Summary:
This patch changes the layout of DoubleAPFloat, and adjust all
operations to do either:
1) (IEEEdouble, IEEEdouble) -> (uint64_t, uint64_t) -> PPCDoubleDoubleImpl,
then run the old algorithm.
2) Do the right thing directly.
1) includes multiply, divide, remainder, mod, fusedMultiplyAdd, roundToIntegral,
convertFromString, next, convertToInteger, convertFromAPInt,
convertFromSignExtendedInteger, convertFromZeroExtendedInteger,
convertToHexString, toString, getExactInverse.
2) includes makeZero, makeLargest, makeSmallest, makeSmallestNormalized,
compare, bitwiseIsEqual, bitcastToAPInt, isDenormal, isSmallest,
isLargest, isInteger, ilogb, scalbn, frexp, hash_value, Profile.
I could split this into two patches, e.g. use
1) for all operatoins first, then incrementally change some of them to
2). I didn't do that, because 1) involves code that converts data between
PPCDoubleDoubleImpl and (IEEEdouble, IEEEdouble) back and forth, and may
pessimize the compiler. Instead, I find easy functions and use
approach 2) for them directly.
Next step is to implement move multiply and divide from 1) to 2). I don't
have plans for other functions in 1).
Differential Revision: https://reviews.llvm.org/D27872
llvm-svn: 292839
For a << b (as original vec_sl does), if b >= sizeof(a) * 8, the
behavior is undefined. However, Power instructions do define the
behavior, which is equivalent to a << (b % (sizeof(a) * 8)).
This patch changes altivec.h to use a << (b % (sizeof(a) * 8)), to
ensure the consistent semantic of the instructions. Then it combines
the generated multiple instructions back to a single shift.
This patch handles left shift only. Right shift, on the other hand, is
more complicated, considering arithematic/logical right shift.
Differential Revision: https://reviews.llvm.org/D28037
llvm-svn: 292659
Summary: LTO backend will not invoke SampleProfileLoader pass even if -fprofile-sample-use is specified. This patch passes the flag down so that pass manager can add the SampleProfileLoader pass correctly.
Reviewers: mehdi_amini, tejohnson
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D28588
llvm-svn: 291870
There is a synchronization point between the reference count of a block dropping to zero and it's destruction, which TSan does not observe. Do not report errors in the compiler-emitted block destroy method and everything called from it.
This is similar to https://reviews.llvm.org/D25857
Differential Revision: https://reviews.llvm.org/D28387
llvm-svn: 291868
clang has generated correct IR for char/short decrement since r126816,
but we didn't have any test coverage for decrement.
Patch by Andrew Rogers.
llvm-svn: 291805
Summary: LTO backend will not invoke SampleProfileLoader pass even if -fprofile-sample-use is specified. This patch passes the flag down so that pass manager can add the SampleProfileLoader pass correctly.
Reviewers: mehdi_amini, tejohnson
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D28588
llvm-svn: 291774
Summary:
In order to simplify distributed build system integration, where actions
may be scheduled before the Thin Link which determines the list of
objects selected by the linker. The gold plugin currently will emit
0-sized index files for objects not selected by the link, to enable
checking for expected output files by the build system. If the build
system then schedules a backend action for these bitcode files, we want
to be able to fall back to normal compilation instead of failing.
Fallback is enabled under an option in LLVM (D28410), in which case a
nullptr is returned from llvm::getModuleSummaryIndexForFile. Clang can
just proceed with non-ThinLTO compilation in that case.
I am investigating whether this can be addressed in our build system,
but that is a longer term fix and so this enables a workaround in the
meantime.
Reviewers: mehdi_amini
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D28362
llvm-svn: 291303
Summary:
This patch makes the type_mismatch static data 7 bytes smaller (and it
ends up being 16 bytes smaller due to alignment restrictions, at least
on some x86-64 environments).
It revs up the type_mismatch handler version since we're breaking binary
compatibility. I will soon post a patch for the compiler-rt side.
Reviewers: rsmith, kcc, vitalybuka, pgousseau, gbedwell
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28242
llvm-svn: 291236
This test would force the execution of the backend. However, the
backend already has a test for this. Effectively, this was trying to
test that an API call was made properly. We do not have a good way to
really test this. The test itself tested very little.
Addresses post-commit review comments from Eric Christopher.
llvm-svn: 291208
Add builtins for the functions and custom codegen mapping the builtins to their
corresponding intrinsics and handling the endian related swapping.
https://reviews.llvm.org/D26546
llvm-svn: 291179
It seems that the ARM buildbots do not include x86 support. However,
other x86 targets do not support the ARM target. Use a x86 triple and
require the registered target.
llvm-svn: 291142
inline assembly may use the `.include` directive to include other
content into the file. Without the integrated assembler, the `-I` group
gets passed to the assembler. Emulate this by collecting the header
search paths and passing them to the IAS.
Resolves PR24811!
llvm-svn: 291123
Front end component (back end changes are D27392). The vectorcall
calling convention was broken subtly in two cases. First,
it didn't properly handle homogeneous vector aggregates (HVAs).
Second, the vectorcall specification requires that only the
first 6 parameters be eligible for register assignment.
This patch fixes both issues.
