {
        console.log("DEBUG is true");
    }
}

JavaScript:

if (false) {
    console.log("DEBUG is true");
}

console.log("The language is : " + "en_US");

function someFunction() {
    var DEBUG = true;
    if (DEBUG) {
        console.log("DEBUG is true");
    }
}

As you can see, the second if(DEBUG) in someFunction is not replaced because it is defined in scope.

Note that the compiler does not dropping the unreachable code, because it is can be easily done by other tools like UglifyJS or Google Closure Compiler.

@isiahmeadows

This comment has been minimized.

Copy link
Contributor

isiahmeadows commented Apr 24, 2017

Here's my idea:

  • Introduce @@guard(cond) { ... } and @@guard(version, range) { ... } constructs that works in most declarative contexts, provided its members are syntactically correct. Additionally, within that block, if it's not matched, only check for balanced parentheses and brackets within template interpolations and outside strings. The top brackets are optional, and each inner expression must be a constant expression (i.e. same restriction as with const enums).

    The first form, @@guard(cond) { ... }, evaluates to its block if and only if cond evaluates to true.

    The second form, @@guard(version, range) { ... }, evaluates to its block if and only if version is satisfied with a semver range.

    // TS define
    @@guard(env ts.version, 'version') { ... }
    
    // Compile option
    @@guard(env ts.options.module) { ... }
    
    // User define
    @@guard(env foo.bar) { ... }
    ```0
    
  • Introduce a env customOpt construct that evaluates to the constant result of that option. The option name must be a valid identifier, custom ones must not be ts, and it returns a constant expression.

This is intentionally made such that it could be resolved at parse time, too.

@kevmeister68

This comment has been minimized.

Copy link

kevmeister68 commented May 22, 2017

Cool me old-school (actually, I am old-school), so my vote is for simple, C++ style pre-processor [like] directives to control conditional compilation, with the ability to control those values from the tsconfig.json or (in the case of IDEs like Visual Studio), from the IDE.

I am not a fan of trying to make conditional compilation "look like" program code, and I believe that the most flexible form of conditional compilation, which allows switching in eg. different modules, classes, interfaces, or whatever, needs that conditional compilation to sit "in front" of the TS compilation.

I use C# heavily and I like the ConditionalAttribute behaviour that they have for cases such as removing Debug.WriteLine calls, and I can see it working nicely for Assert-style semantics also, but it is a much more limited capability in my mind.

@RobertoMalatesta

This comment has been minimized.

Copy link

RobertoMalatesta commented Jun 19, 2017

I'm hitting this problem with a lot of universal code that could be made working in the browser and in node simply by having a:
#IFDEF _CONDITION_SPECIFIED_IN_TSCONFIG
... do something -say- browser-specific, like open a pop-up
#ELIF
... do something that has a meaning only in node, like post the message to an AMQP queue
#ENDIF

Not having this forces to have two quasi-identical classes/functions in the client and the server code.
This, happening in many places, poses serious maintenance problems on a medium code size (>50k lines).

With a tsconfig-driven #IFDEF we could use two separate tsconfigs and each parted build could reference libraries that otherwise would generate globs of compilation errors on the other.
I'd say it's rather simple to implement, since active code can be inferred before compilation.

Badly badly needed indeed.
I see this is lingering since 2014 with ish #449

R

@aleksey-bykov

This comment has been minimized.

Copy link

aleksey-bykov commented Oct 24, 2017

so on our project we need to hide some of the latest features from the release version this is what we do:

// release branch
declare global {
     declare const enum NotFor { Release = 0 }
}
if (NotFor.Release) {
    // here comes the feature which is hidden in release 
}

in the ongoing development branch we use

// develop branch
declare global {
     declare const enum NotFor { Release = 1 }
}
if (NotFor.Release) {
    // here comes the feature visible in the 
}

it's ugly as hell, but we eat it since don't have anything better

@kitsonk

This comment has been minimized.

Copy link
Contributor Author

kitsonk commented Oct 24, 2017

Not that this solves it at a TypeScript level, but we (@dojo) have continued to implement a solution that we can use to give us build time optimisation as well. We use a the has API to provide in code feature switching. We then have a webpack loader that will statically replace constructs (as well as elide imports) based on asserted feature flags.

This gives us run-time feature switching we want but also, by replacing the statically asserted features, we can remove imports via the loader and uglify can perform dead code removal inside of modules when the bundle is minified.

