Its `__has_declspec_attribute()` is not reliable. For instance,
while `__has_declspec_attribute(noalias)` is true but 'noalias'
attribute is warned as unknown.
It is reported that on a system of i386 System V ABI, GCC returns 8 for
__alignof__(double). OTOH the ABI defines alignments of double to be 4,
and ISO/IEC 9899:2011 reads that _Alignof(double) shall return 4 on such
machine. What we want in ruby is 4 instead of 8 there. We cannot use
__alignof__.
Additionally, both old GCC / old clang return 8 for _Alignof(double) on
such platforms. They are their bugs, and already fixed in recent
versions. But we have to support older compilers for a while. Shall
check sanity of _Alignof.
Availability of `alignas` is checked in include/ruby/internal/stdalign.h
now. That does not need this configure check. Also as commented in the
header, we see `_Alignas` being inadequate for our purpose.
AC_PROG_CXX checks for several C++ compilers INCLUDING g++. When none
of them were found it sets the CXX variable to be g++. This doesn't
make any sense. Absense of g++ has already been made sure.
Because we don't want insanity (that's the whole reason we test the
environment using autoconf), we need to swipe such insane variable out.
The previous attempt to fix this in
b87df1bf24 reversed the argument
order instead of just fixing the quote escaping.
From Takahiro Kambe.
Fixes [Bug #15178]
AC_PROG_CC_C99 calls AC_PROC_CC, which calls _AC_PROG_CC_G, which sets
default CFLAGS, whose contents are not customisable. We don't welcome
this behaviour.
The exact same thing happens for CXXFLAGS in AC_PROG_CXX.
In case of cross-compilation, GCC might provide its own linker. Its
behaviour seems slightly different from that of gcc(1). This is not a
big deal for normal situations, but the difference can cause libtool to
go mad.
We ship bundled libffi for windows users, and libffi uses libtool. If
we use cross-compiler version of gcc instead of its ld conterpart, we
fail to compile fiddle. That should not be what we want.
Historically `$program_transform_name` has been put in front of
`AC_CANONICAL_TARGET`. Previous commit changes it, which affects the
name of generated ruby binary when cross-compiling. I _guess_ the
historical behaviour is a bug (name of ruby binary shall honour --target
configure option I think), but anyways here I preserve that questionable
way.
These days as link-time optimisations spread accross compilers, they
tend to ship their own version of ld, ar, etc. Why not detect such
things if any. Users can select compilers by ./configure --with-gcc=
whatever, or select individual tool by e.g. ./configure NM=whatever.
The added AC_ARG_VAR macros enrich ./configure --help output.
The RubyVM uses C macro defines to feature detect whether
`backtrace(2)` support is available, and if so it includes C level backtraces
when the RubyVM itself crashes.
But on my machine, C level backtraces from `vm_dump.c` didn't work when
using a version of Ruby buillt on the machine, but worked fine when using a
version of Ruby built on another machine and copied to my machine.
The default autoconf test for backtraces uses a sigaltstack size that is
too small, so the SIGSEGV signal handler itself causes a SIGSEGV).
I noticed that signal.c uses a larger sigaltstack size:
https://github.com/ruby/ruby/blob/v2_6_5/signal.c#L568
The specific variables it looks at:
- `HAVE_BACKTRACE`
this is a macro defined by autoconf because there is a line in the
configure script like `AC_CHECK_FUNCS(backtrace)` (see the autoconf
docs for more).
- `BROKEN_BACKTRACE`
this comes from a custom program that Ruby's configure script runs to
attempt to figure out whether actually using backtrace(2) in a real
program works. You can see the autoconf program here.
<https://github.com/ruby/ruby/blob/v2_6_5/configure.ac#L2817-L2863>
It uses sigaltstack and SA_ONSTACK to create a seperate stack for
handling signals.
The problem was: SIGSTKSZ (which comes from a system header!) was not
suggesting a large enough stack size. When checking on an Ubuntu 16.04
box, we found that SIGSTKSZ was 8192 and MINSIGSTKSZ was 2048.
BASERUBY needs to be at least Ruby 2.2 since
46acd0075d.
I think it's better to explicitly fail early as soon as BASERUBY
is used in this case, versus trying to debug later failures.
This modifies things to check both implicitly use of ruby in the
PATH as BASERUBY, and explicitly specified older versions of ruby
when using --with-baseruby.
Fixes [Bug #16668]
Allow the 'Dir.home' method to reliably locate the user's home directory when
all three of the following are true at the same time:
1. Ruby is running on a Unix-like OS
2. The $HOME environment variable is not set
3. The process is not a descendant of login(1) (or a work-alike)
The prior behavior was that the lookup could only work for login-descended
processes.
