Match reductions starting at binary operation feeding into a phi. The code
handles trees like
r += v1 + v2 + v3 ...
and
r += v1
r += v2
...
and
r *= v1 + v2 + ...
We currently only handle associative operations (add, fadd fast).
The code can now also handle reductions feeding into stores.
a[i] = v1 + v2 + v3 + ...
The code is currently disabled behind the flag "-slp-vectorize-hor". The cost
model for most architectures is not there yet.
I found one opportunity of a horizontal reduction feeding a phi in TSVC
(LoopRerolling-flt) and there are several opportunities where reductions feed
into stores.
radar://14607682
llvm-svn: 191108
is no need to go through the driver indirection here, and it clutters
things up as dependencies can sneak in for specific things the driver is
doing.
llvm-svn: 191107
Specifically, allows the unwinder to handle the case where sc.function
gets resolved with a pc that is one past the address range of the function
(consistent with a tail call). However, there is no matching symbol.
Adds eSymbolContextTailCall to provide callers with control over the scope
of symbol resolution and to allow ResolveSymbolContextForAddress to handle
tail calls since this routine is common to unwind and disassembly.
llvm-svn: 191102
This solves the problem of fallback onto ourselves if clang-cl
has been renamed to cl.exe and put on the PATH, as happens with
the VS integration.
Differential Revision: http://llvm-reviews.chandlerc.com/D1731
llvm-svn: 191099
Apart from being more compact and already implemented, this also handles the
case where the parent is null. (It does also ignore all casts, not just
implicit ones, but this is more efficient to test and in the case we care
about---a message in a PseudoObjectExpr---there should only be implicit casts
anyway.
This should fix our internal buildbot.
llvm-svn: 191094
The GEP pattern is what SCEV expander emits for "ugly geps". The latter is what
you get for pointer subtraction in C code. The rest of instcombine already
knows how to deal with that so just canonicalize on that.
llvm-svn: 191090
Due to bugs in the interposition of resolver functions on 10.7 and 10.8 both memcpy() and memmove()
were previously intercepted by INTERCEPTOR(memcpy), which led to false positives and crashes (see http://llvm.org/bugs/show_bug.cgi?id=16362)
We choose to use a memmove-like function in both cases. This effectively disables the overlap checks in memcpy(), but the overlaps
aren't possible in practice, because memcpy() and memmove() are the same function on Darwin.
Once the interposition is fixed in 10.9, we'll need to revisit this issue, see https://code.google.com/p/address-sanitizer/issues/detail?id=226
llvm-svn: 191081
Instead add the ASan runtime to the linker command line so that only the ASan API functions can be undefined in the target library.
Fixes http://llvm.org/bugs/show_bug.cgi?id=17275
llvm-svn: 191076
The global registry is used to allow command line override of the
scheduler selection, but does not work well as the normal selection
API. For example, the same LLVM process should be able to target
multiple targets or subtargets.
llvm-svn: 191071
GNU LD driver only understood no_such_file_or_directory error and was showing
just "Unknown Error" for any other type of error. With this patch, the driver
now prints file name and error message string by default.
llvm-svn: 191070
fields in the class. This allows a better checking of member intiailizers and
in class initializers in regards to initialization ordering.
For instance, this code will now produce warnings:
class A {
int x;
int y;
A() : x(y) {} // y is initialized after x, warn here
A(int): y(x) {} // default initialization of leaves x uninitialized, warn here
};
Several test cases were updated with -Wno-uninitialized to silence this warning.
llvm-svn: 191068
variable from being the function to being the enclosing namespace scope (in
C++) or the TU (in C). This allows us to fix a selection of related issues
where we would build incorrect redeclaration chains for such declarations, and
fail to notice type mismatches.
Such declarations are put into a new IdentifierNamespace, IDNS_LocalExtern,
which is only found when searching scopes, and not found when searching
DeclContexts. Such a declaration is only made visible in its DeclContext if
there are no non-LocalExtern declarations.
llvm-svn: 191064
/incremental is an option to enable incremental linking. We will eventually
want to implement the feature for better performance, but in the meantime,
we want to just ignore the option so that the linker does not output unknown
option error when it sees /incremental option.
llvm-svn: 191063