Differential Revision: https://reviews.llvm.org/D27529
llvm-svn: 291041
The special case to widen the integer literal zero when passed to
variadic function calls should only apply to variadic functions, not
unprototyped functions. This is consistent with what MSVC does. In this
test case, MSVC uses a 4-byte store to pass the 5th argument to 'kr' and
an 8-byte store to pass the zero to 'v':
void v(int, ...);
void kr();
void f(void) {
v(1, 2, 3, 4, 0);
kr(1, 2, 3, 4, 0);
}
Aaron Ballman discovered this issue in https://reviews.llvm.org/D28166
llvm-svn: 290906
Our newly aggressive constant folding logic makes it possible for
CGExprConstant to see the same CompoundLiteralExpr more than once. So,
emitting a new GlobalVariable every time we see a CompoundLiteral is no
longer correct.
We had a similar issue with BlockExprs that was caught while testing
said aggressive folding, so I applied the same style of fix (see D26410)
here. If we find yet another case where this needs to happen, we should
probably refactor this so we don't have a third DenseMap+getter+setter.
As a design note: getAddrOfConstantCompoundLiteralIfEmitted is really
only intended to be called by ConstExprEmitter::EmitLValue. So,
returning a GlobalVariable* instead of a ConstantAddress costs us
effectively nothing, and saves us either a few bytes per entry in our
map or a bit of code duplication.
llvm-svn: 290661
This is kind of funny because I specifically did work to make this easy
and then it didn't actually get implemented.
I've also ported a set of tests that rely on this functionality to run
with the new PM as well as the old PM so that we don't mess this up in
the future.
llvm-svn: 290558
manager, and a code path to use it.
The option is actually a top-level option but does contain
'experimental' in the name. This is the compromise suggested by Richard
in discussions. We expect this option will be around long enough and
have enough users towards the end that it merits not being relegated to
CC1, but it still needs to be clear that this option will go away at
some point.
The backend code is a fresh codepath dedicated to handling the flow with
the new pass manager. This was also Richard's suggested code structuring
to essentially leave a clean path for development rather than carrying
complexity or idiosyncracies of how we do things just to share code with
the parts of this in common with the legacy pass manager. And it turns
out, not much is really in common even though we use the legacy pass
manager for codegen at this point.
I've switched a couple of tests to run with the new pass manager, and
they appear to work. There are still plenty of bugs that need squashing
(just with basic experiments I've found two already!) but they aren't in
this code, and the whole point is to expose the necessary hooks to start
experimenting with the pass manager in more realistic scenarios.
That said, I want to *strongly caution* anyone itching to play with
this: it is still *very shaky*. Several large components have not yet
been shaken down. For example I have bugs in both the always inliner and
inliner that I have already spotted and will be fixing independently.
Still, this is a fun milestone. =D
One thing not in this patch (but that might be very reasonable to add)
is some level of support for raw textual pass pipelines such as what
Sean had a patch for some time ago. I'm mostly interested in the more
traditional flow of getting the IR out of Clang and then running it
through opt, but I can see other use cases so someone may want to add
it.
And of course, *many* features are not yet supported!
- O1 is currently more like O2
- None of the sanitizers are wired up
- ObjC ARC optimizer isn't wired up
- ...
So plenty of stuff still lef to do!
Differential Revision: https://reviews.llvm.org/D28077
llvm-svn: 290450
Summary:
We compile user opencl kernel code with spir triple. But built-ins are written in OpenCL and we compile it with triple x86_64 to be able to use x86 intrinsics. And we need address spaces to match in both cases. So, we change fake address space map in OpenCL for matching with spir.
On CPU address spaces are not really important but we'd like to preserve address space information in order to perform optimizations relying on this info like enhanced alias analysis.
Reviewers: pekka.jaaskelainen, Anastasia
Subscribers: pekka.jaaskelainen, yaxunl, bader, cfe-commits
Differential Revision: https://reviews.llvm.org/D28048
llvm-svn: 290436
-fno-inline-functions, -O0, and optnone.
These were really, really tangled together:
- We used the noinline LLVM attribute for -fno-inline
- But not for -fno-inline-functions (breaking LTO)
- But we did use it for -finline-hint-functions (yay, LTO is happy!)
- But we didn't for -O0 (LTO is sad yet again...)
- We had weird structuring of CodeGenOpts with both an inlining
enumeration and a boolean. They interacted in weird ways and
needlessly.
- A *lot* of set smashing went on with setting these, and then got worse
when we considered optnone and other inlining-effecting attributes.
- A bunch of inline affecting attributes were managed in a completely
different place from -fno-inline.
- Even with -fno-inline we failed to put the LLVM noinline attribute
onto many generated function definitions because they didn't show up
as AST-level functions.
- If you passed -O0 but -finline-functions we would run the normal
inliner pass in LLVM despite it being in the O0 pipeline, which really
doesn't make much sense.
- Lastly, we used things like '-fno-inline' to manipulate the pass
pipeline which forced the pass pipeline to be much more
parameterizable than it really needs to be. Instead we can *just* use
the optimization level to select a pipeline and control the rest via
attributes.
Sadly, this causes a bunch of churn in tests because we don't run the
optimizer in the tests and check the contents of attribute sets. It
would be awesome if attribute sets were a bit more FileCheck friendly,
but oh well.
I think this is a significant improvement and should remove the semantic
need to change what inliner pass we run in order to comply with the
requested inlining semantics by relying completely on attributes. It
also cleans up tho optnone and related handling a bit.
One unfortunate aspect of this is that for generating alwaysinline
routines like those in OpenMP we end up removing noinline and then
adding alwaysinline. I tried a bunch of other approaches, but because we
recompute function attributes from scratch and don't have a declaration
here I couldn't find anything substantially cleaner than this.
Differential Revision: https://reviews.llvm.org/D28053
llvm-svn: 290398
Much to my surprise, '-disable-llvm-optzns' which I thought was the
magical flag I wanted to get at the raw LLVM IR coming out of Clang
deosn't do that. It still runs some passes over the IR. I don't want
that, I really want the *raw* IR coming out of Clang and I strongly
suspect everyone else using it is in the same camp.
