In the runtime's implementation of floating-point SUM, the
implementation of Kahan's algorithm for increased precision is
incorrect. The running correction factor should be subtracted from each
new data item, not added to it. This fix ensures that the sum of 100M
random default real values between 0. and 1. is close to 5.E7.
See https://en.wikipedia.org/wiki/Kahan_summation_algorithm.
(This is a big patch, but it's nearly an NFC. No test results have
changed and all Fortran tests in the LLVM test suites work as expected.)
Allow a parser::Message for a warning to be marked with the
common::LanguageFeature or common::UsageWarning that controls it. This
will allow a later patch to add hooks whereby a driver will be able to
decorate warning messages with the names of its options that enable each
particular warning, and to add hooks whereby a driver can map those
enumerators by name to command-line options that enable/disable the
language feature and enable/disable the messages.
The default settings in the constructor for LanguageFeatureControl were
moved from its header file into its C++ source file.
Hooks for a driver to use to map the name of a feature or warning to its
enumerator were also added.
To simplify the tagging of warnings with their corresponding language
feature or usage warning, to ensure that they are properly controlled by
ShouldWarn(), and to ensure that warnings never issue at code sites in
module files, two new Warn() member function templates were added to
SemanticsContext and other contextual frameworks. Warn() can't be used
before source locations can be mapped to scopes, but the bulk of
existing code blocks testing ShouldWarn() and FindModuleFile() before
calling Say() were convertible into calls to Warn(). The ones that were
not convertible were extended with explicit calls to
Message::set_languageFeature() and set_usageWarning().
Conversions of infinities from other kinds to real(10) were incorrect,
and comparisons of real(2) vs real(3) are dicey as conversions in one
direction can overflow and conversions in the other can lose precision.
Use real(16) as the common type for comparisons in IEEE_NEAREST_AFTER.
The generation of 80-bit x87 floating-point infinities was incorrect in
Normalize(), the comparison for IEEE_NEXT_AFTER needs to use the most
precise type of its arguments, and we don't need to warn about overflows
from +/-HUGE() to infinity. Warnings about NaN arguments remain in
place, and enabled by default, as their usage may or may not be
portable, and their appearance in a real code seems most likely to
signify an earlier error.
A couple of intrinsic functions have optional arguments. Don't insert
type conversions on those arguments when the actual arguments may not be
present at execution time, due to being OPTIONAL, allocatables, or
pointers.
When the second argument to these intrinsic functions is a scalar
constant zero, emit a warning (if enabled) even if the first argument is
not a constant.
…Warn()
Many warning messages were being emitted unconditionally. Ensure that
all warnings are conditional on a true result from a call to
common::LanguageFeatureControl::ShouldWarn() so that it is easy for a
driver to disable them all, or, in the future, to provide per-warning
control over them.
The transformational intrinsic functions MATMUL, DOT_PRODUCT, and NORM2
all involve summing up intermediate products into accumulators. In the
constant folding library, this is done with extended precision Kahan
summation for REAL and COMPLEX arguments, but in the runtime
implementations it is not, and this leads to discrepancies between
folded results and dynamic results.
Disable the use of Kahan summation in folding to resolve these
discrepancies, but don't discard the code, in case we want to add Kahan
summation in the runtime for some or all of these intrinsic functions.
Detect NaN elements in data and handle them like gfortran does (at
runtime); namely, NaN can be returned if all the data are NaNs, but any
non-NaN value is preferable. Ensure that folding returns the same
results as runtime computation.
Fixes llvm-test-suite/Fortran/gfortran/regression/maxloc_2.f90 (and
probably others).
The code that folds the relatively new NORM2 intrinsic function can
produce overflow in cases where it's not warranted. Rearrange to NORM2 =
M * SQRT((A(:)/M)**2) where M is MAXVAL(ABS(A)).
Fold references to the (relatively new) intrinsic function NORM2 at
compilation time when the argument(s) are all constants. (Getting this
done right involved some changes to the API of the accumulator function
objects used by the DoReduction<> template, which rippled through some
other reduction function folding code.)
Generalize FoldMerge() to accommodate derived type arguments and results,
rename it into Folder<T>::MERGE(), and remove it from the various
FoldIntrinsicFunction() routines for intrinsic types.
Fixes llvm-test-suite/Fortran/gfortran/regression/merge_init_expr_2.f90.
