maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/AArch64/machine-cp-clobbers.mir

52 lines
1.2 KiB
Plaintext
Raw Normal View History

Reapply "[MachineCopyPropagation] Reimplement CopyTracker in terms of register units" Recommits r342942, which was reverted in r343189, with a fix for an issue where we would propagate unsafely if we defined only the upper part of a register. Original message: Change the copy tracker to keep a single map of register units instead of 3 maps of registers. This gives a very significant compile time performance improvement to the pass. I measured a 30-40% decrease in time spent in MCP on x86 and AArch64 and much more significant improvements on out of tree targets with more registers. llvm-svn: 344942
2018-10-23 03:51:31 +08:00
# RUN: llc -march=aarch64 -o - %s -run-pass=machine-cp | FileCheck %s
---
name: dont_propagate_past_lower_subreg_kill
tracksRegLiveness: true
body: |
bb.0:
; CHECK-LABEL: name: dont_propagate_past_lower_subreg_kill
; CHECK: HINT 0, implicit-def $q0
; CHECK: HINT 0, implicit-def $d1
; CHECK: HINT 0, implicit killed $d1
; CHECK: $q1 = COPY killed $q0
; CHECK: $q2 = COPY $q1
; CHECK: HINT 0, implicit $q2
HINT 0, implicit-def $q0
$q1 = COPY killed $q0
$q0 = COPY killed $q1
HINT 0, implicit-def $d1
HINT 0, implicit killed $d1
$q1 = COPY killed $q0
$q2 = COPY $q1
HINT 0, implicit $q2
...
---
name: dont_propagate_past_upper_subreg_kill
tracksRegLiveness: true
body: |
bb.0:
; CHECK-LABEL: name: dont_propagate_past_upper_subreg_kill
; CHECK: HINT 0, implicit-def $z0
; CHECK: HINT 0, implicit-def $z1_hi
; CHECK: HINT 0, implicit killed $z1_hi
; CHECK: $z1 = COPY killed $z0
; CHECK: $z2 = COPY $z1
; CHECK: HINT 0, implicit $z2
HINT 0, implicit-def $z0
$z1 = COPY killed $z0
$z0 = COPY killed $z1
HINT 0, implicit-def $z1_hi
HINT 0, implicit killed $z1_hi
$z1 = COPY killed $z0
$z2 = COPY $z1
HINT 0, implicit $z2
...