There is actually a flag that does what I want that I didn't know about
called '-disable-llvm-passes'. I suspect many others don't know about it
either. It both does what I want and is much simpler.
This removes the confusing version and makes that spelling of the flag
an alias for '-disable-llvm-passes'. I've also moved everything in Clang
to use the 'passes' spelling as it seems both more accurate (*all* LLVM
passes are disabled, not just optimizations) and much easier to remember
and spell correctly.
This is part of simplifying how Clang drives LLVM to make it cleaner to
wire up to the new pass manager.
Differential Revision: https://reviews.llvm.org/D28047
llvm-svn: 290392
This is a recommit of r290149, which was reverted in r290169 due to msan
failures. msan was failing because we were calling
`isMostDerivedAnUnsizedArray` on an invalid designator, which caused us
to read uninitialized memory. To fix this, the logic of the caller of
said function was simplified, and we now have a `!Invalid` assert in
`isMostDerivedAnUnsizedArray`, so we can catch this particular bug more
easily in the future.
Fingers crossed that this patch sticks this time. :)
Original commit message:
This patch does three things:
- Gives us the alloc_size attribute in clang, which lets us infer the
number of bytes handed back to us by malloc/realloc/calloc/any user
functions that act in a similar manner.
- Teaches our constexpr evaluator that evaluating some `const` variables
is OK sometimes. This is why we have a change in
test/SemaCXX/constant-expression-cxx11.cpp and other seemingly
unrelated tests. Richard Smith okay'ed this idea some time ago in
person.
- Uniques some Blocks in CodeGen, which was reviewed separately at
D26410. Lack of uniquing only really shows up as a problem when
combined with our new eagerness in the face of const.
llvm-svn: 290297
This commit fails MSan when running test/CodeGen/object-size.c in
a confusing way. After some discussion with George, it isn't really
clear what is going on here. We can make the MSan failure go away by
testing for the invalid bit, but *why* things are invalid isn't clear.
And yet, other code in the surrounding area is doing precisely this and
testing for invalid.
George is going to take a closer look at this to better understand the
nature of the failure and recommit it, for now backing it out to clean
up MSan builds.
llvm-svn: 290169
This patch does three things:
- Gives us the alloc_size attribute in clang, which lets us infer the
number of bytes handed back to us by malloc/realloc/calloc/any user
functions that act in a similar manner.
- Teaches our constexpr evaluator that evaluating some `const` variables
is OK sometimes. This is why we have a change in
test/SemaCXX/constant-expression-cxx11.cpp and other seemingly
unrelated tests. Richard Smith okay'ed this idea some time ago in
person.
- Uniques some Blocks in CodeGen, which was reviewed separately at
D26410. Lack of uniquing only really shows up as a problem when
combined with our new eagerness in the face of const.
Differential Revision: https://reviews.llvm.org/D14274
llvm-svn: 290149
Summary:
D27549 (partial fix for PR26619) emits a constant value in the debug
metadata for a floating-point static const that does not exceed 64
bits in size. Whether or not a long double exceeds 64 bits in size
depends on the target. Modify the test case so that it expects a
constant value for long double if and only if the long double is no
larger than 64 bits.
Reviewers: cfe-commits, probinson
Differential Revision: https://reviews.llvm.org/D27597
llvm-svn: 289686
r289225 broke AST invariants by reparenting enumerators into function
decl contexts. This improves things by only reparenting TagDecls while
also attempting to preserve the lexical declcontext chain. The
interesting example here is:
int f(struct S { enum E { a = 1 } b; } c);
The semantic contexts of E and S should be f, and the lexical context of
S should be f and the lexical context of E should be S. We didn't do
that with r289225, but now we should.
This change should also improve our behavior on this example:
void f() {
extern void ext(struct S { } o);
// S injected here
}
Before r289225 we would only remove 'S' from the surrounding tag
injection context if it was the TU, but now we properly reparent S from
f to ext.
Fixes PR31366
llvm-svn: 289678
This will allow the backend to constant fold these to generic shuffle vectors like 128-bit and 256-bit without having to working about handling masking.
llvm-svn: 289351
This will allow the backend to constant fold these to generic shuffle vectors like 128-bit and 256-bit without having to working about handling masking.
llvm-svn: 289345
Summary:
D27549 (partial fix for PR26619) emits a constant value in the debug
metadata for a floating-point static const that does not exceed 64
bits in size. The regression test accompanying that fix assumes that
a long double exceeds 64 bits in size and hence does not get a
constant value in the debug metadata. However, for some targets --
such as "--target=hexagon-unknown-elf" -- a long double does not
exceed 64 bits in size, and hence the test fails.
As a temporary fix, modify the regression test to no longer inspect
the debug metadata for a long double.
Reviewers: cfe-commits, probinson
Differential Revision: https://reviews.llvm.org/D27589
llvm-svn: 289103
Summary:
Partial fix for PR26619.
Prior to this change, a DIGlobalVariable corresponding to a static
const was marked with an expression corresponding to its constant
value only if it is of integral type. With this change, we now do the
same if it is of __fp16, float, or double type (that is,
floating-point types that do not exceed 64 bits in size, and hence are
supported easily by the existing LLVM machinery for creating constant
expressions in debug info).
Reviewers: llvm-commits
Differential Revision: https://reviews.llvm.org/D27549
llvm-svn: 289094
This solves PR23715 in a way that is compatible with LTO.
MSVC supports jumping to source-level labels and between inline asm
blocks, but we don't.