Differential Revision: https://reviews.llvm.org/D157345
DPROD(x,y) is defined as DBLE(x)*DBLE(y) and that's exactly how
the implementation of its rewriting and possible folding should
be implemented, instead of the current code that only works when
both arguments are scalar and crashes otherwise.
Fixes https://github.com/llvm/llvm-project/issues/63991.
Differential Revision: https://reviews.llvm.org/D156754
The code that folds the intrinsic function HYPOT was neglecting to
warn the programmer about overflow when it occurs.
Differential Revision: https://reviews.llvm.org/D154371
The intrinsic function DIM can overflow when its second argument
is negative. Detect this case for real and integer arguments and
emit a warning when necessary.
Differential Revision: https://reviews.llvm.org/D143798
Emit a warning when the result of folding a call to
ABS() with a complex argument results in an overflow.
Differential Revision: https://reviews.llvm.org/D136904
The implementation of the folding code for SET_EXPONENT() was written
in such a fashion as to convert the I= actual argument value to a 32-bit
integer. Which is usually not a problem, but it's not always correct
and a test case ran into trouble with it. Fix to allow any kind of
INTEGER without conversion.
Differential Revision: https://reviews.llvm.org/D135203
Complex component references (z%RE, z%IM) of complex named constants
should be evaluated at compilation time.
Differential Revision: https://reviews.llvm.org/D125341
Evaluate real-valued references to the intrinsic functions MODULO
and MOD at compilation time without recourse to an external math
library.
Differential Revision: https://reviews.llvm.org/D125151
Fold references to the intrinsic function DIM with constant real
arguments. And clean up folding of comparisons with NaNs to address
a problem noticed in testing -- NaNs should successfully compare
unequal to all values, including themselves, instead of failing all
comparisons.
Differential Revision: https://reviews.llvm.org/D125146
The related real number system inquiry functions SPACING()
and RRSPACING() can be folded for constant arguments.
See 16.9.164 & 16.9.180 in Fortran 2018.
Differential Revision: https://reviews.llvm.org/D125100
Transformational bessel intrinsic functions require the same math runtime
as elemental bessel intrinsics.
Currently elemental bessels could be folded if f18 was linked with pgmath
(cmake -DLIBPGMATH_DIR option). `j0`, `y0`, ... C libm functions were not
used because they are not standard C functions: they are Posix
extensions.
This patch enable:
- Using the Posix bessel host runtime functions when available.
- folding the transformational bessel using the elemental version.
Differential Revision: https://reviews.llvm.org/D124167
Adds flang/include/flang/Common/log2-visit.h, which defines
a Fortran::common::visit() template function that is a drop-in
replacement for std::visit(). Modifies most use sites in
the front-end and runtime to use common::visit().
The C++ standard mandates that std::visit() have O(1) execution
time, which forces implementations to build dispatch tables.
This new common::visit() is O(log2 N) in the number of alternatives
in a variant<>, but that N tends to be small and so this change
produces a fairly significant improvement in compiler build
memory requirements, a 5-10% improvement in compiler build time,
and a small improvement in compiler execution time.
Building with -DFLANG_USE_STD_VISIT causes common::visit()
to be an alias for std::visit().
Calls to common::visit() with multiple variant arguments
are referred to std::visit(), pending further work.
This change is enabled only for GCC builds with GCC >= 9;
an earlier attempt (D122441) ran into bugs in some versions of
clang and was reverted rather than simply disabled; and it is
not well tested with MSVC. In non-GCC and older GCC builds,
common::visit() is simply an alias for std::visit().
Implement constant folding for the intrinsic function NEAREST()
and the related functions IEEE_NEXT_AFTER(), IEEE_NEXT_UP(), and
IEEE_NEXT_DOWN().
Differential Revision: https://reviews.llvm.org/D122510
Adds flang/include/flang/Common/visit.h, which defines
a Fortran::common::visit() template function that is a drop-in
replacement for std::visit(). Modifies most use sites in
the front-end and runtime to use common::visit().
The C++ standard mandates that std::visit() have O(1) execution
time, which forces implementations to build dispatch tables.
This new common::visit() is O(log2 N) in the number of alternatives
in a variant<>, but that N tends to be small and so this change
produces a fairly significant improvement in compiler build
memory requirements, a 5-10% improvement in compiler build time,
and a small improvement in compiler execution time.
