Add a pseudo instruction REG_SEQUENCE that takes a list of registers and
authorEvan Cheng <evan.cheng@apple.com>
Sat, 1 May 2010 00:28:44 +0000 (00:28 +0000)
committerEvan Cheng <evan.cheng@apple.com>
Sat, 1 May 2010 00:28:44 +0000 (00:28 +0000)
commitb55c8bed9d0f3eaa454a657746d8ec11aae9dea3
treea711f3e3c0ce07662b161f653e4e8b4151d5b1b4
parent95140a4cc168bb75a54a434e3e53f9060742dc41
Add a pseudo instruction REG_SEQUENCE that takes a list of registers and
sub-register indices and outputs a single super register which is formed from
a consecutive sequence of registers.

This is used as register allocation / coalescing aid and it is useful to
represent instructions that output register pairs / quads. For example,
v1024, v1025 = vload <address>
where v1024 and v1025 forms a register pair.

This really should be modelled as
v1024<3>, v1025<4> = vload <address>
but it would violate SSA property before register allocation is done.

Currently we use insert_subreg to form the super register:
v1026 = implicit_def
v1027 - insert_subreg v1026, v1024, 3
v1028 = insert_subreg v1027, v1025, 4
...
      = use v1024
      = use v1028

But this adds pseudo live interval overlap between v1024 and v1025.

We can now modeled it as
v1024, v1025 = vload <address>
v1026 = REG_SEQUENCE v1024, 3, v1025, 4
...
      = use v1024
      = use v1026

After coalescing, it will be
v1026<3>, v1025<4> = vload <address>
...
      = use v1026<3>
      = use v1026

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102815 91177308-0d34-0410-b5e6-96231b3b80d8
include/llvm/Target/Target.td
include/llvm/Target/TargetOpcodes.h
utils/TableGen/CodeEmitterGen.cpp
utils/TableGen/CodeGenTarget.cpp