Though it is incorrect from point of view of OpenMP standard to have
dependent iteration space in OpenMP loops, compiler should not crash.
Patch fixes this problem.
llvm-svn: 319700
There are 20 LLVM math intrinsics that correspond to mathlib calls according to the LangRef:
http://llvm.org/docs/LangRef.html#standard-c-library-intrinsics
We were only converting 3 mathlib calls (sqrt, fma, pow) and 12 builtin calls (ceil, copysign,
fabs, floor, fma, fmax, fmin, nearbyint, pow, rint, round, trunc) to their intrinsic-equivalents.
This patch pulls the transforms together and handles all 20 cases. The switch is guarded by a
check for const-ness to make sure we're not doing the transform if errno could possibly be set by
the libcall or builtin.
Differential Revision: https://reviews.llvm.org/D40044
llvm-svn: 319593
Previously we emitted `__tgt_target_teams` only for standalone teams
directives. This patch allows emit this function for all teams-based
directives.
llvm-svn: 319585
These command line options are not intended for public use, and often
don't even make sense in the context of a particular tool anyway. About
90% of them are already hidden, but when people add new options they
forget to hide them, so if you were to make a brand new tool today, link
against one of LLVM's libraries, and run tool -help you would get a
bunch of junk that doesn't make sense for the tool you're writing.
This patch hides these options. The real solution is to not have
libraries defining command line options, but that's a much larger effort
and not something I'm prepared to take on.
Differential Revision: https://reviews.llvm.org/D40674
llvm-svn: 319505
The basic idea behind this patch is that since in strict aliasing
mode all accesses to union members require their outermost
enclosing union objects to be specified explicitly, then for a
couple given accesses to union members of the form
p->a.b.c...
q->x.y.z...
it is known they can only alias if both p and q point to the same
union type and offset ranges of members a.b.c... and x.y.z...
overlap. Note that the actual types of the members do not matter.
Specifically, in this patch we do the following:
* Make unions to be valid TBAA base access types. This enables
generation of TBAA type descriptors for unions.
* Encode union types as structures with a single member of a
special "union member" type. Currently we do not encode
information about sizes of types, but conceptually such union
members are considered to be of the size of the whole union.
* Encode accesses to direct and indirect union members, including
member arrays, as accesses to these special members. All
accesses to members of a union thus get the same offset, which
is the offset of the union they are part of. This means the
existing LLVM TBAA machinery is able to handle such accesses
with no changes.
While this is already an improvement comparing to the current
situation, that is, representing all union accesses as may-alias
ones, there are further changes planned to complete the support
for unions. One of them is storing information about access sizes
so we can distinct accesses to non-overlapping union members,
including accesses to different elements of member arrays.
Another change is encoding type sizes in order to make it
possible to compute offsets within constant-indexed array
elements. These enhancements will be addressed with separate
patches.
Differential Revision: https://reviews.llvm.org/D39455
llvm-svn: 319413
Summary:
The -fxray-always-emit-customevents flag instructs clang to always emit
the LLVM IR for calls to the `__xray_customevent(...)` built-in
function. The default behaviour currently respects whether the function
has an `[[clang::xray_never_instrument]]` attribute, and thus not lower
the appropriate IR code for the custom event built-in.
This change allows users calling through to the
`__xray_customevent(...)` built-in to always see those calls lowered to
the corresponding LLVM IR to lay down instrumentation points for these
custom event calls.
Using this flag enables us to emit even just the user-provided custom
events even while never instrumenting the start/end of the function
where they appear. This is useful in cases where "phase markers" using
__xray_customevent(...) can have very few instructions, must never be
instrumented when entered/exited.
Reviewers: rnk, dblaikie, kpw
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D40601
llvm-svn: 319388
Emit a gap area starting after the r-paren location and ending at the
start of the body for the braces-optional statements (for, for-each,
while, etc). The count for the gap area equal to the body's count. This
extends the fix in r317758.
Fixes PR35387, rdar://35570345
Testing: stage2 coverage-enabled build of clang, check-clang
llvm-svn: 319373
Fixes regression introduced by r319297. MSVC environments still use SEH
unwind opcodes but they should use the Microsoft C++ EH personality, not
the mingw one.
llvm-svn: 319363
This is a re-apply of r319294.