Building with -DFLANG_USE_STD_VISIT causes common::visit()
to be an alias for std::visit().
Calls to common::visit() with multiple variant arguments
are referred to std::visit(), pending further work.
Differential Revision: https://reviews.llvm.org/D122441
Using recently established message severity codes, upgrade
non-fatal messages to usage and portability warnings as
appropriate.
Differential Revision: https://reviews.llvm.org/D121246
Fold references to the intrinsic function SCALE().
(Also work around some MSVC headaches somehow exposed by
this patch: disable a bogus MSVC warning that began to appear
in unrelated source files, and avoid the otherwise-necessary
use of the "template" keyword in a call to a template member
function of a class template.)
Differential Revision: https://reviews.llvm.org/D117150
These functions were missing from the standard intrinsic module
IEEE_ARITHMETIC. IEEE_SCALB is an alias for the standard intrinsic
function SCALE(), and the others are defined as new builtin intrinsic
functions.
Differential Revision: https://reviews.llvm.org/D111253
Implement IEEE Real::SQRT() operation, then use it to
also implement Real::HYPOT(), which can then be used directly
to implement Complex::ABS().
Differential Revision: https://reviews.llvm.org/D109250
Refactor the recently-implemented MAXVAL/MINVAL folding so
that the parts that can be used to implement other reduction
transformational intrinsic function folding are exposed.
Use them to implement folding of IALL, IANY, IPARITY,
SUM. and PRODUCT. Replace the folding of ALL & ANY to
use the new infrastructure and become able to handle DIM=
arguments.
Differential Revision: https://reviews.llvm.org/D104562
Implement constant folding for the reduction transformational
intrinsic functions MAXVAL and MINVAL.
In anticipation of more folding work to follow, with (I hope)
some common infrastructure, these two have been implemented in a
new header file.
Differential Revision: https://reviews.llvm.org/D104337
- Rework the host runtime table so that it is constexpr to avoid
having to construct it and to store/propagate it.
- Make the interface simpler (remove many templates and a file)
- Enable 16bits float folding using 32bits float host runtime
- Move StaticMultimapView into its own header to use it for host
folding
Reviewed By: klausler, PeteSteinfeld
Differential Revision: https://reviews.llvm.org/D88981
This change prepares usage of lipgmath description in lowering.
- Removes the static variable templates that were used to abstract
libpgmath description
- Move the description to pgmath.h.inc header and rework the macros
so that they can both be used to declare pgmath functions and use
them.
The way they are to be used is left to pgmath.h.inc user that
must define PGMATH_USE_XX macros that will be called for all pgmath
functions in pgmath.h.inc.
- In intrinsic-library.cpp define PGMATH_USE_XX macro callbacks in
order to capture function pointers to pgmath functions as well as
a description of their type. This will be used for constant folding
using pgmath.
- Change atan/atan2 handling to use atan2 instead of atan when there are two
arguments because it is easier to handle in the runtime description.
Also fixes lipgmath linking regression after D78215 cmake changes.
This change is motivated by the need to use a similar pgmath
description in lowering. The difference is that no function pointers will
be taken there, and instead only the function name and type are needed.
Reviewed By: schweitz, sscalpone
Differential Revision: https://reviews.llvm.org/D83051
Summary:
This patch changes speficic extremum functions rewrite to generic MIN/MAX.
It applies to AMAX0, AMIN0, AMAX1, AMIN1, MAX0, MIN0, MAX1, MIN1, DMAX1,
and DMIN1.
- Do not re-write specific extremums to MAX/MIN in intrinsic Probe and let
folding rewrite it and introduc the conversion on the MIN/MAX result.
- Also make operand promotion explicit in MIN/MAX folding.
For instance, after this patch:
AMAX0(int8, int4) is rewritten to REAL(MAX(int8, INT(int4, 8)))
All this care is to avoid rewritting it to MAX(REAL(int8), REAL(int4))
that may not always be numerically equivalent to the first rewrite.
Reviewers: klausler, schweitz, sscalpone, jdoerfert, DavidTruby
Reviewed By: klausler, schweitz
Subscribers: llvm-commits, flang-commits
Tags: #flang, #llvm
Differential Revision: https://reviews.llvm.org/D81940
