Currently all variables used in OpenMP regions are captured into a record and passed to outlined functions in this record. It may result in some poor performance because of too complex analysis later in optimization passes. Patch makes to emit outlined functions for parallel-based regions with a list of captured variables. It reduces code for 2*n GEPs, stores and loads at least.
Codegen for task-based regions remains unchanged because runtime requires that all captured variables are passed in captured record.
llvm-svn: 247251
This flag causes the compiler to emit bit set entries for functions as well
as runtime bitset checks at indirect call sites. Depends on the new function
bitset mechanism.
Differential Revision: http://reviews.llvm.org/D11857
llvm-svn: 247238
Summary:
Currently clang provides no general way to generate nontemporal loads/stores.
There are some architecture specific builtins for doing so (e.g. in x86), but
there is no way to generate non-temporal store on, e.g. AArch64. This patch adds
generic builtins which are expanded to a simple store with '!nontemporal'
attribute in IR.
Differential Revision: http://reviews.llvm.org/D12313
llvm-svn: 247104
Introduce an Address type to bundle a pointer value with an
alignment. Introduce APIs on CGBuilderTy to work with Address
values. Change core APIs on CGF/CGM to traffic in Address where
appropriate. Require alignments to be non-zero. Update a ton
of code to compute and propagate alignment information.
As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment
helper function to CGF and made use of it in a number of places in
the expression emitter.
The end result is that we should now be significantly more correct
when performing operations on objects that are locally known to
be under-aligned. Since alignment is not reliably tracked in the
type system, there are inherent limits to this, but at least we
are no longer confused by standard operations like derived-to-base
conversions and array-to-pointer decay. I've also fixed a large
number of bugs where we were applying the complete-object alignment
to a pointer instead of the non-virtual alignment, although most of
these were hidden by the very conservative approach we took with
member alignment.
Also, because IRGen now reliably asserts on zero alignments, we
should no longer be subject to an absurd but frustrating recurring
bug where an incomplete type would report a zero alignment and then
we'd naively do a alignmentAtOffset on it and emit code using an
alignment equal to the largest power-of-two factor of the offset.
We should also now be emitting much more aggressive alignment
attributes in the presence of over-alignment. In particular,
field access now uses alignmentAtOffset instead of min.
Several times in this patch, I had to change the existing
code-generation pattern in order to more effectively use
the Address APIs. For the most part, this seems to be a strict
improvement, like doing pointer arithmetic with GEPs instead of
ptrtoint. That said, I've tried very hard to not change semantics,
but it is likely that I've failed in a few places, for which I
apologize.
ABIArgInfo now always carries the assumed alignment of indirect and
indirect byval arguments. In order to cut down on what was already
a dauntingly large patch, I changed the code to never set align
attributes in the IR on non-byval indirect arguments. That is,
we still generate code which assumes that indirect arguments have
the given alignment, but we don't express this information to the
backend except where it's semantically required (i.e. on byvals).
This is likely a minor regression for those targets that did provide
this information, but it'll be trivial to add it back in a later
patch.
I partially punted on applying this work to CGBuiltin. Please
do not add more uses of the CreateDefaultAligned{Load,Store}
APIs; they will be going away eventually.
llvm-svn: 246985
Fix processing of shared variables with reference types in OpenMP constructs. Previously, if the variable was not marked in one of the private clauses, the reference to this variable was emitted incorrectly and caused an assertion later.
llvm-svn: 246846
Added codegen for array section in 'depend' clause of 'task' directive. It emits to pointers, one for the begin of array section and another for the end of array section. Size of the section is calculated as (end + 1 - start) * sizeof(basic_element_type).
llvm-svn: 246422
Added codegen for array section in 'depend' clause of 'task' directive. It emits to pointers, one for the begin of array section and another for the end of array section. Size of the section is calculated as (end + 1 - start) * sizeof(basic_element_type).
llvm-svn: 246278
Summary:
float_cast_overflow is the only UBSan check without a source location attached.
This patch propagates SourceLocations where necessary to get them to the
EmitCheck() call.
Reviewers: rsmith, ABataev, rjmccall
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D11757
llvm-svn: 244568
tools/clang/test/CodeGen/packed-nest-unpacked.c contains this test:
struct XBitfield {
unsigned b1 : 10;
unsigned b2 : 12;
unsigned b3 : 10;
};
struct YBitfield {
char x;
struct XBitfield y;
} __attribute((packed));
struct YBitfield gbitfield;
unsigned test7() {
// CHECK: @test7
// CHECK: load i32, i32* getelementptr inbounds (%struct.YBitfield, %struct.YBitfield* @gbitfield, i32 0, i32 1, i32 0), align 4
return gbitfield.y.b2;
}
The "align 4" is actually wrong. Accessing all of "gbitfield.y" as a single
i32 is of course possible, but that still doesn't make it 4-byte aligned as
it remains packed at offset 1 in the surrounding gbitfield object.
This alignment was changed by commit r169489, which also introduced changes
to bitfield access code in CGExpr.cpp. Code before that change used to take
into account *both* the alignment of the field to be accessed within the
current struct, *and* the alignment of that outer struct itself; this logic
was removed by the above commit.