This is accomplished by looking up the user's record in the password database
by uid (getpwuid_r(3)) as a fallback to the lookup by name (getpwname_r(3))
which is still attempted first (based on the name, if any, returned by
getlogin_r(3)).
If getlogin_r(3), getpwnam_r(3), and/or getpwuid_r(3) is not available at
compile time, will fallback on using their respective non-*_r() variants:
getlogin(3), getpwnam(3), and/or getpwuid(3).
The rationale for attempting to do the lookup by name prior to doing it by uid
is to accommodate the possibility of multiple login names (each with its own
record in the password database, so each with a potentially different home
directory) being mapped to the same uid (as is explicitly allowed for by
POSIX; see getlogin(3posix)).
Preserves the existing behavior for login-descended processes, and adds the
new capability of having Dir.home being able to find the user's home directory
for non-login-descended processes.
Fixes [Bug #16787]
Related discussion:
https://bugs.ruby-lang.org/issues/16787https://github.com/ruby/ruby/pull/3034
GCC/Clang can optimize to calculate `sin(x)` and `cos(x)` at once,
when the both are closely called on the same argument.
Similar optimization is possible for `__sinpi(x)` and `__cospi(x)`
if available, which calculate arguments in radian, i.e.
`sin(x*M_PI)` and `cos(x*M_PI)` respectively.
On some architectures (like RISC-V) sub-word atomics are only available
when linking against -latomic, but the configure script doesn't do that,
causing the atomic checks to fail and the resulting ruby binary is
non-functional. Ruby does not use sub-word atomic operations, rb_atomic_t
is defined to unsigned int, so use unsigned int when checking for atomic
operations.
To this date there is no way for Oracle developer Studio to suppress
warnings about unreachable codes (12.6 manual says it implemented
__builtin_unreachable "as a no-op to C++. It might be added to C.")
There is no way but globally kill the warning.
Every time a pointer to/from VALUE conversion happens, these two
warnings are issued:
- warning #1684: conversion from pointer to same-sized integral type (potential portability problem)
- warning #2312: pointer cast involving 64-bit pointed-to type
Thank you, but we are well aware of the "potential portability problem".
Let us ignore them all.
C++ keyword `nullptr` represents a null pointer (note also that NULL is
an integer in C++ due to its design flaw). Its type is `std::nullptr_t`,
defined in <cstddef> standard header. Why not support it when the
backend implementation can take a null pointer as an argument.
ISO/IEC 9899:1999 section 6.5.7 states that "If the value of the right
operand is negative or is greater than or equal to the width of the
promoted left operand, the behavior is undefined". So we have to take
care of such situations.
This has not been a problem because contemporary C compilers are
extraordinary smart to compile the series of shifts into a single
ROTLQ/ROTRQ machine instruction. In contrast to what C says those
instructions have fully defined behaviour for all possible inputs.
Hence it has been quite difficult to observe the undefined-ness of such
situations. But undefined is undefined. We should not rely on such
target-specific assumptions.
We are fixing the situation by carefully avoiding shifts with out-of-
range values. At least GCC since 4.6.3 and Clang since 8.0 can issue
the exact same instructions like before the changeset.
Also in case of Intel processors, there supposedly be intrinsics named
_rotr/_rotl that do exactly what we need. They, in practice, are absent
on Clang before 9.x so we cannot blindly use. But we can at least save
MSVC.
See also:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57157https://bugs.llvm.org/show_bug.cgi?id=17332
MJIT with ccache has a problem on docker environment, so
we need to use original CC (/usr/bin/gcc, for example).
Ubuntu system provides /usr/lib/ccache/gcc and so on to use
gcc with ccache. It is easy to setup ccache by adding
/usr/lib/ccache to $PATH. However we need to use /usr/bin/gcc
(and so on) for MJIT_CC. We can specify MJIT_CC option at
configure, but specifying them is troublesome.
This patch choose original $CC (/usr/bin/gcc, for example)
if $CC is /usr/lib/ccache/$CC.
The use of RUBY_WERROR_FLAG causes this test to fail on OpenBSD
platforms that use the binutils linker (ld.bfd) instead of the
llvm linker (ld.lld), due to warnings added to the binutils linker
in OpenBSD.
Removing the RUBY_WERROR_FLAG would probably also fix it, but that
would affect other platforms.
This should also be backported to Ruby 2.7.
Analysis and similar fix recommended by George Koehler.
Nobuyoshi Nakada
卜部昌平
Masaki Matsushita
Jeremy Evans
Jake Zimmerman
Watson
Alan D. Salewski
Yusuke Endoh
Andreas Schwab
Paul Jordan
Takashi Kokubun
Koichi Sasada