We want to replace as much custom x86 shuffling via intrinsics
as possible because pushing the code down the generic shuffle
optimization path allows for better codegen and less complexity
in LLVM.
This is the sibling patch for the LLVM half of this change:
http://reviews.llvm.org/D8086
Differential Revision: http://reviews.llvm.org/D8088
llvm-svn: 231792
This is a recommit of r231150, reverted in r231409. Turns out
that -fsanitize=shift-base check implementation only works if the
shift exponent is valid, otherwise it contains undefined behavior
itself.
Make sure we check that exponent is valid before we proceed to
check the base. Make sure that we actually report invalid values
of base or exponent if -fsanitize=shift-base or
-fsanitize=shift-exponent is specified, respectively.
llvm-svn: 231711
When passing a type with large alignment byval, we were specifying the type's
alignment rather than the alignment that the backend is actually capable of
producing (ABIAlign).
This would be OK (if odd) assuming the backend dealt with it prooperly,
unfortunately it doesn't and trying to pass types with "byval align 16" can
cause it to set fp incorrectly and trash the stack during the prologue. I'll be
fixing that in a separate patch, but Clang should still be emitting IR that's
as close to its intent as possible.
rdar://20059039
llvm-svn: 231706
It's not that easy. If we're only checking -fsanitize=shift-base we
still need to verify that exponent has sane value, otherwise
UBSan-inserted checks for base will contain undefined behavior
themselves.
llvm-svn: 231409
Opt in Win64 to supporting sjlj lowering. We have the backend lowering,
so I think this was just an oversight because WinX86_64TargetCodeGenInfo
doesn't inherit from X86_64TargetCodeGenInfo.
llvm-svn: 231280
This test doesn't provide any value (it just checks that the frontend
produces exactly one compile unit), and it certainly isn't doing what
the comment says. Noticed via IRC review of my update to it in r231083.
llvm-svn: 231152
-fsanitize=shift is now a group that includes both these checks, so
exisiting users should not be affected.
This change introduces two new UBSan kinds that sanitize only left-hand
side and right-hand side of shift operation. In practice, invalid
exponent value (negative or too large) tends to cause more portability
problems, including inconsistencies between different compilers, crashes
and inadequeate results on non-x86 architectures etc. That is,
-fsanitize=shift-exponent failures should generally be addressed first.
As a bonus, this change simplifies CodeGen implementation for emitting left
shift (separate checks for base and exponent are now merged by the
existing generic logic in EmitCheck()), and LLVM IR for these checks
(the number of basic blocks is reduced).
llvm-svn: 231150
Originally we were using the same GCC builtins to lower this AVX2 vector
intrinsic. Instead we will now lower it directly to a vector shuffle.
This will not only allow LLVM to generate better code, but it will also allow us
to remove the GCC intrinsics.
Reviewed by Andrea
This is related to rdar://problem/18742778.
llvm-svn: 231081
For global reg lvalue - use regular store through global register.
For simple lvalue - use simple atomic store.
For bitfields, vector element, extended vector elements - the original value of the whole storage (for vector elements) or of some aligned value (for bitfields) is atomically read, the part of this value for the given lvalue is modified and then use atomic compare-and-exchange operation to try to atomically write modified value (if it was not modified).
Also, changes in this patch fix the bug for '#pragma omp atomic read' applied to extended vector elements.
Differential Revision: http://reviews.llvm.org/D7369
llvm-svn: 230736
The __finally emission block tries to be clever by removing unused continuation
edges if there's an unconditional jump out of the __finally block. With
exception edges, the EH continuation edge isn't always unused though and we'd
crash in a few places.
Just don't be clever. That makes the IR for __finally blocks a bit longer in
some cases (hence small and behavior-preserving changes to existing tests), but
it makes no difference in general and it fixes the last crash from PR22553.
http://reviews.llvm.org/D7918
llvm-svn: 230697
Currently, the NaN values emitted for MIPS architectures do not cover
non-IEEE754-2008 compliant case. This change fixes the issue.
Patch by Vladimir Radosavljevic.
Differential Revision: http://reviews.llvm.org/D7882
llvm-svn: 230653
Original CL description:
Produce less broken basic block sequences for __finally blocks.