@porfirioribeiro

This comment has been minimized.

Copy link

porfirioribeiro commented Nov 27, 2017

Maybe implement it like Haxe: https://code.haxe.org/category/beginner/conditional-compilation.html
Work's good and it's simple and easy to use

@gwicksted

This comment has been minimized.

Copy link
Contributor

gwicksted commented Nov 28, 2017

Suggestion if implementing preprocessor directives

If this is to be implemented (I'm on the fence -- see below), take a look at the npm package preprocessor. It has desirable syntax mentioned by @JohnWeisz and plenty of features. The only change I would make is being able to configure the path resolver. Currently it is fixed as:

Includes (always relative to the baseDirectory, defaults to ".")

But it would be nice if it used require.resolve() or a callback as an alternative.

Why do we need this?

Common uses for an #ifdef are:

  • toggle/change an import,
  • prevent a reference (to that import); or,
  • toggle a behavior (such as trace logging).

Common uses for a #define are usually complimentary to the above:

  • later used in #ifdefs without adding constants to code
  • configuration (where constants are often more appropriate)
  • small macro functions (which are inherently dangerous)

And while conditional compilation is a convenient hack, what you really need to keep the code clean long-term is IoC + DI

Currently supported alternative

DI has already been implemented in TypeScript via ES7 decorators (TypeScript link). This prevents import statements to missing files which, unlike require statements, cannot be placed within conditionals. Bonus: it also decouples your code and can reduce the number of imports per module.

DI + different sets of source and configuration files covers the majority of cases where conditional compilation would be used.

Steps involved

  1. Install an IOC container such as decoration-ioc (based on the IOC in vscode) or typescript-ioc (which has great readme docs).

(disclaimer: I have not used either in production)

  1. Enable decorators in your tsconfig.json via:
    "experimentalDecorators": true,
    "emitDecoratorMetadata": true
  1. You would then configure different sets of sources via include + differing hierarchies of tsconfigs using extends.

Admittedly it's quite a bit of initial complexity for a handful of #ifdefs but there are major long-term maintainability advantages over conditional compilation and it becomes very pleasant to work with over time (my experience with DI is primarily in C# but the concept is universal).

@gwicksted

This comment has been minimized.

Copy link
Contributor

gwicksted commented Nov 28, 2017

Input greatly appreciated -- @Nimikolh if you get a chance, please expand on why you believe DI (or the npm preprocessor package) are less desirable than other proposals. I'm looking into this setup for an existing codebase so all input is welcome!

@Nemikolh

This comment has been minimized.

Copy link

Nemikolh commented Nov 28, 2017

@gwicksted The main reason I would like to see this supported as a first class citizen feature in TypeScript is to get the full power of the type system. Being able to turn on / off certain features in my code or to predefine a set of targets can indeed be done easily with any preprocessor. However this is not the only issue that needs to be addressed in my opinion. For instance getting your IDE (or an editor using ts-server) understand that it should be run after the preprocessor is also a valid concern.

So now, would using DI be the solution to rule them all? I don't think so. Assuming the conditional compilation design is driven by the use cases, one could be "how to get the full assistance of the type system while using React Native" for instance. In that example, using DI to resolve a cross-platform component is just overkill. Would conditional compilation be sufficient to make that experience better? Maybe not, given how import works in that space. However it would definitely helps.

So the TLDR; is multiple platforms.

@Pauan

This comment has been minimized.

Copy link

Pauan commented Nov 29, 2017

@gwicksted We are creating a large TypeScript library, and the reason we want conditional compilation is so that we can have two different builds of our library:

  • A build with expensive run-time checking + asserts (for extra safety and correctness)

  • A build without the run-time checks or asserts (for extra performance)

Our users can use the "debug" build while developing, and then when their code is finished, they can switch to the "production" build for extra speed.

This is a common pattern in many languages and libraries (e.g. React and Angular both do this).

We can accomplish this right now with runtime switches, but ideally we would want no runtime overhead at all for the production build. Also, with runtime switches all of the code is included even though the runtime asserts aren't actually being run. So there is a lot of dead code (which bloats up the file size).

We could also accomplish this by having two separate code bases, one for the "debug" build and one for the "production" build. But that's an incredibly bad idea: trying to manage two code bases and ensuring that both code bases remain in sync is impractical.

We could try to use something like preprocessor, except those tools haven't been designed for TypeScript, so they don't understand TypeScript code. In many cases they don't even understand ES6 code!

