|
| 1 | +# vs. CMake |
| 2 | + |
| 3 | +GYP was originally created to generate native IDE project files (Visual Studio, Xcode) for building [Chromium](http://www.chromim.org). |
| 4 | + |
| 5 | +The functionality of GYP is very similar to the [CMake](http://www.cmake.org) |
| 6 | +build tool. Bradley Nelson wrote up the following description of why the team |
| 7 | +created GYP instead of using CMake. The text below is copied from |
| 8 | +http://www.mail-archive.com/[email protected]/msg11029.html |
| 9 | + |
| 10 | +``` |
| 11 | +
|
| 12 | +Re: [webkit-dev] CMake as a build system? |
| 13 | +Bradley Nelson |
| 14 | +Mon, 19 Apr 2010 22:38:30 -0700 |
| 15 | +
|
| 16 | +Here's the innards of an email with a laundry list of stuff I came up with a |
| 17 | +while back on the gyp-developers list in response to Mike Craddick regarding |
| 18 | +what motivated gyp's development, since we were aware of cmake at the time |
| 19 | +(we'd even started a speculative port): |
| 20 | +
|
| 21 | +
|
| 22 | +I did an exploratory port of portions of Chromium to cmake (I think I got as |
| 23 | +far as net, base, sandbox, and part of webkit). |
| 24 | +There were a number of motivations, not all of which would apply to other |
| 25 | +projects. Also, some of the design of gyp was informed by experience at |
| 26 | +Google with large projects built wholly from source, leading to features |
| 27 | +absent from cmake, but not strictly required for Chromium. |
| 28 | +
|
| 29 | +1. Ability to incrementally transition on Windows. It took us about 6 months |
| 30 | +to switch fully to gyp. Previous attempts to move to scons had taken a long |
| 31 | +time and failed, due to the requirement to transition while in flight. For a |
| 32 | +substantial period of time, we had a hybrid of checked in vcproj and gyp generated |
| 33 | +vcproj. To this day we still have a good number of GUIDs pinned in the gyp files, |
| 34 | +because different parts of our release pipeline have leftover assumptions |
| 35 | +regarding manipulating the raw sln/vcprojs. This transition occurred from |
| 36 | +the bottom up, largely because modules like base were easier to convert, and |
| 37 | +had a lower churn rate. During early stages of the transition, the majority |
| 38 | +of the team wasn't even aware they were using gyp, as it integrated into |
| 39 | +their existing workflow, and only affected modules that had been converted. |
| 40 | +
|
| 41 | +2. Generation of a more 'normal' vcproj file. Gyp attempts, particularly on |
| 42 | +Windows, to generate vcprojs which resemble hand generated projects. It |
| 43 | +doesn't generate any Makefile type projects, but instead produces msvs |
| 44 | +Custom Build Steps and Custom Build Rules. This makes the resulting projects |
| 45 | +easier to understand from the IDE and avoids parts of the IDE that simply |
| 46 | +don't function correctly if you use Makefile projects. Our early hope with |
| 47 | +gyp was to support the least common denominator of features present in each |
| 48 | +of the platform specific project file formats, rather than falling back on |
| 49 | +generated Makefiles/shell scripts to emulate some common abstraction. CMake by |
| 50 | +comparison makes a good faith attempt to use native project features, but |
| 51 | +falls back on generated scripts in order to preserve the same semantics on |
| 52 | +each platforms. |
| 53 | +
|
| 54 | +3. Abstraction on the level of project settings, rather than command line |
| 55 | +flags. In gyp's syntax you can add nearly any option present in a hand |
| 56 | +generated xcode/vcproj file. This allows you to use abstractions built into |
| 57 | +the IDEs rather than reverse engineering them possibly incorrectly for |
| 58 | +things like: manifest generation, precompiled headers, bundle generation. |
| 59 | +When somebody wants to use a particular menu option from msvs, I'm able to |
| 60 | +do a web search on the name of the setting from the IDE and provide them |
| 61 | +with a gyp stanza that does the equivalent. In many cases, not all project |
| 62 | +file constructs correspond to command line flags. |
| 63 | +
|
| 64 | +4. Strong notion of module public/private interface. Gyp allows targets to |
| 65 | +publish a set of direct_dependent_settings, specifying things like |
| 66 | +include_dirs, defines, platforms specific settings, etc. This means that |
| 67 | +when module A depends on module B, it automatically acquires the right build |
| 68 | +settings without module A being filled with assumptions/knowledge of exactly |
| 69 | +how module B is built. Additionally, all of the transitive dependencies of |
| 70 | +module B are pulled in. This avoids their being a single top level view of |
| 71 | +the project, rather each gyp file expresses knowledge about its immediate |
| 72 | +neighbors. This keep local knowledge local. CMake effectively has a large |
| 73 | +shared global namespace. |
| 74 | +
|
| 75 | +5. Cross platform generation. CMake is not able to generate all project |
| 76 | +files on all platforms. For example xcode projects cannot be generated from |
| 77 | +windows (cmake uses mac specific libraries to do project generation). This |
| 78 | +means that for instance generating a tarball containing pregenerated |
| 79 | +projects for all platforms is hard with Cmake (requires distribution to |
| 80 | +several machine types). |
| 81 | +
|
| 82 | +6. Gyp has rudimentary cross compile support. Currently we've added enough |
| 83 | +functionality to gyp to support x86 -> arm cross compiles. Last I checked |
| 84 | +this functionality wasn't present in cmake. (This occurred later). |
| 85 | +
|
| 86 | +
|
| 87 | +That being said there are a number of drawbacks currently to gyp: |
| 88 | +
|
| 89 | +1. Because platform specific settings are expressed at the project file |
| 90 | +level (rather than the command line level). Settings which might otherwise |
| 91 | +be shared in common between platforms (flags to gcc on mac/linux), end up |
| 92 | +being repeated twice. Though in fairness there is actually less sharing here |
| 93 | +than you'd think. include_dirs and defines actually represent 90% of what |
| 94 | +can be typically shared. |
| 95 | +
|
| 96 | +2. CMake may be more mature, having been applied to a broader range of |
| 97 | +projects. There a number of 'tool modules' for cmake, which are shared in a |
| 98 | +common community. |
| 99 | +
|
| 100 | +3. gyp currently makes some nasty assumptions about the availability of |
| 101 | +chromium's hermetic copy of cygwin on windows. This causes you to either |
| 102 | +have to special case a number of rules, or swallow this copy of cygwin as a |
| 103 | +build time dependency. |
| 104 | +
|
| 105 | +4. CMake includes a fairly readable imperative language. Currently Gyp has a |
| 106 | +somewhat poorly specified declarative language (variable expansion happens |
| 107 | +in sometimes weird and counter-intuitive ways). In fairness though, gyp assumes |
| 108 | +that external python scripts can be used as an escape hatch. Also gyp avoids |
| 109 | +a lot of the things you'd need imperative code for, by having a nice target |
| 110 | +settings publication mechanism. |
| 111 | +
|
| 112 | +5. (Feature/drawback depending on personal preference). Gyp's syntax is |
| 113 | +DEEPLY nested. It suffers from all of Lisp's advantages and drawbacks. |
| 114 | +
|
| 115 | +-BradN |
| 116 | +``` |
0 commit comments