Also revert the old solution, r255201, which was to mark these calls as
noduplicate.
llvm-svn: 288059
(commit again after fixing the buildbot failures)
This adds various overloads of the following builtins to altivec.h:
vec_neg
vec_nabs
vec_adde
vec_addec
vec_sube
vec_subec
vec_subc
Note that for vec_sub builtins on 32 bit integers, the semantics is similar to
what ISA describes for instructions like vsubecuq that work on quadwords: the
first operand is added to the one's complement of the second operand. (As
opposed to two's complement which I expected).
llvm-svn: 287872
(commit again after fixing the buildbot failures)
This adds various overloads of the following builtins to altivec.h:
vec_neg
vec_nabs
vec_adde
vec_addec
vec_sube
vec_subec
vec_subc
Note that for vec_sub builtins on 32 bit integers, the semantics is similar to
what ISA describes for instructions like vsubecuq that work on quadwords: the
first operand is added to the one's complement of the second operand. (As
opposed to two's complement which I expected).
llvm-svn: 287795
This adds various overloads of the following builtins to altivec.h:
vec_neg
vec_nabs
vec_adde
vec_addec
vec_sube
vec_subec
vec_subc
Note that for vec_sub builtins on 32 bit integers, the semantics is similar to
what ISA describes for instructions like vsubecuq that work on quadwords: the
first operand is added to the one's complement of the second operand. (As
opposed to two's complement which I expected).
llvm-svn: 287772
Both the (V)CVTDQ2PD (i32 to f64) and (V)CVTUDQ2PD (u32 to f64) conversion instructions are lossless and can be safely represented as generic __builtin_convertvector calls instead of x86 intrinsics without affecting final codegen.
This patch removes the clang builtins and their use in the headers - a future patch will deal with removing the llvm intrinsics.
This is an extension patch to D20528 which dealt with the equivalent sse/avx cases.
Differential Revision: https://reviews.llvm.org/D26686
llvm-svn: 287088
Instead of always displaying the mangled name, try to do better
and get something closer to regular functions.
Recommit r287039 (that was reverted in r287039) with a tweak to
be more generic, and test fixes!
Differential Revision: https://reviews.llvm.org/D26522
llvm-svn: 287085
Instead of always displaying the mangled name, try to do better
and get something closer to regular functions.
Differential Revision: https://reviews.llvm.org/D26522
llvm-svn: 287039
This patch implements all the overloads for vec_xl_be and vec_xst_be. On BE,
they behaves exactly the same with vec_xl and vec_xst, therefore they are
simply implemented by defining a matching macro. On LE, they are implemented
by defining new builtins and intrinsics. For int/float/long long/double, it
is just a load (lxvw4x/lxvd2x) or store(stxvw4x/stxvd2x). For char/char/short,
we also need some extra shuffling before or after call the builtins to get the
desired BE order. For int128, simply call vec_xl or vec_xst.
llvm-svn: 286971
Adds 2 vector functions for converting from a vector of unsigned short to a
vector of float. One converts the low 4 halfwords and one converts the high
4 halfwords.
Differential Revision: https://reviews.llvm.org/D26534
llvm-svn: 286863
Add vector extract exponent/significand functions to altivec.h, as well as
functions (and related constants) to test the data class of vector float
and vector double.
Differential Revision: https://reviews.llvm.org/D26271
llvm-svn: 286830
This is part of a set of changes to allow InstCombine in the backend to optimize variable shifts without having to know about masking.
llvm-svn: 286757
Summary: Inverting the mask argument does not reflect the intended semantics of the intrinsic.
Reviewers: igorb, delena
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D26019
llvm-svn: 286733
This introduces a function annotation that disables TSan checking for the
function at run time. The benefit over attribute((no_sanitize("thread")))
is that the accesses within the callees will also be suppressed.
The motivation for this attribute is a guarantee given by the objective C
language that the calls to the reference count decrement and object
deallocation will be synchronized. To model this properly, we would need to
intercept all ref count decrement calls (which are very common in ObjC due
to use of ARC) and also every single message send. Instead, we propose to
just ignore all accesses made from within dealloc at run time. The main
downside is that this still does not introduce any synchronization, which
means we might still report false positives if the code that relies on this
synchronization is not executed from within dealloc. However, we have not
seen this in practice so far and think these cases will be very rare.
(This problem is similar in nature to https://reviews.llvm.org/D21609;
unfortunately, the same solution does not apply here.)
Differential Revision: https://reviews.llvm.org/D25857
llvm-svn: 286672
Add a check to the DeclCache before emitting debug info for a
GlobalVariable a second time and just attach the previsously created one to it.
<rdar://problem/26721101>
llvm-svn: 286322
This patch implements the register call calling convention, which ensures
as many values as possible are passed in registers. CodeGen changes
were committed in https://reviews.llvm.org/rL284108.
Differential Revision: https://reviews.llvm.org/D25204
llvm-svn: 285849
Commit on behalf of: Coby Tayree
1.'v' constraint for (x86) non-avx arch imitates the already implemented 'x' constraint, i.e. allows XMM{0-15} & YMM{0-15} depending on the apparent arch & mode (32/64).
2.for the avx512 arch it allows [X,Y,Z]MM{0-31} (mode dependent)
This patch applies the needed changes to clang
LLVM patch: https://reviews.llvm.org/D25005
Differential Revision: https://reviews.llvm.org/D25005
llvm-svn: 285688
Unfortunately, the backend currently doesn't fold masks into the instructions correctly when they come from these shufflevectors. I'll work on that in a future commit.
llvm-svn: 285667
Commit on behalf of mharoush
Extending inline assembly support, compatible with GCC as folowing:
"k" constraint hints the compiler to select any of AVX512 k0-k7 registers.