Conditional compilation gives us what we want with no overhead. Runtime switches, or DI, or whatever else doesn't work as well.

@gwicksted

This comment has been minimized.

Copy link
Contributor

gwicksted commented Nov 29, 2017

@Pauan That's understandable since you will have very coupled asserts intertwined with your logic.

Fortunately for you, uglify-es has drop_console which will get rid of the entire console.assert statement including parameters so you can emit a second release-mode file -- which is likely going to be minified already -- by piping it through uglify. Zero overhead, no dead code, and it understands ES6+ code and source maps.

@Pauan

This comment has been minimized.

Copy link

Pauan commented Nov 29, 2017

@gwicksted We need to be able to remove our own custom asserts, and runtime checking code, we don't use console.assert at all.

Also, we don't distribute our library as a bundled + minified + optimized blob of JS code. We distribute our library as ES6 modules, and our users are then responsible for bundling.

So we can't rely upon a minifier which is run in the final stages of compilation, because we can't control how our users choose to compile their code.

@gwicksted

This comment has been minimized.

Copy link
Contributor

gwicksted commented Nov 29, 2017

@Pauan that makes sense as does not wanting a third party step like a preprocessor/minifier involved.

@RichiCoder1

This comment has been minimized.

Copy link

RichiCoder1 commented Nov 29, 2017

@Pauan Correct me if I'm wrong, but doesn't React get around this my isolating non-production code in such a way that Webpack and Browserify can shake it out automatically? By wrapping that code in process.env.NODE_ENV?

@gwicksted

This comment has been minimized.

Copy link
Contributor

gwicksted commented Nov 29, 2017

@RichiCoder1 I know dead code elimination + tree shaking can do what you are describing and that several transpilers (or a combination of them) are capable of this. Tree shaking is easy thanks to ES6 imports and it looks like dead code elimination has finally become mature enough to use in production.

I'm more familiar with browserify so I'll take a crack at this with tinyifyify + envify at some point to see how it goes.

At @Pauan 's defense, they want the benefits of this without imposing restrictions on the end user's build system. If TS could perform tree shaking and dead code elimination natively then I'm sure we're all ok with that over a preprocessor!

@isiahmeadows

This comment has been minimized.

Copy link
Contributor

isiahmeadows commented Nov 30, 2017

@RichiCoder1 IIRC, React uses a special __DEV__ global variable that they shake out via either UglifyJS, Webpack, or Babel. (I can't remember exactly which one they shake it out with, though.)

@Telem

This comment has been minimized.

Copy link

Telem commented Mar 16, 2018

I'd like to add another use-case for this feature that didn't seem to get mentionned yet : shaking out log messages from production builds. The goal would be to allow devs to provide verbose logging for development builds that does not make its way to production builds, thus reducing the package size.

It certainly does not need the full power (and complexity) of conditional compiles, but I am not certain that dead code removal would be able to solve this on its own either.

@gwicksted

This comment has been minimized.

Copy link
Contributor

gwicksted commented Mar 16, 2018

@Telem yes, it would be much easier with native language support rather than relying on the third-party uglify-es + drop_console (which not only drops asserts that @Pauan requires but also logs).

I have a codebase with custom logging wrappers so shaking them out would require not using those directly (eg. by replacing console) as well as modifying tslint rules to allow their usage without warning.

Would also be nice if it wasn't just conditional compilation but also macro functions so non-error logging could include the file and line without specifying them explicitly (verbose) nor generating a stack trace (slow + requires filtering) and would not require a map file to point back to the original TypeScript source from the minified js bundle.

@JonathanWilbur

This comment has been minimized.

Copy link

JonathanWilbur commented Jun 1, 2018

I upvote this as well. My use case is making a TypeScript library that compiles to both NodeJS and Browser dialects. In particular, I would like the ability to alternate between Uint8Array (which works in the web browser) and Buffer, which works in NodeJS.

I imagine other people would also benefit from the ability to compile for specific browsers as well.

@xdvarpunen

This comment has been minimized.

Copy link

xdvarpunen commented Jul 25, 2018

Upcoming deno has targeted the code to be usable both in node and browser. The creator of NodeJS wanted to remove the whole separation of node and browser.

In webpack there is target option to set particular target platform for the code.

Currently to separate to different platform would be by having two separate project with their own tsconfig files and webpack files. If you have code that is reusable, but need separate platform types, then having one common folder that is imported to platform specific index file or main would be the possible solution.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment