Annotation of wikisrc/pkgsrc/gcc.mdwn, revision 1.7

1.1       gdt         1: On many systems pkgsrc supports, gcc is the standard compiler.  In
                      2: general, different versions of each OS have different gcc versions,
1.2       gdt         3: and some packages require newer gcc versions, in order to support
1.1       gdt         4: newer language standards (e.g. c++11, written in the style of
                      5: USE_LANGUAGES), or because older versions don't work (infrequently).
                      7: This page discusses issues related to version selection, and intends
                      8: to be a design document for how pkgsrc should address this problem, to
                      9: be converted into historical design rationale once implemented.  It
                     10: freely takes content from extensive mailinglist discussions, and
                     11: attempts to follow the rough consensus that has emerged.
                     13: ## Base system gcc vs pkgsrc gcc
                     15: Systems using gcc (e.g. NetBSD) have a compiler as /usr/bin/gcc, and
                     16: this is usable by pkgsrc without any bootstrapping activity.  One can
                     17: build gcc versions (typically newer versions) from pkgsrc, resulting
                     18: in a compiler within ${PREFIX}, e.g. /usr/pkg/gcc6/bin/gcc.  This
                     19: compiler can then be used to compile other packages.
1.3       gdt        21: The Issue with using base system gcc is typically that it is too old,
1.6       gdt        22: such as gcc 4.5 with NetBSD 6, which cannot compile c++11.  Another
                     23: example is gcc 4.8 with NetBSD 7.  While this can compile most c++11
                     24: programs, it cannot be used for firefox or glibmm (and therefore any
                     25: package that links against glibmm).
1.1       gdt        26: 
                     27: Issues when using pkgsrc gcc are that
                     29:   - it must be bootstrapped, requiring compiling a number of packages
                     30:     with the system compiler
                     31:   - C++ packages that are linked together should be built with the
                     32:     same compiler, because the standard library ABI is not necessarily
                     33:     the same for each compiler version
                     34:   - While C packages can be built with mixed versions, the binary
                     35:     should be linked with the higher version because the support
                     36:     library is backwards compatible but not forward compatible.
                     38: ## Specific constraints and requirements
                     40: This section attempts to gather all the requirements.
                     42:   - By default, pkgsrc should be able to build working packages, even
                     43:     for packages that need a newer compiler than that provided in the
                     44:     base system.
                     46:   - The set of packages that are needed when building a bootstrap
                     47:     compiler should be minimized.
                     49:   - All packages that use C should have final linking with the highest
                     50:     version used in any included library.
1.2       gdt        52:   - All packages that use C++ should be built with the same compiler
1.3       gdt        53:     version.  Because these in the general case may include C, the
                     54:     version used for C++ must be at least as new as the version used
                     55:     for any used C package.
1.2       gdt        56: 
1.1       gdt        57:   - pkgsrc should avoid building gcc unless it is more or less
                     58:     necessary to build packges.  (As an example, if the base system
                     59:     gcc can build c99 but not c++11, building a c99-only program
                     60:     should not trigger building a gcc version adequate for c++11.)
1.2       gdt        62:   - The compiler selection logic should work on NetBSD 6 and newer,
                     63:     and other systems currently supported by pkgsrc, including in-use
                     64:     LTS GNU/Linux systems.  It should work on systems that default to
                     65:     clang, when set to use GCC, at least as well as the current
                     66:     scheme.  It is desirable for this logic to work on NetBSD 5.
1.1       gdt        67: 
                     68:   - The compiler selection logic should be understandable and not brittle.
                     70: ## Design
                     72: The above requirements could in theory be satisfied in many ways, but
1.2       gdt        73: most of them are too complicated.  We present a design that aims to be
                     74: sound while mimimizing complexity.
1.1       gdt        75: 
                     76:   - Packages declare what languages they need, with c++, c++11, and
1.2       gdt        77:     c++14 being expressed differently.  (This is exactly current
                     78:     practice and just noted for completeness.)
1.1       gdt        79: 
                     80:   - The package-settable variable GCC_REQD will be used only when a
                     81:     compiler that generally can compile the declared language version
1.2       gdt        82:     is insufficient.  These cases are expected to be relatively rare;
                     83:     an example is firefox that is in c++ (but not c+11) and needs gcc
                     84:     4.9.
1.1       gdt        85: 
                     86:   - A user-settable variable PKGSRC_GCC_VERSION will declare the
                     87:     version of gcc to be used for C programs, with an OS- and
                     88:     version--specific default.
1.6       gdt        90:   - A user-settable variable PKGSRC_GXX_VERSION will declare the
                     91:     version of gcc to be used for all C++ programs, again with an OS-
                     92:     and version-specific default.  It must be at least
                     93:     PKGSRC_GCC_VERSION.
1.1       gdt        94: 
                     95:   - Each of c99, c++, c++11, and c++14 will be associated with a
                     96:     minimum gcc version, such that almost all programs declaring that
                     97:     language can be built with that version.  (This avoids issues of
                     98:     strict compliance with c++11, which requires a far higher version
                     99:     of gcc than the version required to compile almost all actual
                    100:     programs in c++11.)
                    102:   - The minimum version inferred from the language tag will be
                    103:     combined with any GCC_REQD declarations to find a minimum version
                    104:     for a specific package.  If that is greater than
                    105:     PKGSRC_GCC_VERSION (programs using only C) or PKGSRC_GXX_VERSION,
                    106:     package building will fail.  We call the resulting
                    107:     PKGSRC_GCC_VERSION or PKGSRC_GXX_VERSION the chosen version.
1.3       gdt       109:   - When building a program using C or C++, if the chosen version is
                    110:     not provided by the base system, and the chosen version is not
1.1       gdt       111:     installed via pkgsrc, then it (and its dependencies) will be built
                    112:     from pkgsrc in a special bootstrap mode.  When building in
                    113:     bootstrap mode, the version selection logic is ignored and the
                    114:     base system compiler is used.  Consistency and reproducible builds
                    115:     require that a package built with the normal prefix must be the
                    116:     same whether built because of compiler bootstrapping or normal
                    117:     use.
                    119:     There are thus two choices for dealing with bootstrapping.  One is
                    120:     to use a distinct prefix, which will ensure that all packages that
                    121:     are part of the compiler bootstrap will not be linked into normal
                    122:     pkgsrc programs.  This implies that any dependencies of gcc may
                    123:     exist twice, once in bootstrap mode and once if built normally.  A
                    124:     gcc version itself will be built twice, if it is desired for
                    125:     regular use.  This double building and the complexity of a second
                    126:     prefix are the negatives of this approach.
                    128:     The other choice is to mark gcc and all depending packages as used
                    129:     for compiler bootstrapping, and to always build those with the
                    130:     base compiler.  We use the package-settable variable
                    131:     PKGSRC_GCC_BOOTSTRAP=yes to denote this.  The negative with this
                    132:     approach is possible inconsistency with gcc's dependencies being
                    133:     built with the base compiler and used later.
1.2       gdt       135:   - We hope that the chosen version can be built using the base system
                    136:     version, and hope to avoid multi-stage bootstrapping.
1.1       gdt       138:   - We expect that any program containing C++ will undergo final
                    139:     linking with a C++ compiler.  This is not a change from the
                    140:     current situation.
                    142: ## Remaining issues
1.6       gdt       144: ### gcc dependencies introduction
1.1       gdt       145: 
                    146: Because gcc can have dependencies, there could be packages built with
                    147: the system compiler that are then later used with the chosen version.
                    148: For now, we defer worrying about these problems (judging that they
                    149: will be less serious than the current situation where all c++11
                    150: programs fail to build on NetBSD 6).
1.6       gdt       152: \todo: Change gcc 4.8 and 4.9 to enable gcc-inplace-math by default.
1.1       gdt       154: \todo: Analyze what build-time and install-time dependencies actually
1.6       gdt       155: exist.  Include old GNU/Linux in this analysis.
