File: bfe_uint.ll

package info (click to toggle)
llvm-toolchain-11 1%3A11.0.1-2
  • links: PTS, VCS
  • area: main
  • in suites: bullseye
  • size: 995,808 kB
  • sloc: cpp: 4,767,656; ansic: 760,916; asm: 477,436; python: 170,940; objc: 69,804; lisp: 29,914; sh: 23,855; f90: 18,173; pascal: 7,551; perl: 7,471; ml: 5,603; awk: 3,489; makefile: 2,573; xml: 915; cs: 573; fortran: 503; javascript: 452
file content (26 lines) | stat: -rw-r--r-- 784 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
26
 
; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s

; CHECK: {{^}}bfe_def:
; CHECK: BFE_UINT
define amdgpu_kernel void @bfe_def(i32 addrspace(1)* %out, i32 %x) {
entry:
  %0 = lshr i32 %x, 5
  %1 = and i32 %0, 15 ; 0xf
  store i32 %1, i32 addrspace(1)* %out
  ret void
}

; This program could be implemented using a BFE_UINT instruction, however
; since the lshr constant + number of bits in the mask is >= 32, it can also be
; implmented with a LSHR instruction, which is better, because LSHR has less
; operands and requires less constants.

; CHECK: {{^}}bfe_shift:
; CHECK-NOT: BFE_UINT
define amdgpu_kernel void @bfe_shift(i32 addrspace(1)* %out, i32 %x) {
entry:
  %0 = lshr i32 %x, 16
  %1 = and i32 %0, 65535 ; 0xffff
  store i32 %1, i32 addrspace(1)* %out
  ret void
}