xref: /llvm-project/llvm/test/CodeGen/AMDGPU/fp-min-max-num-global-atomics.ll (revision 3277c7cd28154e33637a168acb26cea7ac1f7fff)
107a6d736SMirko Brkušanin; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
29e9907f1SFangrui Song; RUN: llc < %s -global-isel=0 -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs | FileCheck %s -check-prefixes=GFX12,GFX12-SDAG
39e9907f1SFangrui Song; RUN: llc < %s -global-isel=1 -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs | FileCheck %s -check-prefixes=GFX12,GFX12-GISEL
407a6d736SMirko Brkušanin
507a6d736SMirko Brkušanindeclare float @llvm.amdgcn.global.atomic.fmin.num.f32.p1.f32(ptr addrspace(1) %ptr, float %data)
607a6d736SMirko Brkušanindeclare float @llvm.amdgcn.global.atomic.fmax.num.f32.p1.f32(ptr addrspace(1) %ptr, float %data)
707a6d736SMirko Brkušanin
807a6d736SMirko Brkušanindefine amdgpu_cs void @global_atomic_fmin_num_f32_noret(ptr addrspace(1) %ptr, float %data) {
907a6d736SMirko Brkušanin; GFX12-LABEL: global_atomic_fmin_num_f32_noret:
1007a6d736SMirko Brkušanin; GFX12:       ; %bb.0:
1107a6d736SMirko Brkušanin; GFX12-NEXT:    global_atomic_min_num_f32 v[0:1], v2, off
1207a6d736SMirko Brkušanin; GFX12-NEXT:    s_endpgm
1307a6d736SMirko Brkušanin  %ret = call float @llvm.amdgcn.global.atomic.fmin.num.f32.p1.f32(ptr addrspace(1) %ptr, float %data)
1407a6d736SMirko Brkušanin  ret void
1507a6d736SMirko Brkušanin}
1607a6d736SMirko Brkušanin
1707a6d736SMirko Brkušanindefine amdgpu_cs void @global_atomic_fmax_num_f32_noret(ptr addrspace(1) %ptr, float %data) {
1807a6d736SMirko Brkušanin; GFX12-LABEL: global_atomic_fmax_num_f32_noret:
1907a6d736SMirko Brkušanin; GFX12:       ; %bb.0:
2007a6d736SMirko Brkušanin; GFX12-NEXT:    global_atomic_max_num_f32 v[0:1], v2, off
2107a6d736SMirko Brkušanin; GFX12-NEXT:    s_endpgm
2207a6d736SMirko Brkušanin  %ret = call float @llvm.amdgcn.global.atomic.fmax.num.f32.p1.f32(ptr addrspace(1) %ptr, float %data)
2307a6d736SMirko Brkušanin  ret void
2407a6d736SMirko Brkušanin}
2507a6d736SMirko Brkušanin
2607a6d736SMirko Brkušanindefine amdgpu_cs void @global_atomic_fmax_num_f32_rtn(ptr addrspace(1) %ptr, float %data, ptr addrspace(1) %out) {
2707a6d736SMirko Brkušanin; GFX12-LABEL: global_atomic_fmax_num_f32_rtn:
2807a6d736SMirko Brkušanin; GFX12:       ; %bb.0:
2907a6d736SMirko Brkušanin; GFX12-NEXT:    global_atomic_max_num_f32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN
30*ba52f06fSJay Foad; GFX12-NEXT:    s_wait_loadcnt 0x0
3107a6d736SMirko Brkušanin; GFX12-NEXT:    global_store_b32 v[3:4], v0, off
3207a6d736SMirko Brkušanin; GFX12-NEXT:    s_endpgm
3307a6d736SMirko Brkušanin  %ret = call float @llvm.amdgcn.global.atomic.fmax.num.f32.p1.f32(ptr addrspace(1) %ptr, float %data)
3407a6d736SMirko Brkušanin  store float %ret, ptr addrspace(1) %out
3507a6d736SMirko Brkušanin  ret void
3607a6d736SMirko Brkušanin}
3707a6d736SMirko Brkušanin
3807a6d736SMirko Brkušanindefine amdgpu_cs void @global_atomic_fmin_num_f32_rtn(ptr addrspace(1) %ptr, float %data, ptr addrspace(1) %out) {
3907a6d736SMirko Brkušanin; GFX12-LABEL: global_atomic_fmin_num_f32_rtn:
4007a6d736SMirko Brkušanin; GFX12:       ; %bb.0:
4107a6d736SMirko Brkušanin; GFX12-NEXT:    global_atomic_min_num_f32 v0, v[0:1], v2, off th:TH_ATOMIC_RETURN
42*ba52f06fSJay Foad; GFX12-NEXT:    s_wait_loadcnt 0x0
4307a6d736SMirko Brkušanin; GFX12-NEXT:    global_store_b32 v[3:4], v0, off
4407a6d736SMirko Brkušanin; GFX12-NEXT:    s_endpgm
4507a6d736SMirko Brkušanin  %ret = call float @llvm.amdgcn.global.atomic.fmin.num.f32.p1.f32(ptr addrspace(1) %ptr, float %data)
4607a6d736SMirko Brkušanin  store float %ret, ptr addrspace(1) %out
4707a6d736SMirko Brkušanin  ret void
4807a6d736SMirko Brkušanin}
4907a6d736SMirko Brkušanin;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
5007a6d736SMirko Brkušanin; GFX12-GISEL: {{.*}}
5107a6d736SMirko Brkušanin; GFX12-SDAG: {{.*}}
52