                    157: \todo: Consider if dropping nls would help.  (On NetBSD, it seems that
                    158: base system libraries are used, so it would not help.)
1.7     ! gdt       160: \todo: Consider failing if optins that we want one way are another,
        !           161: when bootstrapping.
        !           162: 
1.6       gdt       163: ### managing gcc dependencies
                    165: There are multiple paths forward.
                    167: \todo Choose one.  Straw proposal is "Don't worry" and recursive
                    168: variable for the initial implementation.
                    170: #### Separate prefix
                    172: Build compilers in a separate prefix, or a subprefix, so that the
                    173: compiler and the packages needed to build it will not be used by any
                    174: normal packages.  This completely avoids the issue of building a
                    175: package one way in bootstrap and another not in bootstrap, at the cost
                    176: of two builds and writing the separate-prefix code.
                    178: #### Don't worry
                    180: Don't worry that packages used to bootstrap the needed compiler are
                    181: compiled with an older compiler.  Don't worry that they might be
                    182: different depending on build order.  If we have an actual problem,
                    183: deal with it.  This requires choosing an approach to omit compiler
                    184: selection logic when building the compiler:
                    186: ##### Mark bootstrap packages
                    188: Mark packages used to build gcc as PKGSRC_GCC_BOOTSTRAP=yes.
                    189: Conditionalize this on OPSYS if necessary.  Don't force the compiler
                    190: if this is set.
1.1       gdt       191: 
1.6       gdt       192: ##### Pass a recursive variable
                    194: As above, but set PKGSRC_GCC_BOOTSTRAP=yes in the evniroment of the
                    195: call to build the compiler, so that all dependencies inherit
                    196: permission to skip compiler selection logic.  (Alternatively, use some
                    197: other mechanism such as passing a make variable explicitly.)
1.1       gdt       198: 
1.3       gdt       199: ### Differing GCC and GXX versions
                    201: Perhaps it is a mistake to allow the chosen GCC and GXX versions to
                    202: differ.  If we require them to be the same, then essentially all
                    203: systems with a base system compiler older than gcc 5 will have to
                    204: bootstrap the compiler.  For now, we allow them to differ and will
                    205: permit the defaults to differ.
1.6       gdt       207: ### gcc versions and number of buildable packages
                    209: A gcc version that is too old will not build a number of packages.
                    210: Anything older than 4.8 fails for c++11.  4.8 fails on some c++11
                    211: packages, such as firefox and glibmm.
                    213: A version that is too new also fails to build packages.  Analyses
                    214: posted to tech-pkg indicate that 5 is close to 4.9 in the number of
                    215: packages built, and that moving to 6 causes hundreds of additional
                    216: failures.
                    218: Therefore, the current answer to "What is the best version to use" is
                    219: 5.
                    221: \todo Check this with Jason Bacon.
1.1       gdt       223: ### Default versions for various systems
1.3       gdt       225: Note that if for any particular system's set of installed packages (or
                    226: bulk build), a newer gcc has to be built, it does not hurt to have
                    227: built it earlier.
1.1       gdt       228: 
                    229: When the base system is old (e.g., gcc 4.5 in NetBSD 6, or 4.1, in
                    230: NetBSD 5), then it is clear that a newer version must be built.  For
                    231: these, PKGSRC_GXX_VERSION should default to a newish gcc, avoiding
1.4       gdt       232: being so new as to cause building issues.  Currently, gcc5 is probably
                    233: a good choice, with gcc6 compiling significantly but not vastly fewer
                    234: packages.  PKGSRC_GCC_VERSION should probably default to the system
                    235: version if it can build all C99 programs, or match PKGSRC_GXX_VERSION,
                    236: if the system version is too old.  Perhaps gcc 4.5 would be used, but
                    237: 4.1 not used.  \todo Discuss.