"Yk" constraint is a subset of "k" excluding k0 which is not allowd to be used as a mask.
Reviewer: 1. rnk
Differential Revision: https://reviews.llvm.org/D25063
llvm-svn: 285604
Commit on behalf of mharoush
After LGTM and check all:
This patch is a compatibility fix for clang, matching GCC support for charter escape when using extended in-line assembly (i.e, "%{" ,"%}" --> "{" ,"}" ).
It is meant to enable support for advanced features such as AVX512 conditional\masked vector instructions/broadcast assembly syntax.
Reviewer: 1. rnk
Differential Revision: https://reviews.llvm.org/D25012
llvm-svn: 285585
For compatibility with other compilers on the platform, allow specifying
levels of the z/Architecture instead of model names with -march. In
particular, the following aliases are now supported:
-march=arch8 equals -march=z10
-march=arch9 equals -march=z196
-march=arch10 equals -march=zEC12
-march=arch11 equals -march=z13
This parallels the equivalent (and prerequisite) LLVM change in r285577.
llvm-svn: 285578
Commit on behalf of mharoush
After LGTM and check all:
This patch enables usage of k registers in inline assembly syntax.
Adding triple
Reviewer: 1. rnk
2. delena
Differential Revision: https://reviews.llvm.org/D25011
llvm-svn: 285563
Commit on behalf of mharoush
After LGTM and check all:
This patch enables usage of k registers in inline assembly syntax.
Reviewer: 1. rnk
2. delena
Differential Revision: https://reviews.llvm.org/D25011
llvm-svn: 285555
__builtin_alloca always uses __BIGGEST_ALIGNMENT__ for the alignment of
the allocation. __builtin_alloca_with_align allows the programmer to
specify the alignment of the allocation.
This fixes PR30658.
llvm-svn: 285544
Unfortunately, the backend currently doesn't fold masks into the instructions correctly when they come from these shufflevectors. I'll work on that in a future commit.
llvm-svn: 285540
After LGTM and Check-all
Vector-reduction arithmetic accepts vectors as inputs and produces
scalars as outputs.This class of vector operation forms the basis
of many scientific computations. In vector-reduction arithmetic,
the evaluation off is independent of the order of the input elements of V.
Reviewer: 1. craig.topper
2. igorb
Differential Revision: https://reviews.llvm.org/D25988
llvm-svn: 285493
GCC documents __builtin_alloca as aligning the storage to at least
__BIGGEST_ALIGNMENT__.
MSVC documents essentially the same for the x64 ABI:
https://msdn.microsoft.com/en-us/library/x9sx5da1.aspx
The 32-bit ABI follows the same rule: it emits a call to _alloca_probe_16
Differential Revision: https://reviews.llvm.org/D24378
llvm-svn: 285316
Summary:
Current generation of lifetime intrinsics does not handle cases like:
```
{
char x;
l1:
bar(&x, 1);
}
goto l1;
```
We will get code like this:
```
%x = alloca i8, align 1
call void @llvm.lifetime.start(i64 1, i8* nonnull %x)
br label %l1
l1:
%call = call i32 @bar(i8* nonnull %x, i32 1)
call void @llvm.lifetime.end(i64 1, i8* nonnull %x)
br label %l1
```
So the second time bar was called for x which is marked as dead.
Lifetime markers here are misleading so it's better to remove them at all.
This type of bypasses are rare, e.g. code detects just 8 functions building
clang (2329 targets).
PR28267
Reviewers: eugenis
Subscribers: beanz, mgorny, cfe-commits
Differential Revision: https://reviews.llvm.org/D24693
llvm-svn: 285176
Committed after LGTM and check-all
Vector-reduction arithmetic accepts vectors as inputs and produces scalars as outputs.
This class of vector operation forms the basis of many scientific computations.
In vector-reduction arithmetic, the evaluation off is independent of the order of the input elements of V.
Used bisection method. At each step, we partition the vector with previous
step in half, and the operation is performed on its two halves.
This takes log2(n) steps where n is the number of elements in the vector.
Reviwer: 1. igorb
2. craig.topper
Differential Revision: https://reviews.llvm.org/D25527
llvm-svn: 285054
This reverts commit r285007 and reapply r284990, with a fix for the
opencl test that I broke. Original commit message follows:
These new builtins support a mechanism for logging OS events, using a
printf-like format string to specify the layout of data in a buffer.
The _buffer_size version of the builtin can be used to determine the size
of the buffer to allocate to hold the data, and then __builtin_os_log_format
can write data into that buffer. This implements format checking to report
mismatches between the format string and the data arguments. Most of this
code was written by Chris Willmore.
Differential Revision: https://reviews.llvm.org/D25888
llvm-svn: 285019
These new builtins support a mechanism for logging OS events, using a
printf-like format string to specify the layout of data in a buffer.
The _buffer_size version of the builtin can be used to determine the size
of the buffer to allocate to hold the data, and then __builtin_os_log_format
can write data into that buffer. This implements format checking to report
mismatches between the format string and the data arguments. Most of this
code was written by Chris Willmore.
Differential Revision: https://reviews.llvm.org/D25888
llvm-svn: 284990
Committed after LGTM and check-all
Vector-reduction arithmetic accepts vectors as inputs and produces scalars as outputs.
This class of vector operation forms the basis of many scientific computations.
In vector-reduction arithmetic, the evaluation off is independent of the order of the input elements of V.
Used bisection method. At each step, we partition the vector with previous
step in half, and the operation is performed on its two halves.