The way cleanups (such as PerformSEHFinally) get emitted is that codegen
generates some initialization code, then calls the cleanup's Emit() with the
insertion point set to a good place, then the cleanup is supposed to emit its
stuff, and then codegen might tack in a jump or similar to where the insertion
point is after the cleanup.
The PerformSEHFinally cleanup tries to just stash away the block it's supposed
to codegen into, and then does codegen later, into that stashed block. However,
after codegen'ing the __finally block, it used to set the insertion point to
the finally's continuation block (where the __finally cleanup goes when its body
is completed after regular, non-exceptional control flow). That's not correct,
as that block can (and generally does) already ends in a jump. Instead,
remember the insertion point that was current before the __finally got emitted,
and restore that.
Fixes two of the crashes in PR22553.
llvm-svn: 230503
The way cleanups (such as PerformSEHFinally) get emitted is that codegen
generates some initialization code, then calls the cleanup's Emit() with the
insertion point set to a good place, then the cleanup is supposed to emit its
stuff, and then codegen might tack in a jump or similar to where the insertion
point is after the cleanup.
The PerformSEHFinally cleanup tries to just stash away the block it's supposed
to codegen into, and then does codegen later, into that stashed block. However,
after codegen'ing the __finally block, it used to set the insertion point to
the finally's continuation block (where the __finally cleanup goes when its body
is completed after regular, non-exceptional control flow). That's not correct,
as that block can (and generally does) already ends in a jump. Instead,
remember the insertion point that was current before the __finally got emitted,
and restore that.
Fixes two of the crashes in PR22553.
llvm-svn: 230460
This is a necessary prerequisite for debugging with modules.
The .pcm files become containers that hold the serialized AST which allows
us to store debug information in the module file that can be shared by all
object files that were built importing the module.
This reapplies r230044 with a fixed configure+make build and updated
dependencies and testcase requirements. Over the last iteration this
version adds
- missing target requirements for testcases that specify an x86 triple,
- a missing clangCodeGen.a dependency to libClang.a in the make build.
rdar://problem/19104245
llvm-svn: 230423
The backend should now be able to handle all AAPCS rules based on argument
type, which means Clang no longer has to duplicate the register-counting logic
and the CodeGen can be significantly simplified.
llvm-svn: 230349
MSVC does not support C99 _Complex.
ICC, however, does support it on windows x86_64, and treats it, for purposes of parameter passing, as equivalent to a struct containing two fields (for the real and imaginary part).
Differential Revision: http://reviews.llvm.org/D7825
llvm-svn: 230315
llvm.eh.sjlj.setjmp / llvm.eh.sjlj.longjmp, if the backend is known to
support them outside the Exception Handling context. The default
handling in LLVM codegen doesn't work and will create incorrect code.
The ARM backend on the other hand will assert if the intrinsics are
used.
llvm-svn: 230255
For now -funique-section-names is the default, so no change in default behavior.
The total .o size in a build of llvm and clang goes from 241687775 to 230649031
bytes if -fno-unique-section-names is used.
llvm-svn: 230031
Summary:
The definition for _mm256_insert_epi64 was taking an int, which would get
truncated before being inserted in the vector.
Original patch by Joshua Magee!
Reviewers: bruno, craig.topper
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D7179
llvm-svn: 229811
Not all targets generate 'store atomic' instructions for
'_Atomic(_Complex int)'. Some targets use the __atomic_store builtin instead.
This commit makes the test accept either one.
llvm-svn: 229676
This is a patch for PR22563 ( http://llvm.org/bugs/show_bug.cgi?id=22563 ).
We were not correctly unwrapping a single 256-bit AVX vector that was defined as an array of 1 inside a struct.
We would generate a <4 x float> param/return value instead of <8 x float> and lose half of the vector.
Differential Revision: http://reviews.llvm.org/D7614
llvm-svn: 229408
For #pragma comment(linker, ...) MSVC expects the comment string to be quoted, but for #pragma comment(lib, ...) the compiler itself quotes the library name.
Since this distinction disappears by the time the directive reaches the backend, move quoting for the "lib" version to the frontend.
Differential Revision: http://reviews.llvm.org/D7653
llvm-svn: 229376
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 first change won't touch GEPOperators such as these, but the update
script only identifies them by the leading '(' after getelementptr or
'getelementptr inbounds', so update this test to at least have those
features to allow auto-migrating.
llvm-svn: 229198
Sanjay Patel