File: arm64-vselect.ll

package info (click to toggle)
llvm-toolchain-13 1%3A13.0.1-11
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 1,418,840 kB
  • sloc: cpp: 5,290,826; ansic: 996,570; asm: 544,593; python: 188,212; objc: 72,027; lisp: 30,291; f90: 25,395; sh: 24,898; javascript: 9,780; pascal: 9,398; perl: 7,484; ml: 5,432; awk: 3,523; makefile: 2,913; xml: 953; cs: 573; fortran: 539
file content (25 lines) | stat: -rw-r--r-- 672 bytes parent folder | download | duplicates (15) 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
 
; RUN: llc < %s -mtriple=arm64-eabi -aarch64-neon-syntax=apple | FileCheck %s

;CHECK: @func63
;CHECK: cmeq.4h v0, v0, v1

;FIXME: currently, it will generate 3 instructions:
; ushll.4s	v0, v0, #0
; shl.4s	v0, v0, #31
; sshr.4s	v0, v0, #31
;But these instrucitons can be optimized into 1 instruction:
; sshll.4s  v0, v0, #0

;CHECK: bsl.16b v0, v2, v3
;CHECK: str  q0, [x0]
;CHECK: ret

%T0_63 = type <4 x i16>
%T1_63 = type <4 x i32>
%T2_63 = type <4 x i1>
define void @func63(%T1_63* %out, %T0_63 %v0, %T0_63 %v1, %T1_63 %v2, %T1_63 %v3) {
  %cond = icmp eq %T0_63 %v0, %v1
  %r = select %T2_63 %cond, %T1_63 %v2, %T1_63 %v3
  store %T1_63 %r, %T1_63* %out
  ret void
}