This takes log2(n) steps where n is the number of elements in the vector.
Differential Revision: https://reviews.llvm.org/D25527
llvm-svn: 284963
Tests fall into one of the following categories:
- The requirement was unnecessary
- Additional quoting was required for backslashes in paths (see "sed -e
's/\\/\\\\/g'") in the sanitizer tests.
- OpenMP used 'REQUIRES: shell' as a proxy for the test failing on
Windows. Those tests fail there reliably, so use XFAIL instead.
I tried not to remove shell requirements that were added to suppress
flaky test failures, but if I screwed up, we can add it back as needed.
llvm-svn: 284793
Preparation to implement DW_AT_alignment support:
- We pass non-zero align value to DIBuilder only when alignment was forced
- Modify tests to match this change
Differential Revision: https://reviews.llvm.org/D24426
llvm-svn: 284679
This patch teaches clang to range check immediates for MIPS MSA instrinsics.
This checking is done strictly in comparison to some existing GCC
implementations. E.g. msa_andvi_b(var, 257) does not result in andvi $wX, 1.
Similarily msa_ldi_b takes a range of -128 to 127.
As part of this effort, correct the existing MSA test as it has both illegal
types and immediates.
Reviewers: vkalintiris
Differential Revision: https://reviews.llvm.org/D25017
llvm-svn: 284620
Gcc prints error if elements of left and right parts of a shift have different
sizes. This patch is provided the GCC compatibility.
Patch by Vladimir Yakovlev.
Differential Revision: https://reviews.llvm.org/D24669
llvm-svn: 284579
Summary: We need `__stosb` to be an intrinsic, because SecureZeroMemory function uses it without including intrin.h. Implementing it as a volatile memset is not consistent with MSDN specification, but it gives us target-independent IR while keeping the most important properties of `__stosb`.
Reviewers: rnk, hans, thakis, majnemer
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D25334
llvm-svn: 284253
The backtrace on the bot does not give me any indication what is wrong.
The test case interestingly passes in stage2 of the build.
I don't have a way of debugging this.
Disable the test on windows and hope if there is truly a bug in the code that
was causing we will eventually run into this on other platforms.
llvm-svn: 284174
Summary: Previously global 64-bit versions of _Interlocked functions broke buildbots on i386, so now I'm adding them as builtins for x86-64 and ARM only (should they be also on AArch64? I had problems with testing it for AArch64, so I left it)
Reviewers: hans, majnemer, mstorsjo, rnk
Subscribers: cfe-commits, aemerson
Differential Revision: https://reviews.llvm.org/D25576
llvm-svn: 284172
Summary: _BitScan intrinsics (and some others, for example _Interlocked and _bittest) are supposed to work on both ARM and x86. This is an attempt to isolate them, avoiding repeating their code or writing separate function for each builtin.
Reviewers: hans, thakis, rnk, majnemer
Subscribers: RKSimon, cfe-commits, aemerson
Differential Revision: https://reviews.llvm.org/D25264
llvm-svn: 284060
The backend now has the capability to save information from optimizations, the
same information that can be used to generate optimization diagnostics but in
machine-consumable form, into an output file. This can be enabled when using
opt (see r282539), and this change enables it when using clang. The idea is
that other tools will be able to consume these files, and perhaps in
combination with the original source code, produce various kinds of
optimization reports for users (and for compiler developers).
We now have at-least two tools that can consume these files:
* tools/llvm-opt-report
* utils/opt-viewer
Using the flag -fsave-optimization-record will cause the YAML file to be
generated; the file name will be based on the output file name (if we're using
-c or -S and have an output name), or the input file name. When we're using
CUDA, or some other offloading mechanism, separate files are generated for each
backend target. The output file name can be specified by the user using
-foptimization-record-file=filename.
Differential Revision: https://reviews.llvm.org/D25225
llvm-svn: 283834
Commit in the name of: Coby Tayree
1.'v' constraint for (x86) non-avx arch imitates the already implemented 'x' constraint, i.e. allows XMM{0-15} & YMM{0-15} depending on the apparent arch & mode (32/64).
2.for the avx512 arch it allows [X,Y,Z]MM{0-31} (mode dependent)
This patch applies the needed changes to clang
LLVM patch: https://reviews.llvm.org/D25005
Differential Revision: D25004
llvm-svn: 283716
Summary: We need x86-64-specific builtins if we want to implement some of the MS intrinsics - winnt.h contains definitions of some functions for i386, but not for x86-64 (for example _InterlockedOr64), which means that we cannot treat them as builtins for both i386 and x86-64, because then we have definitions of builtin functions in winnt.h on i386.
Reviewers: thakis, majnemer, hans, rnk
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D24598
llvm-svn: 283264
The motivation for the change is that we can't have pseudo-global settings
for codegen living in TargetOptions because that doesn't work with LTO.
Ideally, these reciprocal attributes will be moved to the instruction-level
via FMF, metadata, or something else. But making them function attributes is
at least an improvement over the current state.
I'm committing this patch ahead of the related LLVM patch to avoid bot failures,
but if that patch needs to be reverted, then this should be reverted too.
Differential Revision: https://reviews.llvm.org/D24815
llvm-svn: 283251
Enable soft-float support on PPC64, as the backend now supports it. Also, the
backend now uses -hard-float instead of +soft-float, so set the target features
accordingly.
Fixes PR26970.
llvm-svn: 283061
These are supposed to produce the same as normal volatile
pointer loads/stores. When -volatile:ms is specified,
normal volatile pointers are forced to have atomic semantics
(as is the default on x86 in MSVC mode). In that case,
these builtins should still produce non-atomic volatile
loads/stores without acquire/release semantics, which
the new test verifies.
