This actually exposed an infinite recursion bug in ComputeValueKnownInPredecessors which theoretically already existed (in JumpThreading's
handling of and/or of i1's), but never manifested before. This patch adds a tracking set to prevent this case.
llvm-svn: 112589
getMagicNumber was treating the _binary_ data it read in as a
null terminated string. This resulted in the std::string
calculating the length, and causing an assert in other code that
assumed that the length it passed was the same as the length of
the string it would get back.
llvm-svn: 112586
declaration send or a variadic function call, collapse the ", ..."
into the parameter before it, so that we don't get a second
placeholder.
llvm-svn: 112579
Eventually, we want to disable physreg coalescing completely, and let the
register allocator do its job using hints.
This option makes it possible to measure the impact of disabling physreg
coalescing.
llvm-svn: 112567
instantiating the parameters. In a perfect world, this wouldn't
matter, and compilers are free to instantiate in any order they
want. However, every other compiler seems to instantiate the return
type first, and some code (in this case, Boost.Polygon) depends on
this and SFINAE to avoid instantiating something that shouldn't be
instantiated.
We could fight this battle, and insist that Clang is allowed to do
what it does, but it's not beneficial: it's more predictable to
instantiate this way, in source order. When we implement
late-specified return types, we'll need to instantiate the return type
last when it was late-specified, hence the FIXME.
We now compile Boost.Polygon properly.
llvm-svn: 112561
of that parameter, reduce the level by the number of active template
argument lists rather than by 1. The number of active template
argument lists is only > 1 when we have a class template partial
specialization of a member template of a class template that itself is
a member template of another class template.
... and Boost.MSM does this. Fixes PR7669.
llvm-svn: 112551
persistent variables were staying around too long.
This caused the following problem:
- A persistent result variable is created for the
result of an expression. The pointer to the
corresponding Decl is stored in the variable.
- The persistent variable is looked up during
struct generation (correctly) using its Decl.
- Another expression defines a new result variable
which happens to have a Decl in the same place
as the original result variable.
- The persistent variable is looked up during
struct generation using its Decl, but the old
result variable appears first in the list and
has the same Decl pointer.
The fix is to destroy parser-specific data when
it is no longer valid.
Also improved some logging as I diagnosed the
bug.
llvm-svn: 112540
Bill Wendling