a NaN-test prior to the call to the library function.
This should automatically make fastmath (including just non-NaNs) able to avoid
the expensive libcalls and also open the door to more advanced folding in LLVM
based on the rules for complex math.
Two important notes to remember: first is that this isn't yet a proper
limited range mode, it's still just improving the unlimited range mode.
Also, it isn't really perfecet w.r.t. what an unlimited range mode
should be doing because it isn't quite handling the flags produced by
all the operations in the way desirable for that mode, but then neither
is compiler-rt's libcall. When the compiler-rt libcall is improved to
carefully manage flags, the code emitted here should be improved
correspondingly. And it is still a long-term desirable thing to add
a limited range mode to Clang that would be able to use direct math
without library calls here.
Special thanks to Steve Canon for the careful review on this patch and
teaching me about these issues. =D
Differential Revision: http://reviews.llvm.org/D5756
llvm-svn: 220167
for complex math.
This should fix the windows build bots that started having trouble here
and generally fix complex libcall emission on targets which use sret for
complex data types. It also makes the code a bit simpler (despite
calling into a much more complex bucket of code).
llvm-svn: 219565
operators where one type is a C complex type, and to emit both the
efficient and correct implementation for complex arithmetic according to
C11 Annex G using this extra information.
For both multiply and divide the old code was writing a long-hand
reduced version of the math without any of the special handling of inf
and NaN recommended by the standard here. Instead of putting more
complexity here, this change does what GCC does which is to emit
a libcall for the fully general case.
However, the old code also failed to do the proper minimization of the
set of operations when there was a mixed complex and real operation. In
those cases, C provides a spec for much more minimal operations that are
valid. Clang now emits the exact suggested operations. This change isn't
*just* about performance though, without minimizing these operations, we
again lose the correct handling of infinities and NaNs. It is critical
that this happen in the frontend based on assymetric type operands to
complex math operations.
The performance implications of this change aren't trivial either. I've
run a set of benchmarks in Eigen, an open source mathematics library
that makes heavy use of complex. While a few have slowed down due to the
libcall being introduce, most sped up and some by a huge amount: up to
100% and 140%.
In order to make all of this work, also match the algorithm in the
constant evaluator to the one in the runtime library. Currently it is
a broken port of the simplifications from C's Annex G to the long-hand
formulation of the algorithm.
Splitting this patch up is very hard because none of this works without
the AST change to preserve non-complex operands. Sorry for the enormous
change.
Follow-up changes will include support for sinking the libcalls onto
cold paths in common cases and fastmath improvements to allow more
aggressive backend folding.
Differential Revision: http://reviews.llvm.org/D5698
llvm-svn: 219557
Previously, we made one traversal of the AST prior to codegen to assign
counters to the ASTs and then propagated the count values during codegen. This
patch now adds a separate AST traversal prior to codegen for the
-fprofile-instr-use option to propagate the count values. The counts are then
saved in a map from which they can be retrieved during codegen.
This new approach has several advantages:
1. It gets rid of a lot of extra PGO-related code that had previously been
added to codegen.
2. It fixes a serious bug. My original implementation (which was mailed to the
list but never committed) used 3 counters for every loop. Justin improved it to
move 2 of those counters into the less-frequently executed breaks and continues,
but that turned out to produce wrong count values in some cases. The solution
requires visiting a loop body before the condition so that the count for the
condition properly includes the break and continue counts. Changing codegen to
visit a loop body first would be a fairly invasive change, but with a separate
AST traversal, it is easy to control the order of traversal. I've added a
testcase (provided by Justin) to make sure this works correctly.
3. It improves the instrumentation overhead, reducing the number of counters for
a loop from 3 to 1. We no longer need dedicated counters for breaks and
continues, since we can just use the propagated count values when visiting
breaks and continues.
To make this work, I needed to make a change to the way we count case
statements, going back to my original approach of not including the fall-through
in the counter values. This was necessary because there isn't always an AST node
that can be used to record the fall-through count. Now case statements are
handled the same as default statements, with the fall-through paths branching
over the counter increments. While I was at it, I also went back to using this
approach for do-loops -- omitting the fall-through count into the loop body
simplifies some of the calculations and make them behave the same as other
loops. Whenever we start using this instrumentation for coverage, we'll need
to add the fall-through counts into the counter values.
llvm-svn: 201528
adjustFallThroughCount isn't a good name, and the documentation was
even worse. This commit attempts to clarify what it's for and when to
use it.
llvm-svn: 199139
With the introduction of explicit address space casts into LLVM, there's
a need to provide a new cast kind the front-end can create for C/OpenCL/CUDA
and code to produce address space casts from those kinds when appropriate.
Patch by Michele Scandale!
llvm-svn: 197036
This is the same way GenericSelectionExpr works, and it's generally a
more consistent approach.
A large part of this patch is devoted to caching the value of the condition
of a ChooseExpr; it's needed to avoid threading an ASTContext into
IgnoreParens().
Fixes <rdar://problem/14438917>.
llvm-svn: 186738
_Complex load/store didn't have their alignment set properly, which was visible when GCC's torture tests use volatile _Complex.
Update some existing tests to check for alignment, and add a new test which also has over-aligned volatile _Complex (since the imaginary part shouldn't be overaligned, only the real part).
llvm-svn: 186490
Add a CXXDefaultInitExpr, analogous to CXXDefaultArgExpr, and use it both in
CXXCtorInitializers and in InitListExprs to represent a default initializer.
There's an additional complication here: because the default initializer can
refer to the initialized object via its 'this' pointer, we need to make sure
that 'this' points to the right thing within the evaluation.
llvm-svn: 179958
aggregate types in a profoundly wrong way that has to be
worked around in every call site, to getEvaluationKind,
which classifies and distinguishes between all of these
cases.
Also, normalize the API for loading and storing complexes.
I'm working on a larger patch and wanted to pull these
changes out, but it would have be annoying to detangle
them from each other.
llvm-svn: 176656
uncovered.
This required manually correcting all of the incorrect main-module
headers I could find, and running the new llvm/utils/sort_includes.py
script over the files.
I also manually added quite a few missing headers that were uncovered by
shuffling the order or moving headers up to be main-module-headers.
llvm-svn: 169237
(__builtin_* etc.) so that it isn't possible to take their address.
Specifically, introduce a new type to represent a reference to a builtin
function, and a new cast kind to convert it to a function pointer in the
operand of a call. Fixes PR13195.
llvm-svn: 162962
Chandler Carruth
Joerg Sonnenberger
Craig Topper
Bob Wilson
Justin Bogner
David Tweed
Nick Lewycky
Eli Friedman
JF Bastien
Richard Smith
John McCall
Guy Benyei
David Blaikie