These are only available on ARM (and on AArch64,
although clang doesn't support AArch64/Windows yet).
This implements what is missing for PR30394, making it possible
to compile C++ for ARM in MSVC mode with MSVC headers.
Differential Revision: https://reviews.llvm.org/D24986
llvm-svn: 282900
The X86 clang/test/CodeGen/*builtins.c tests define the mm_malloc.h include
guard as a hack for avoiding its inclusion (mm_malloc.h requires a hosted
environment since it expects stdlib.h to be available - which is not the case
in these internal clang codegen tests).
This patch removes this hack and instead passes -ffreestanding to clang cc1.
Differential Revision: https://reviews.llvm.org/D24825
llvm-svn: 282581
This patch corresponds to review:
https://reviews.llvm.org/D24397
It adds the __POWER9_VECTOR__ macro and the -mpower9-vector option along with
a number of altivec.h functions (refer to the code review for a list).
llvm-svn: 282481
Clang has the default FP contraction setting of “-ffp-contract=on”, which
doesn't really mean “on” in the conventional sense of the word, but rather
really means “according to the per-statement effective value of the relevant
pragma”.
Before this patch, Clang has that pragma defaulting to “off”. Since the
“-ffp-contract=on” mode is really an AND of two booleans and the second of them
defaults to “off”, the whole thing effectively defaults to “off”. This patch
changes the default value of the pragma to “on”, thus making the default pair of
booleans (on, on) rather than (on, off). This makes FP optimization slightly
more aggressive than before when not using either “-Ofast”, “-ffast-math”, or
“-ffp-contract=fast”. Even with this patch the compiler still respects
“-ffp-contract=off”.
As per a suggestion by Steve Canon, the added code does _not_ require “-O3” or
higher. This is so as to try our best to preserve identical floating-point
results for unchanged source code compiling for an unchanged target when only
changing from any optimization level in the set (“-O0”, “-O1”, “-O2”, “-O3”) to
any other optimization level in that set. “-Os” and “-Oz” seem to be behaving
identically, i.e. should probably be considered a part of the aforementioned
set, but I have not reviewed this rigorously. “-Ofast” is explicitly _not_ a
member of that set.
Patch authored by Abe Skolnik [a.skolnik@samsung.com] and Stephen Canon [scanon@apple.com].
Differential Revision: https://reviews.llvm.org/D24481
llvm-svn: 282259
This patch adds the msa.h header file containing the shorter names for the
MSA instrinsics, e.g. msa_sll_b for builtin_msa_sll_b.
Reviewers: vkalintiris, zoran.jovanovic
Differential Review: https://reviews.llvm.org/D24674
llvm-svn: 281975
Summary: We previously relies on InstructionCombining pass to remove invoke instructions. Now that we can inline invoke instructions correctly, we do not need these passes any more.
Reviewers: dnovillo
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D24730
llvm-svn: 281910
r278501 inadvertently introduced a bug in which it disallowed shifting
scalar operands by vector operands when not compiling for OpenCL. This
commit fixes it.
Patch by Vladimir Yakovlev.
Differential Revision: https://reviews.llvm.org/D24467
llvm-svn: 281669
Summary: This patch converts finite/__finite to builtin functions so that it will be inlined by compiler.
Reviewers: hfinkel, davidxl, efriedma
Subscribers: efriedma, llvm-commits
Differential Revision: https://reviews.llvm.org/D24483
llvm-svn: 281509
This patch makes us act more conservatively when trying to determine
the objectsize for an array at the end of an object. This is in
response to code like the following:
```
struct sockaddr {
/* snip */
char sa_data[14];
};
void foo(const char *s) {
size_t slen = strlen(s) + 1;
size_t added_len = slen <= 14 ? 0 : slen - 14;
struct sockaddr *sa = malloc(sizeof(struct sockaddr) + added_len);
strcpy(sa->sa_data, s);
// ...
}
```
`__builtin_object_size(sa->sa_data, 1)` would return 14, when there
could be more than 14 bytes at `sa->sa_data`.
Code like this is apparently not uncommon. FreeBSD's manual even
explicitly mentions this pattern:
https://www.freebsd.org/doc/en/books/developers-handbook/sockets-essential-functions.html
(section 7.5.1.1.2).
In light of this, we now just give up on any array at the end of an
object if we can't find the object's initial allocation.
I lack numbers for how much more conservative we actually become as a
result of this change, so I chose the fix that would make us as
compatible with GCC as possible. If we want to be more aggressive, I'm
happy to consider some kind of whitelist or something instead.
llvm-svn: 281277
r280553 introduced an issue where we'd emit ambiguity errors for code
like:
```
void foo(int *, int);
void foo(unsigned int *, unsigned int);
void callFoo() {
unsigned int i;
foo(&i, 0); // ambiguous: int->unsigned int is worse than int->int,
// but unsigned int*->unsigned int* is better than
// int*->int*.
}
```
This patch fixes this issue by changing how we handle ill-formed (but
valid) implicit conversions. Candidates with said conversions now always
rank worse than candidates without them, and two candidates are
considered to be equally bad if they both have these conversions for
the same argument.
Additionally, this fixes a case in C++11 where we'd complain about an
ambiguity in a case like:
```
void f(char *, int);
void f(const char *, unsigned);
void g() { f("abc", 0); }
```
...Since conversion to char* from a string literal is considered
ill-formed in C++11 (and deprecated in C++03), but we accept it as an
extension.
llvm-svn: 280847
Clang tests for verifying the following syntaxes:
1. 0xNN and NNh are accepted as valid hexadecimal numbers, but 0xNNh is not.
