[GlobalISel][AArch64] Combine unmerge(G_EXT v, undef) to unmerge(v). #65263
Conversation
dzhidzhoev
requested review from
aemerson,
arsenm,
TNorthover,
RKSimon,
ornata,
dsandersllvm,
davemgreen and
a team
dzhidzhoev
force-pushed
the
unmerge-ext
branch
3 times, most recently
from
2fe29e8
to
8e44ad1
Compare
|
Should be a generic combine. Can you generalize to N unused defs? |
G_EXT is an opcode for AArch64's |
| if (!DstTy.isVector()) | ||
| return false; | ||
|
|
||
| MachineInstr *Ext = getDefIgnoringCopies( |
There was a problem hiding this comment.
getOpcodeDef() with the GUnmerge source reg helper here?
| if (!LowestVal || LowestVal->Value.getZExtValue() != DstTy.getSizeInBytes()) | ||
| return false; | ||
|
|
||
| MachineInstr *Undef = getDefIgnoringCopies(ExtSrc2, MRI); |
G_EXT should be G_AARCH64_EXT then. I assumed this was shorthand for all 3 extends |
dzhidzhoev
force-pushed
the
unmerge-ext
branch
from
8e44ad1
to
b8e671f
Compare
| return false; | ||
| if (!MRI.use_nodbg_empty(MI.getOperand(1).getReg())) | ||
| if (!MRI.use_nodbg_empty(Unmerge.getOperand(1).getReg())) |
There was a problem hiding this comment.
GenericMachineInstr also gives you getReg(N)
| @@ -1071,18 +1071,18 @@ void applyVectorSextInReg(MachineInstr &MI, MachineRegisterInfo &MRI, | |||
| bool matchUnmergeExtToUnmerge(MachineInstr &MI, MachineRegisterInfo &MRI, | |||
| Register &MatchInfo) { | |||
| assert(MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES); | |||
There was a problem hiding this comment.
This assert becomes unnecessary once you have the cast<GUnmerge> below.
When having <N x t> d1, unused = unmerge(G_EXT <2*N x t> v1, undef, N), it is possible to express it just as unused, d1 = unmerge v1. It is useful for tackling regressions in arm64-vcvt_f.ll, introduced in https://reviews.llvm.org/D144670.
dzhidzhoev
force-pushed
the
unmerge-ext
branch
from
a3cd7f7
to
bb1a03d
Compare
When having
<N x t> d1, unused = unmerge(G_EXT <2*N x t> v1, undef, N), it is possible to express it just asunused, d1 = unmerge v1.It is useful for tackling regressions in arm64-vcvt_f.ll, introduced in https://reviews.llvm.org/D144670.