diff options
author | Simon Tatham <simon.tatham@arm.com> | 2019-12-11 12:07:16 +0000 |
---|---|---|
committer | Simon Tatham <simon.tatham@arm.com> | 2019-12-11 12:07:26 +0000 |
commit | d290424731ede31fd5fd75b929df8fe0adb547c7 (patch) | |
tree | e330c03a8673f9884918574a7213e4373ac89cea /clang/lib/CodeGen/CodeGenModule.cpp | |
parent | 9a5c849991fd4088b90c56cb402e6e9fa6e9141d (diff) | |
download | bcm5719-llvm-d290424731ede31fd5fd75b929df8fe0adb547c7.tar.gz bcm5719-llvm-d290424731ede31fd5fd75b929df8fe0adb547c7.zip |
[ARM][MVE] Factor out an IntrinsicMX multiclass.
Summary:
The ACLE intrinsics for MVE contain a lot of pairs of functions with
`_m` and `_x` in the name, wrapping a predicated MVE instruction which
only partially overwrites its output register. They have the common
pattern that the `_m` variant takes an initial argument called
'inactive', of the same type as the return value, supplying the input
value of the output register, so that lanes disabled by the
predication will be taken from that parameter; the `_x` variant omits
that initial argument, and simply sets it to undef.
That common pattern is simple enough to wrap into a multiclass, which
should save a lot of effort in setting up all the rest of the `_x`
variants. In this commit I introduce `multiclass IntrinsicMX` in
`arm_mve_defs.td`, and convert existing generation of m/x pairs to use
it.
This allows me to remove the `PredicatedImmediateVectorShift`
multiclass (from D71065) completely, because the new multiclass makes
it so much simpler that it's not worth bothering to define it at all.
Reviewers: MarkMurrayARM, miyuki
Reviewed By: MarkMurrayARM, miyuki
Subscribers: kristof.beyls, dmgreen, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D71335
Diffstat (limited to 'clang/lib/CodeGen/CodeGenModule.cpp')
0 files changed, 0 insertions, 0 deletions