0xNN and NNh may come with optional U or L suffix.
2. NNb is accepted as a valid binary (base-2) number, but 0bNN is not.
NNb may come with optional U or L suffix.
Differential Revision: https://reviews.llvm.org/D22112
llvm-svn: 280556
This patch allows us to perform incompatible pointer conversions when
resolving overloads in C. So, the following code will no longer fail to
compile (though it will still emit warnings, assuming the user hasn't
opted out of them):
```
void foo(char *) __attribute__((overloadable));
void foo(int) __attribute__((overloadable));
void callFoo() {
unsigned char bar[128];
foo(bar); // selects the char* overload.
}
```
These conversions are ranked below all others, so:
A. Any other viable conversion will win out
B. If we had another incompatible pointer conversion in the example
above (e.g. `void foo(int *)`), we would complain about
an ambiguity.
Differential Revision: https://reviews.llvm.org/D24113
llvm-svn: 280553
Since some profiling tools, such as gprof, ftrace, and uftrace, use
-pg option to generate a mcount function call at the entry of each
function. Function invocation can be detected by this hook function.
But mcount insertion is done before function inlining phase in clang,
sometime a function that already has a mcount call can be inlined in the
middle of another function.
This patch adds an attribute "counting-function" to each function
rather than emitting the mcount call directly in frontend so that this
attribute can be processed in backend. Then the mcount calls can be
properly inserted in backend after all the other optimizations are
completed.
Link: https://llvm.org/bugs/show_bug.cgi?id=28660
Reviewers: hans, rjmccall, hfinkel, rengolin, compnerd
Subscribers: shenhan, cfe-commits
Differential Revision: https://reviews.llvm.org/D22666
llvm-svn: 280355
-fprofile-dir=path allows the user to specify where .gcda files should be
emitted when the program is run. In particular, this is the first flag that
causes the .gcno and .o files to have different paths, LLVM is extended to
support this. -fprofile-dir= does not change the file name in the .gcno (and
thus where lcov looks for the source) but it does change the name in the .gcda
(and thus where the runtime library writes the .gcda file). It's different from
a GCOV_PREFIX because a user can observe that the GCOV_PREFIX_STRIP will strip
paths off of -fprofile-dir= but not off of a supplied GCOV_PREFIX.
To implement this we split -coverage-file into -coverage-data-file and
-coverage-notes-file to specify the two different names. The !llvm.gcov
metadata node grows from a 2-element form {string coverage-file, node dbg.cu}
to 3-elements, {string coverage-notes-file, string coverage-data-file, node
dbg.cu}. In the 3-element form, the file name is already "mangled" with
.gcno/.gcda suffixes, while the 2-element form left that to the middle end
pass.
llvm-svn: 280306
These will be reused when removing some builtins from avx512vldqintrin.h and this will make the tests for that change show a better number of vector elements.
llvm-svn: 280196
to CC1, which are translated to function attributes and can e.g. be mapped on
build attributes FP_exceptions and FP_denormal. Setting these build attributes
allows better selection of floating point libraries.
Differential Revision: https://reviews.llvm.org/D23840
llvm-svn: 280064
The PPC64 DWARF register-size table did not match the ABI specification (or
GCC, for that matter). Fix that, and add a regression test.
Fixes PR27931.
llvm-svn: 280053
In cases where .dwo/.dwp files are guaranteed to be available, skipping
the extra online (in the .o file) inline info can save a substantial
amount of space - see the original r221306 for more details there.
llvm-svn: 279651
We processed unnamed bitfields after our logic for non-vector field
elements in records larger than 128 bits. The vector logic would
determine that the bit-field disqualifies the record from occupying a
register despite the unnamed bit-field not participating in the record
size nor its alignment.
N.B. This behavior matches GCC and ICC.
llvm-svn: 278656
An __m512 vector type wrapped in a structure should be passed in a
vector register.
Our prior implementation was based on a draft version of the psABI.
This fixes PR28975.
N.B. The update to the ABI was made here:
https://github.com/hjl-tools/x86-psABI/commit/30f9c9
llvm-svn: 278655
constraints were added to _mm256_broadcast_{pd,ps} intel intrinsics.
The spec for these intrinics is ... pretty much silent on alignment.
This is especially frustrating considering the amount of discussion of
alignment in the load and store instrinsics. So I was forced to rely on
the specification for the VBROADCASTF128 instruction.
That instruction's spec is *also* completely silent on alignment.
Fortunately, when it comes to the instruction's spec, silence is enough.
There is no #GP fault option for an underaligned address so this
instruction, and by inference the intrinsic, can read any alignment.
As it happens, the old code worked exactly this way and in fact we have
plenty of code that hands pointers with less than 16-byte alignment to
these intrinsics. This code broke pretty spectacularly with this commit.
Fortunately, the fix is super simple! Change a 16 to a 1, and ta da!
Anyways, a lot of debugging for a really boring fix. =]
llvm-svn: 278202
Summary:
Based on a patch by Michael Mueller.
This attribute specifies that a function can be hooked or patched. This
mechanism was originally devised by Microsoft for hotpatching their
binaries (which they're constantly updating to stay ahead of crackers,
script kiddies, and other ne'er-do-wells on the Internet), but it's now
commonly abused by Windows programs that want to hook API functions. It
is for this reason that this attribute was added to GCC--hence the name,
`ms_hook_prologue`.
Depends on D19908.
Reviewers: rnk, aaron.ballman
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D19909
llvm-svn: 278050
Tim Shen