Matt Arsenault | 09b2c4a | 2016-07-15 21:26:52 +0000 | [diff] [blame] | 1 | ; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG %s |
| 2 | |
| 3 | declare float @llvm.r600.recipsqrt.ieee.f32(float) nounwind readnone |
| 4 | |
| 5 | ; EG-LABEL: {{^}}recipsqrt.ieee_f32: |
| 6 | ; EG: RECIPSQRT_IEEE |
Matt Arsenault | 3dbeefa | 2017-03-21 21:39:51 +0000 | [diff] [blame] | 7 | define amdgpu_kernel void @recipsqrt.ieee_f32(float addrspace(1)* %out, float %src) nounwind { |
Matt Arsenault | 09b2c4a | 2016-07-15 21:26:52 +0000 | [diff] [blame] | 8 | %recipsqrt.ieee = call float @llvm.r600.recipsqrt.ieee.f32(float %src) nounwind readnone |
| 9 | store float %recipsqrt.ieee, float addrspace(1)* %out, align 4 |
| 10 | ret void |
| 11 | } |
| 12 | |
| 13 | ; TODO: Really these should be constant folded |
| 14 | ; EG-LABEL: {{^}}recipsqrt.ieee_f32_constant_4.0 |
| 15 | ; EG: RECIPSQRT_IEEE |
Matt Arsenault | 3dbeefa | 2017-03-21 21:39:51 +0000 | [diff] [blame] | 16 | define amdgpu_kernel void @recipsqrt.ieee_f32_constant_4.0(float addrspace(1)* %out) nounwind { |
Matt Arsenault | 09b2c4a | 2016-07-15 21:26:52 +0000 | [diff] [blame] | 17 | %recipsqrt.ieee = call float @llvm.r600.recipsqrt.ieee.f32(float 4.0) nounwind readnone |
| 18 | store float %recipsqrt.ieee, float addrspace(1)* %out, align 4 |
| 19 | ret void |
| 20 | } |
| 21 | |
| 22 | ; EG-LABEL: {{^}}recipsqrt.ieee_f32_constant_100.0 |
| 23 | ; EG: RECIPSQRT_IEEE |
Matt Arsenault | 3dbeefa | 2017-03-21 21:39:51 +0000 | [diff] [blame] | 24 | define amdgpu_kernel void @recipsqrt.ieee_f32_constant_100.0(float addrspace(1)* %out) nounwind { |
Matt Arsenault | 09b2c4a | 2016-07-15 21:26:52 +0000 | [diff] [blame] | 25 | %recipsqrt.ieee = call float @llvm.r600.recipsqrt.ieee.f32(float 100.0) nounwind readnone |
| 26 | store float %recipsqrt.ieee, float addrspace(1)* %out, align 4 |
| 27 | ret void |
| 28 | } |