77 lines · plain
1; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s2; RUN: llc -mtriple=amdgcn -mcpu=bonaire < %s | FileCheck -check-prefix=CI -check-prefix=FUNC %s3 4declare i32 @llvm.amdgcn.workitem.id.x() #15declare double @llvm.fabs.f64(double) #16 7; FUNC-LABEL: @fp_to_sint_f64_i328; SI: v_cvt_i32_f64_e329define amdgpu_kernel void @fp_to_sint_f64_i32(ptr addrspace(1) %out, double %in) {10 %result = fptosi double %in to i3211 store i32 %result, ptr addrspace(1) %out12 ret void13}14 15; FUNC-LABEL: @fp_to_sint_v2f64_v2i3216; SI: v_cvt_i32_f64_e3217; SI: v_cvt_i32_f64_e3218define amdgpu_kernel void @fp_to_sint_v2f64_v2i32(ptr addrspace(1) %out, <2 x double> %in) {19 %result = fptosi <2 x double> %in to <2 x i32>20 store <2 x i32> %result, ptr addrspace(1) %out21 ret void22}23 24; FUNC-LABEL: @fp_to_sint_v4f64_v4i3225; SI: v_cvt_i32_f64_e3226; SI: v_cvt_i32_f64_e3227; SI: v_cvt_i32_f64_e3228; SI: v_cvt_i32_f64_e3229define amdgpu_kernel void @fp_to_sint_v4f64_v4i32(ptr addrspace(1) %out, <4 x double> %in) {30 %result = fptosi <4 x double> %in to <4 x i32>31 store <4 x i32> %result, ptr addrspace(1) %out32 ret void33}34 35; FUNC-LABEL: @fp_to_sint_i64_f6436; CI-DAG: buffer_load_dwordx2 [[VAL:v\[[0-9]+:[0-9]+\]]]37; CI-DAG: v_trunc_f64_e32 [[TRUNC:v\[[0-9]+:[0-9]+\]]], [[VAL]]38; CI-DAG: s_movk_i32 [[K0_EXP:s[0-9]+]], 0xffe039 40; CI-DAG: v_ldexp_f64 [[MUL:v\[[0-9]+:[0-9]+\]]], [[VAL]], [[K0_EXP]]41; CI-DAG: v_floor_f64_e32 [[FLOOR:v\[[0-9]+:[0-9]+\]]], [[MUL]]42 43; CI-DAG: s_mov_b32 s[[K1_HI:[0-9]+]], 0xc1f0000044 45; CI-DAG: v_fma_f64 [[FMA:v\[[0-9]+:[0-9]+\]]], [[FLOOR]], s{{\[[0-9]+}}:[[K1_HI]]], [[TRUNC]]46; CI-DAG: v_cvt_u32_f64_e32 v[[LO:[0-9]+]], [[FMA]]47; CI-DAG: v_cvt_i32_f64_e32 v[[HI:[0-9]+]], [[FLOOR]]48; CI: buffer_store_dwordx2 v[[[LO]]:[[HI]]]49define amdgpu_kernel void @fp_to_sint_i64_f64(ptr addrspace(1) %out, ptr addrspace(1) %in) {50 %tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone51 %gep = getelementptr double, ptr addrspace(1) %in, i32 %tid52 %val = load double, ptr addrspace(1) %gep, align 853 %cast = fptosi double %val to i6454 store i64 %cast, ptr addrspace(1) %out, align 855 ret void56}57 58; FUNC-LABEL: {{^}}fp_to_sint_f64_to_i1:59; SI: v_cmp_eq_f64_e64 s{{\[[0-9]+:[0-9]+\]}}, -1.0, s{{\[[0-9]+:[0-9]+\]}}60define amdgpu_kernel void @fp_to_sint_f64_to_i1(ptr addrspace(1) %out, double %in) #0 {61 %conv = fptosi double %in to i162 store i1 %conv, ptr addrspace(1) %out63 ret void64}65 66; FUNC-LABEL: {{^}}fp_to_sint_fabs_f64_to_i1:67; SI: v_cmp_eq_f64_e64 s{{\[[0-9]+:[0-9]+\]}}, -1.0, |s{{\[[0-9]+:[0-9]+\]}}|68define amdgpu_kernel void @fp_to_sint_fabs_f64_to_i1(ptr addrspace(1) %out, double %in) #0 {69 %in.fabs = call double @llvm.fabs.f64(double %in)70 %conv = fptosi double %in.fabs to i171 store i1 %conv, ptr addrspace(1) %out72 ret void73}74 75attributes #0 = { nounwind }76attributes #1 = { nounwind readnone }77