Neglecting to consider both values can cause incorrect code to be generated
(I've seen an unaligned access crash on SystemZ due to this bug).
In order to always use the best known alignment value, this patch removes
the CGBitFieldInfo::StorageAlignment member and replaces it with a
StorageOffset member specifying the offset from the start of the surrounding
struct to the bitfield's underlying storage. This offset can then be combined
with the best-known alignment for a bitfield access lvalue to determine the
alignment to use when accessing the bitfield's storage.
Differential Revision: http://reviews.llvm.org/D11034
llvm-svn: 241916
This is needed to use clang's command line option "-ftrap-function" for LTO and
enable changing the trap function name on a per-call-site basis.
rdar://problem/21225723
Differential Revision: http://reviews.llvm.org/D10831
llvm-svn: 241306
The patch is generated using this command:
$ tools/extra/clang-tidy/tool/run-clang-tidy.py -fix \
-checks=-*,llvm-namespace-comment -header-filter='llvm/.*|clang/.*' \
work/llvm/tools/clang
To reduce churn, not touching namespaces spanning less than 10 lines.
llvm-svn: 240270
This causes programs compiled with this flag to print a diagnostic when
a control flow integrity check fails instead of aborting. Diagnostics are
printed using UBSan's runtime library.
The main motivation of this feature over -fsanitize=vptr is fidelity with
the -fsanitize=cfi implementation: the diagnostics are printed under exactly
the same conditions as those which would cause -fsanitize=cfi to abort the
program. This means that the same restrictions apply regarding compiling
all translation units with -fsanitize=cfi, cross-DSO virtual calls are
forbidden, etc.
Differential Revision: http://reviews.llvm.org/D10268
llvm-svn: 240109
This flag controls whether a given sanitizer traps upon detecting
an error. It currently only supports UBSan. The existing flag
-fsanitize-undefined-trap-on-error has been made an alias of
-fsanitize-trap=undefined.
This change also cleans up some awkward behavior around the combination
of -fsanitize-trap=undefined and -fsanitize=undefined. Previously we
would reject command lines containing the combination of these two flags,
as -fsanitize=vptr is not compatible with trapping. This required the
creation of -fsanitize=undefined-trap, which excluded -fsanitize=vptr
(and -fsanitize=function, but this seems like an oversight).
Now, -fsanitize=undefined is an alias for -fsanitize=undefined-trap,
and if -fsanitize-trap=undefined is specified, we treat -fsanitize=vptr
as an "unsupported" flag, which means that we error out if the flag is
specified explicitly, but implicitly disable it if the flag was implied
by -fsanitize=undefined.
Differential Revision: http://reviews.llvm.org/D10464
llvm-svn: 240105
The RegionCounter type does a lot of legwork, but most of it is only
meaningful within the implementation of CodeGenPGO. The uses elsewhere
in CodeGen generally just want to increment or read counters, so do
that directly.
llvm-svn: 235664
This scheme checks that pointer and lvalue casts are made to an object of
the correct dynamic type; that is, the dynamic type of the object must be
a derived class of the pointee type of the cast. The checks are currently
only introduced where the class being casted to is a polymorphic class.
Differential Revision: http://reviews.llvm.org/D8312
llvm-svn: 232241
I disabled putting the new global into the same COMDAT as the function for now.
There's a fundamental problem when we inline references to the global but still
have the global in a COMDAT linked to the inlined function. Since this is only
an optimization there may be other versions of the COMDAT around that are
missing the new global and hell breaks loose at link time.
I hope the chromium build doesn't break this time :)
llvm-svn: 231564
This broke the Chromium build. Links were failing with messages like:
obj/dbus/libdbus_test_support.a(obj/dbus/dbus_test_support.mock_object_proxy.o):../../dbus/mock_object_proxy.cc:function dbus::MockObjectProxy::Detach(): warning: relocation refers to discarded section
/usr/local/google/work/chromium/src/third_party/binutils/Linux_x64/Release/bin/ld.gold: error: treating warnings as errors
llvm-svn: 231541
Instead of creating a copy on the stack just stash them in a private
constant global. This saves both the copying overhead and the stack
space, and gives the optimizer more room to constant fold.
This tries to make array temporaries more similar to regular arrays,
they can't use the same logic because a temporary has no VarDecl to be
bound to so we roll our own version here.
The original use case for this optimization was code like
for (int i : {1, 2, 3, 4, 5, 6, 7, 8, 10})
foo(i);
where without this patch (assuming that the loop is not unrolled) we
would alloca an array on the stack, copy the 10 values over and
iterate on that. With this patch we put the array in .text use it
directly. Apart from that case this helps on virtually any passing of
a constant std::initializer_list as a function argument.
Differential Revision: http://reviews.llvm.org/D8034
llvm-svn: 231508
Bools are a little tricky, they are i8 in memory and must be coerced
back to i1 before further operations can be performed on them.
