VPMACSDQH/VPMACSDQL act as VPADDQ( VPMULDQ( x, y ), z ) - multiply+extending either the odd/even 4i32 input elements and adding to v2i64 accumulator
llvm-svn: 292020
Add LIBUNWIND_* directories to include path only if they were actually
found, in order to fix the CMake error. Both of the directories are
usually unnecessary since libcxxabi uses only the common part of
unwind.h that is supplied both by GCC and Clang.
Differential Revision: https://reviews.llvm.org/D25314
llvm-svn: 292018
Tests showing missed opportunities to use XOP's integer fma instructions
Some of these are pretty awkward to match as they often have implicit sext/trunc stages but many just ignore overflow bits which makes things pretty straightforward.
llvm-svn: 292017
Correct handling of the following FileCheck options is implemented in
update_llc_test_checks.py and update_test_checks.py scripts:
1) -check-prefix (with a single dash)
2) -check-prefixes (with multiple prefixes)
Differential Revision: https://reviews.llvm.org/D28572
llvm-svn: 292008
Summary: This patch allows libc++ to be built against the debug MSVC runtimes instead of just the release ones.
Reviewers: rnk, majnemer, compnerd, smeenai
Subscribers: mgorny, cfe-commits
Differential Revision: https://reviews.llvm.org/D28725
llvm-svn: 292006
Isel now selects masked move instructions for vselect instead of blendm. But sometimes it beneficial to register allocation to remove the tied register constraint by using blendm instructions.
This also picks up cases where the masked move was created due to a masked load intrinsic.
Differential Revision: https://reviews.llvm.org/D28454
llvm-svn: 292005
We'll now expand AVX512_128_SET0 to an EVEX VXORD if VLX available. Or if its not, but register allocation has selected a non-extended register we will use VEX VXORPS. And if its an extended register without VLX we'll use a 512-bit XOR. Do the same for AVX512_FsFLD0SS/SD.
This makes it possible for the register allocator to have all 32 registers available to work with.
llvm-svn: 292004
This fallthrough if other cases are added between fabs and default
could cause fabs to fall to the next case resulting in a bug.
Better getting rid of it immediately just to be sure.
llvm-svn: 292003
Summary:
This patch attempts to fix the libc++ build/link so that it doesn't use an default C++ libraries on Windows. This is needed to prevent linking to MSVC's STL library.
Additionally this patch changes libc++ so that it is always linked with the non-debug DLL's (e.g. `/MD`). This is needed so that the test suite can correctly link the same libraries without needing to know which configuration `c++.dll` was linked with.
Reviewers: compnerd, rnk, majnemer, kimgr, awson, halyavin, smeenai
Subscribers: cfe-commits, mgorny
Differential Revision: https://reviews.llvm.org/D28441
llvm-svn: 292001
When -pedantic-errors is specified `__has_extension(<feature>)` is always
false when it would otherwise be true. This causes C++03 <atomic> to break
along with other issues.
This patch avoids the above problem by using __is_identifier(...) instead since
it is not affected by -pedantic-errors. For example instead of checking for
__has_extension(c_atomics) we now check `!__is_identifier(_Atomic)`, which
is only true when _Atomic is not a keyword provided by the compiler.
This patch applies similar changes to the detection logic for __decltype and
__nullptr as well.
Note that it does not apply this change to the C++03
`static_assert` macro since -Wc11-extensions warnings generated by expanding
that macro will appear in user code, and will not be suppressed as part of a
system header.
llvm-svn: 291995
This patch adjusts the out-of-tree CMake configuration so that
the stderr output is ignored when an old llvm-config is found
that doesn't support --cmakedir.
llvm-svn: 291993
This patch adjusts the out-of-tree CMake configuration so that
the stderr output is ignored when an old llvm-config is found
that doesn't support --cmakedir.
llvm-svn: 291992
This patch adjusts the out-of-tree CMake configuration so that
the stderr output is ignored when an old llvm-config is found
that doesn't support --cmakedir.
llvm-svn: 291991
extractProfTotalWeight checks if the profile type is sample profile, but
before that we have to ensure that summary is available. Also expanded
the unittest to test the case where there is no summar
Differential Revision: https://reviews.llvm.org/D28708
llvm-svn: 291982
When GlobalISel is configured to abort rather than fallback the only
thing that resetting the machine function does is make things harder
to debug. If we ever get to this point in the abort configuration it
indicates that we've already hit a bug, so this changes the behaviour
to abort instead.
llvm-svn: 291977
Allows LLVM to optimize sequences like the following:
%add = add nuw i32 %x, 1
%cmp = icmp ugt i32 %add, %y
Into:
%cmp = icmp uge i32 %x, %y
Previously, only signed comparisons were being handled.
Decrements could also be handled, but 'sub nuw %x, 1' is currently canonicalized to
'add %x, -1' in InstCombineAddSub, losing the nuw flag. Removing that canonicalization
seems like it might have far-reaching ramifications so I kept this simple for now.
Patch by Matti Niemenmaa!
Differential Revision: https://reviews.llvm.org/D24700
llvm-svn: 291975
Here, the optimization to not line wrap when it would not lead to a
reduction in columns was overwriting and enforced break that we want to
do no matter what.
Before:
int i = someFunction(
aaaaaaa,
0).aaa(aaaaaaaaaaaaa) *
aaaaaaa +
aaaaaaa;
After:
int i = someFunction(aaaaaaa, 0)
.aaa(aaaaaaaaaaaaa) *
aaaaaaa +
aaaaaaa;
llvm-svn: 291974
Correctly populating Machine PHIs relies on knowing exactly how the IR level
CFG was lowered to MachineIR. This needs to be tracked by any translation
phases that meddle (currently only SwitchInst handling).
llvm-svn: 291973
Summary:
This is a testcase where phi node cycling happens, and because we do
not order the leaders by domination or anything similar, the leader
keeps changing.
Using std::set for the members is too expensive, and we actually don't
need them sorted all the time, only at leader changes.
We could keep both a set and a vector, and keep them mostly sorted and
resort as necessary, or use a set and a fibheap, but all of this seems
premature.
After running some statistics, we are able to avoid the vast majority
of sorting by keeping a "next leader" field. Most congruence classes only have
leader changes once or twice during GVN.
Reviewers: davide
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28594
llvm-svn: 291968
Simon Pilgrim