adds -fseh-exceptions and -fdwarf-exceptions flags
clang will check if the user has specified an exception model flag,
in the absense of specifying the exception model clang will then check
the driver default and append the model flag for that target to cc1
-fno-exceptions has a higher priority then specifying the model
move __SEH__ macro definitions out of Targets into InitPreprocessor
behind the -fseh-exceptions flag
move __ARM_DWARF_EH__ macrodefinitions out of verious targets and into
InitPreprocessor behind the -fdwarf-exceptions flag and arm|thumb check
remove unused USESEHExceptions from the MinGW Driver
fold USESjLjExceptions into a new GetExceptionModel function that
gives the toolchain classes more flexibility with eh models
Reviewers: rnk, mstorsjo
Differential Revision: https://reviews.llvm.org/D39673
llvm-svn: 319297
adds -fseh-exceptions and -fdwarf-exceptions flags
clang will check if the user has specified an exception model flag,
in the absense of specifying the exception model clang will then check
the driver default and append the model flag for that target to cc1
clang cc1 assumes dwarf is the default if none is passed
and -fno-exceptions has a higher priority then specifying the model
move __SEH__ macro definitions out of Targets into InitPreprocessor
behind the -fseh-exceptions flag
move __ARM_DWARF_EH__ macrodefinitions out of verious targets and into
InitPreprocessor behind the -fdwarf-exceptions flag and arm|thumb check
remove unused USESEHExceptions from the MinGW Driver
fold USESjLjExceptions into a new GetExceptionModel function that
gives the toolchain classes more flexibility with eh models
Reviewers: rnk, mstorsjo
Differential Revision: https://reviews.llvm.org/D39673
llvm-svn: 319294
Currently CodeGen is calling std::sort on the features vector in TargetOptions for every function, but I don't think CodeGen should be modifying TargetOptions.
Differential Revision: https://reviews.llvm.org/D40228
llvm-svn: 319195
These functions were defined as static members of TemplateSpecializationType.
Now they are moved to namespace level. Previously there were different
implementations for lists containing TemplateArgument and TemplateArgumentLoc,
now these implementations share the same code.
This change is a result of refactoring patch D40508. NFC.
llvm-svn: 319178
The information about access and type sizes is necessary for
producing TBAA metadata in the new size-aware format. With this
patch, D39955 and D39956 in place we should be able to change
CodeGenTBAA::createScalarTypeNode() and
CodeGenTBAA::getBaseTypeInfo() to generate metadata in the new
format under the -new-struct-path-tbaa command-line option. For
now, this new information remains unused.
Differential Revision: https://reviews.llvm.org/D40176
llvm-svn: 319012
In the future the compiler will analyze whether the OpenMP
runtime needs to be (fully) initialized and avoid that overhead
if possible. The functions already take an argument to transfer
that information to the runtime, so pass in the default value 1.
(This is needed for binary compatibility with libomptarget-nvptx
currently being upstreamed.)
Differential Revision: https://reviews.llvm.org/D40354
llvm-svn: 318836
This clang patch changes the __tgt_* API function signatures in preparation for the new map interface.
Changes are: Device IDs 32bits --> 64bits, Flags 32bits --> 64bits
Differential revision: https://reviews.llvm.org/D40281
llvm-svn: 318789
This is an instrumentation flag that's similar to
-finstrument-functions, but it only inserts calls on function entry, the
calls are inserted post-inlining, and they don't take any arugments.
This is intended for users who want to instrument function entry with
minimal overhead.
(-pg would be another alternative, but forces frame pointer emission and
affects link flags, so is probably best left alone to be used for
generating gcov data.)
Differential revision: https://reviews.llvm.org/D40276
llvm-svn: 318785
OpenMP 5.0 introduces asynchronous data update/dependecies clauses on
target data directives. Patch adds initial support for outer task
regions to use task-based codegen for future async target data
directives.
llvm-svn: 318781
Summary:
This patch is part of the development effort to add support in the current OpenMP GPU offloading implementation for implicitly sharing variables between a target region executed by the team master thread and the worker threads within that team.
This patch is the first of three required for successfully performing the implicit sharing of master thread variables with the worker threads within a team. The remaining two patches are:
- Patch D38978 to the LLVM NVPTX backend which ensures the lowering of shared variables to an device memory which allows the sharing of references;
- Patch (coming soon) is a patch to libomptarget runtime library which ensures that a list of references to shared variables is properly maintained.
A simple code snippet which illustrates an implicit data sharing situation is as follows:
```
#pragma omp target
{
// master thread only
int v;
#pragma omp parallel
{
// worker threads
// use v
}
}
```
Variable v is implicitly shared from the team master thread which executes the code in between the target and parallel directives. The worker threads must operate on the latest version of v, including any updates performed by the master.
The code generated in this patch relies on the LLVM NVPTX patch (mentioned above) which prevents v from being lowered in the thread local memory of the master thread thus making the reference to this variable un-shareable with the workers. This ensures that the code generated by this patch is correct.
Since the parallel region is outlined the passing of arguments to the outlined regions must preserve the original order of arguments. The runtime therefore maintains a list of references to shared variables thus ensuring their passing in the correct order. The passing of arguments to the outlined parallel function is performed in a separate function which the data sharing infrastructure constructs in this patch. The function is inlined when optimizations are enabled.
Reviewers: hfinkel, carlo.bertolli, arpith-jacob, Hahnfeld, ABataev, caomhin
Reviewed By: ABataev
Subscribers: cfe-commits, jholewinski
Differential Revision: https://reviews.llvm.org/D38976
llvm-svn: 318773
This patch introduces a couple of helper functions that make it
possible to handle the caching logic in a single place.
