| Tom Stellard | 2b971eb | 2013-05-10 02:09:45 +0000 | [diff] [blame] | 1 | ; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s |
| 2 | |
| Tom Stellard | 79243d9 | 2014-10-01 17:15:17 +0000 | [diff] [blame] | 3 | ; CHECK: {{^}}bfe_def: |
| Tom Stellard | 2b971eb | 2013-05-10 02:09:45 +0000 | [diff] [blame] | 4 | ; CHECK: BFE_UINT |
| Matt Arsenault | 3dbeefa | 2017-03-21 21:39:51 +0000 | [diff] [blame] | 5 | define amdgpu_kernel void @bfe_def(i32 addrspace(1)* %out, i32 %x) { |
| Tom Stellard | 2b971eb | 2013-05-10 02:09:45 +0000 | [diff] [blame] | 6 | entry: |
| 7 | %0 = lshr i32 %x, 5 |
| 8 | %1 = and i32 %0, 15 ; 0xf |
| 9 | store i32 %1, i32 addrspace(1)* %out |
| 10 | ret void |
| 11 | } |
| 12 | |
| 13 | ; This program could be implemented using a BFE_UINT instruction, however |
| 14 | ; since the lshr constant + number of bits in the mask is >= 32, it can also be |
| 15 | ; implmented with a LSHR instruction, which is better, because LSHR has less |
| 16 | ; operands and requires less constants. |
| 17 | |
| Tom Stellard | 79243d9 | 2014-10-01 17:15:17 +0000 | [diff] [blame] | 18 | ; CHECK: {{^}}bfe_shift: |
| Tom Stellard | 2b971eb | 2013-05-10 02:09:45 +0000 | [diff] [blame] | 19 | ; CHECK-NOT: BFE_UINT |
| Matt Arsenault | 3dbeefa | 2017-03-21 21:39:51 +0000 | [diff] [blame] | 20 | define amdgpu_kernel void @bfe_shift(i32 addrspace(1)* %out, i32 %x) { |
| Tom Stellard | 2b971eb | 2013-05-10 02:09:45 +0000 | [diff] [blame] | 21 | entry: |
| 22 | %0 = lshr i32 %x, 16 |
| 23 | %1 = and i32 %0, 65535 ; 0xffff |
| 24 | store i32 %1, i32 addrspace(1)* %out |
| 25 | ret void |
| 26 | } |