This fixes PR22577.
llvm-svn: 229204
The /volatile:ms semantics turn volatile loads and stores into atomic
acquire and release operations. This distinction is important because
volatile memory operations do not form a happens-before relationship
with non-atomic memory. This means that a volatile store is not
sufficient for implementing a mutex unlock routine.
Differential Revision: http://reviews.llvm.org/D7580
llvm-svn: 229082
This causes things like assignment to refer to the '=' rather than the
LHS when attributing the store instruction, for example.
There were essentially 3 options for this:
* The beginning of an expression (this was the behavior prior to this
commit). This meant that stepping through subexpressions would bounce
around from subexpressions back to the start of the outer expression,
etc. (eg: x + y + z would go x, y, x, z, x (the repeated 'x's would be
where the actual addition occurred)).
* The end of an expression. This seems to be what GCC does /mostly/, and
certainly this for function calls. This has the advantage that
progress is always 'forwards' (never jumping backwards - except for
independent subexpressions if they're evaluated in interesting orders,
etc). "x + y + z" would go "x y z" with the additions occurring at y
and z after the respective loads.
The problem with this is that the user would still have to think
fairly hard about precedence to realize which subexpression is being
evaluated or which operator overload is being called in, say, an asan
backtrace.
* The preferred location or 'exprloc'. In this case you get sort of what
you'd expect, though it's a bit confusing in its own way due to going
'backwards'. In this case the locations would be: "x y + z +" in
lovely postfix arithmetic order. But this does mean that if the op+
were an operator overload, say, and in a backtrace, the backtrace will
point to the exact '+' that's being called, not to the end of one of
its operands.
(actually the operator overload case doesn't work yet for other reasons,
but that's being fixed - but this at least gets scalar/complex
assignments and other plain operators right)
llvm-svn: 227027
This workaround was to provide unique call sites to ensure LLVM's inline
debug info handling would properly unique two calls to the same function
on the same line. Instead, this has now been fixed in LLVM (r226736) and
the workaround here can be removed.
Originally committed in r176895, but this isn't a straight revert due to
all the changes since then. I just searched for anything ForcedColumn*
related and removed them.
We could test this - but it didn't strike me as terribly valuable once
we're no longer adding this workaround everything just works as expected
& it's no longer a special case to test for.
llvm-svn: 226738
Sorry for the noise, I managed to miss a bunch of recent regressions of
include orderings here. This should actually sort all the includes for
Clang. Again, no functionality changed, this is just a mechanical
cleanup that I try to run periodically to keep the #include lines as
regular as possible across the project.
llvm-svn: 225979
Several pieces of code were relying on implicit debug location setting
which usually lead to incorrect line information anyway. So I've fixed
those (in r225955 and r225845) separately which should pave the way for
this commit to be cleanly reapplied.
The reason these implicit dependencies resulted in crashes with this
patch is that the debug location would no longer implicitly leak from
one place to another, but be set back to invalid. Once a call with
no/invalid location was emitted, if that call was ever inlined it could
produce invalid debugloc chains and assert during LLVM's codegen.
There may be further cases of such bugs in this patch - they're hard to
flush out with regression testing, so I'll keep an eye out for reports
and investigate/fix them ASAP if they come up.
Original commit message:
Reapply "DebugInfo: Generalize debug info location handling"
Originally committed in r224385 and reverted in r224441 due to concerns
this change might've introduced a crash. Turns out this change fixes the
crash introduced by one of my earlier more specific location handling
changes (those specific fixes are reverted by this patch, in favor of
the more general solution).
Recommitted in r224941 and reverted in r224970 after it caused a crash
when building compiler-rt. Looks to be due to this change zeroing out
the debug location when emitting default arguments (which were meant to
inherit their outer expression's location) thus creating call
instructions without locations - these create problems for inlining and
must not be created. That is fixed and tested in this version of the
change.
Original commit message:
This is a more scalable (fixed in mostly one place, rather than many
places that will need constant improvement/maintenance) solution to
several commits I've made recently to increase source fidelity for
subexpressions.
This resetting had to be done at the DebugLoc level (not the
SourceLocation level) to preserve scoping information (if the resetting
was done with CGDebugInfo::EmitLocation, it would've caused the tail end
of an expression's codegen to end up in a potentially different scope
than the start, even though it was at the same source location). The
drawback to this is that it might leave CGDebugInfo out of sync. Ideally
CGDebugInfo shouldn't have a duplicate sense of the current
SourceLocation, but for now it seems it does... - I don't think I'm
going to tackle removing that just now.
I expect this'll probably cause some more buildbot fallout & I'll
investigate that as it comes up.
Also these sort of improvements might be starting to show a weakness/bug
in LLVM's line table handling: we don't correctly emit is_stmt for
statements, we just put it on every line table entry. This means one
statement split over multiple lines appears as multiple 'statements' and
two statements on one line (without column info) are treated as one
statement.
I don't think we have any IR representation of statements that would
help us distinguish these cases and identify the beginning of each
statement - so that might be something we need to add (possibly to the
lexical scope chain - a scope for each statement). This does cause some
problems for GDB and possibly other DWARF consumers.
llvm-svn: 225956
Alexey Bataev