dag-combine optimization to implement the ext-load efficiently (using shuffles).
For example the type <4 x i8> is stored in memory as i32, but it needs to
find its way into a <4 x i32> register. Previously we scalarized the memory
access, now we use shuffles.
llvm-svn: 139995
maxps and maxpd). This broke the sse41-blend.ll testcase by causing
maxpd to be produced rather than a cmp+blend pair, which is the reason
I tweaked it. Gives a small speedup on doduc with dragonegg when the
GCC vectorizer is used.
llvm-svn: 139986
ModuleSP
Module::GetSP();
Since we are now using intrusive ref counts, we can easily turn any
pointer to a module into a shared pointer just by assigning it.
llvm-svn: 139984
We had some cases where getting the shared pointer for a module from
the global module list was causing a performance issue when debugging
with DWARF in .o files. Now that the module uses intrusive ref counts,
we can easily convert any pointer to a shared pointer.
llvm-svn: 139983
the AST reader), merge that header file information with whatever
header file information we already have. Otherwise, we might forget
something we already knew (e.g., that the header was #import'd already).
llvm-svn: 139979
Microsoft specific tweaking will now fall into 2 categories:
- fms-extension: Microsoft specific extensions that should never change the meaning of an otherwise well formed code. Currently map to LangOptions::Microsoft. (To be clearer, I am planning to change the name to LangOptions::MicrosoftExt).
- fms-compatibility: Really a MSVC emulation mode. Map to LangOptions::MicrosoftMode. Can change the meaning of an otherwise standard conformant program.
llvm-svn: 139978
data sent back to the debugger. On the debugger side, use the opportunity during the
StopInfoMachException::CreateStopReasonWithMachException() method to set the hardware index
for the very watchpoint location.
llvm-svn: 139975
arbitrary amount of code. This forces us to stage the AST writer more
strictly, ensuring that we don't assign a declaration ID to a
declaration until after we're certain that no more modules will get
loaded.
llvm-svn: 139974
- Speed of "merge()", which merged data flow facts. This was doing a set canonicalization on every insertion, which was super slow.
To fix this, we use ImmutableSetRef.
- Visit CFGBlocks in reverse postorder. This is a huge speedup, as on some test cases the algorithm would take many iterations
to converge.
This contains a bunch of copy-paste from UninitializedValues.cpp and ThreadSafety.cpp. The idea
was to get something working first, and then refactor the common logic for all three files into
a separate analysis/library entry point.
llvm-svn: 139968
are declared with load patterns. This fix the crash in PR10941. No testcases,
since a fold is triggered and then converted back to the register form
afterwards.
llvm-svn: 139953
Bill Wendling