v1024 = REG_SEQUENCE ...
v1025 = EXTRACT_SUBREG v1024, 5
v1026 = EXTRACR_SUBREG v1024, 6
= VSTxx <addr>, v1025, v1026
The REG_SEQUENCE ensures the sources that feed into the VST instruction
are getting the right register allocation so they form a large super-
register. The extract_subreg will be coalesced away all would just work:
v1024 = REG_SEQUENCE ...
= VSTxx <addr>, v1024:5, v1024:6
The problem is if the coalescer isn't run, the extract_subreg instructions
would stick around and there is no assurance v1025 and v1026 will get the
right registers.
As a short term workaround, teach the NEON pre-allocation pass to transfer
the sub-register indices over. An alternative would be do it 2addr pass
when reg_sequence's are eliminated. But that *seems* wrong and require
updating liveness information.
Another alternative is to do this in the scheduler when the instructions are
created. But that would mean somehow the scheduler this has to be done for
correctness reason. That's yucky as well. So for now, we are leaving this
in the target specific pass.
llvm-svn: 103540
be diced into atoms, and adjust getAtom() to take this into account.
- This fixes relocations to symbols in fixed size literal sections, for
example.
llvm-svn: 103532
and the others use the regular addPassesToEmitFile hook now, and
llc no longer needs a bunch of redundant code to handle the
whole-file case.
llvm-svn: 103492
Sorry for the big change. The path leading up to this patch had some TableGen
changes that I didn't want to commit before I knew they were useful. They
weren't, and this version does not need them.
The fast register allocator now does no liveness calculations. Instead it relies
on kill flags provided by isel. (Currently those kill flags are also ignored due
to isel bugs). The allocation algorithm is supposed to work with any subset of
valid kill flags. More kill flags simply means fewer spills inserted.
Registers are allocated from a working set that contains no aliases. That means
most allocations can be done directly without expensive alias checks. When the
working set runs out of registers we do the full alias check to find new free
registers.
llvm-svn: 103488
Move EmitTargetCodeForMemcpy, EmitTargetCodeForMemset, and
EmitTargetCodeForMemmove out of TargetLowering and into
SelectionDAGInfo to exercise this.
llvm-svn: 103481
- This eliminates getAtomForAddress() (which was a linear search) and
simplifies getAtom().
- This also fixes some correctness problems where local labels at the same
address as non-local labels could be assigned to the wrong atom.
llvm-svn: 103480
string of features for that target. However LTO was using that string to pass
into the "create target machine" stuff. That stuff needed the feature string to
be in a particular form. In particular, it needed the CPU specified first and
then the attributes. If there isn't a CPU specified, it required it to be blank
-- e.g., ",+altivec". Yuck.
Modify the getDefaultSubtargetFeatures method to be a non-static member
function. For all attributes for a specific subtarget, it will add them in like
normal. It will also take a CPU string so that it can satisfy this horrible
syntax.
llvm-svn: 103451