Differential Revision: https://reviews.llvm.org/D39953
llvm-svn: 318752
https://reviews.llvm.org/D40187
This patch implements code gen for 'teams distribute parallel for' on the host, including all its clauses and related regression tests.
llvm-svn: 318692
The object is provided by the objc runtime and is never visible in the
module itself, but even so, the address point we compute points into it,
and "+16" is guaranteed not to overflow.
This matches the c++ vtable IRGen.
Note that I'm not entirely convinced the 'i8*' type is correct here: at
the IR level, we're accessing memory that's outside the global object.
But we don't control the allocation, so it's not obviously wrong either.
But either way, this is only in a global initializer, so I don't think
it's going to be mucked with. Filed PR35352 to discuss that.
llvm-svn: 318545
Summary:
The MS ABI convention is that the 'this' pointer on entry is the address
of the vfptr that was used to make the virtual method call. In other
words, the pointer on entry always points to the base subobject that
introduced the virtual method. Consider this hierarchy:
struct A { virtual void f() = 0; };
struct B { virtual void g() = 0; };
struct C : A, B {
void f() override;
void g() override;
};
On entry to C::g, [ER]CX will contain the address of C's B subobject,
and C::g will have to subtract sizeof(A) to recover a pointer to C.
Before this change, we applied this adjustment in the prologue and
stored the new value into the "this" local variable alloca used for
debug info. However, MSVC does not do this, presumably because it is
often profitable to fold the adjustment into later field accesses. This
creates a problem, because the debugger expects the variable to be
unadjusted. Unfortunately, CodeView doesn't have anything like DWARF
expressions for computing variables that aren't in the program anymore,
so we have to declare 'this' to be the unadjusted value if we want the
debugger to see the right value.
This has the side benefit that, in optimized builds, the 'this' pointer
will usually be available on function entry because it doesn't require
any adjustment.
Reviewers: hans
Subscribers: aprantl, cfe-commits
Differential Revision: https://reviews.llvm.org/D40109
llvm-svn: 318440
Summary:
Constant samplers are handled as static variables and clang's code generation
library, which leads to llvm::unreachable. We bypass emitting sampler variable
as static since it's translated to a function call later.
Reviewers: yaxunl, Anastasia
Reviewed By: yaxunl, Anastasia
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D34342
llvm-svn: 318290
LLVM exposes a file in the backend (X86TargetParser.def) that
contains information about the correct list of CpuIs values.
This patch removes 2 of the copied and pasted versions of this
list from clang and instead includes the data from the .def file.
Differential Revision: https://reviews.llvm.org/D40054
llvm-svn: 318234
Lifting from Bob Wilson's notes: The hash value that we compute and
store in PGO profile data to detect out-of-date profiles does not
include enough information. This means that many significant changes to
the source will not cause compiler warnings about the profile being out
of date, and worse, we may continue to use the outdated profile data to
make bad optimization decisions. There is some tension here because
some source changes won't affect PGO and we don't want to invalidate the
profile unnecessarily.
This patch adds a new hashing scheme which is more sensitive to loop
nesting, conditions, and out-of-order control flow. Here are examples
which show snippets which get the same hash under the current scheme,
and different hashes under the new scheme:
Loop Nesting Example
--------------------
// Snippet 1
while (foo()) {
while (bar()) {}
}
// Snippet 2
while (foo()) {}
while (bar()) {}
Condition Example
-----------------
// Snippet 1
if (foo())
bar();
baz();
// Snippet 2
if (foo())
bar();
else
baz();
Out-of-order Control Flow Example
---------------------------------
// Snippet 1
while (foo()) {
if (bar()) {}
baz();
}
// Snippet 2
while (foo()) {
if (bar())
continue;
baz();
}
In each of these cases, it's useful to differentiate between the
snippets because swapping their profiles gives bad optimization hints.
The new hashing scheme considers some logical operators in an effort to
detect more changes in conditions. This isn't a perfect scheme. E.g, it
does not produce the same hash for these equivalent snippets:
// Snippet 1
bool c = !a || b;
if (d && e) {}
// Snippet 2
bool f = d && e;
bool c = !a || b;
if (f) {}
This would require an expensive data flow analysis. Short of that, the
new hashing scheme looks reasonably complete, based on a scan over the
statements we place counters on.
Profiles which use the old version of the PGO hash remain valid and can
be used without issue (there are tests in tree which check this).
rdar://17068282
Differential Revision: https://reviews.llvm.org/D39446
llvm-svn: 318229
This updates -mcount to use the new attribute names (LLVM r318195), and
switches over -finstrument-functions to also use these attributes rather
than inserting instrumentation in the frontend.
It also adds a new flag, -finstrument-functions-after-inlining, which
makes the cygprofile instrumentation get inserted after inlining rather
than before.
Differential Revision: https://reviews.llvm.org/D39331
llvm-svn: 318199
Summary: Currently the -fdebug-pass-manager flag for clang doesn't enable the debug logging in the analysis managers. This is different than what the switch does when passed to opt.
Reviewers: chandlerc
Reviewed By: chandlerc
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D40007
llvm-svn: 318140
Alexey Bataev