1.1       gdt       238: 
                    239: When the base system is almost new enough, the decision about the
                    240: default is more complicated.  A key example is gcc 4.8, found in
1.2       gdt       241: NetBSD 7.  Firefox requires gcc 4.9, and all programs using c++14 also
1.1       gdt       242: need a newer version.  One options is to choose 4.8, resulting in
                    243: firefox failing, as well as all c++14 programs.  Another is to choose
                    244: 4.9, but this makes little sense because c++14 programs will still
                    245: fail, and the general rule of moving to the most recent
                    246: generally-acceptable version applies, which currently leads to gcc6.
                    247: This is in effect a declaration that "almost new enough" does not
                    248: count as new enough.  Thus the plan for NetBSD 7 is to set
1.4       gdt       249: PKGSRC_GCC_VERSION to 4.8 and PKGSRC_GXX_VERSION to 5.
1.1       gdt       250: 
1.4       gdt       251: When the base system is new enough, e.g. gcc 5, 6 or 7 it should
                    252: simply be used.  By "new enough", we mean that almost no programs in
1.6       gdt       253: pkgsrc fail to build with it (because it is too old), which implies
                    254: that it supports (almost all) C++14 programs.  Our current definiton
                    255: of new enough is gcc 5.
1.1       gdt       256: 
1.2       gdt       257: ### Limited mixed versions
                    259: One approach would be to allow limited mixed versions, where
                    260: individual programs could force a specific version to be bootstrapped
                    261: and used, so that e.g. firefox could use 4.9 even though most programs
                    262: use 4.8, which is what happens now on NetBSD 7.  This would rely on
                    263: being able to link c++ with 4.9 including some things built with 4.8
                    264: (which is done presently).  However, this approach would become
                    265: unsound with a library rather than an end program.  We reject this as
                    266: too much complexity for avoiding building a newer compiler in limited
                    267: situations.
1.1       gdt       269: ### Fortran
                    271: Fortran support is currently somewhat troubled..  It seems obvious to
                    272: extend to PGKSRC_GFORTRAN_VERSION, and have that match
                    273: PKGSRC_GCC_VERSION or PKGSRC_GXX_VERSION, but the Fortran situation is
1.6       gdt       274: not worsened by the above design.
                    276: When building a gcc version, we get gfortran.  Perhaps, because of
                    277: fortran, we should require a single version, vs a C and a C++ version.
                    279: \todo Discuss.
                    281: ### C++ programs used by C programs
                    283: The choice of one version for C++ and one for C (e.g. 5, 4.8 on
                    284: netbsd-7) breaks down if a C program links against a library that is
                    285: written in C++ but provides a C API, because we still need the C++
                    286: version's stdlib.
                    288: \todo Define a variable for such packages to have in their buildlink3,
                    289: which will not add c++ to USE_LANGUAGES but will force
                    290: PKGSRC_GXX_VERSION to be used.  Or decide that this is a good reason
                    291: to really just have one compiler version.
1.1       gdt       292: 
                    293: ## Path forward
                    295:  - Modify all gcc packages to have minimal dependencies, and to add
                    296:    PKGSRC_GCC_BOOTSTRAP.
                    298:  - Modify the compiler selection logic to do nothing if
                    299:    PKGSRC_GCC_BOOTSTRAP is set.
                    301:  - Modify the compiler selection logic for LANGUAGES= to fail if
                    302:    PKGSRC_GCC_VERSION/PKGSRC_GXX_VERSION is not new enough.
                    304:  - Modify the compiler selection logic for GCC_REQD to fail if the
                    305:    version of GCC/GXX is not new enough.
                    307:  - Decide on defaults.  The straw proposal is that PKGSRC_GCC_VERSION
1.5       gdt       308:    is the base system version if >= 4.5 (or 4.4?), and otherwise 5,
1.1       gdt       309:    and that PKGSRC_GXX_VERSION is the base system version if >= 5, and
1.5       gdt       310:    otherwise 5.
1.3       gdt       311: 
                    312: ### Later steps
                    314:  - Address fortran.

CVSweb for NetBSD wikisrc <> software: FreeBSD-CVSweb