55 lines · plain
1; RUN: llc -mtriple=amdgcn -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GCN %s2; RUN: llc -mtriple=amdgcn -enable-new-pm -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GCN %s3 4; GCN-LABEL: name: s_ctpop_i325; GCN: S_BCNT1_I32_B326define amdgpu_kernel void @s_ctpop_i32(ptr addrspace(1) noalias %out, i32 %val) nounwind {7 %ctpop = call i32 @llvm.ctpop.i32(i32 %val) nounwind readnone8 store i32 %ctpop, ptr addrspace(1) %out, align 49 ret void10}11 12; GCN-LABEL: name: s_ctpop_i6413; GCN: %[[BCNT:[0-9]+]]:sreg_32 = S_BCNT1_I32_B6414; GCN: %[[SREG1:[0-9]+]]:sreg_32 = COPY killed %[[BCNT]]15; GCN: %[[SREG2:[0-9]+]]:sreg_32 = S_MOV_B32 016; GCN: REG_SEQUENCE killed %[[SREG1]], %subreg.sub0, killed %[[SREG2]], %subreg.sub117define amdgpu_kernel void @s_ctpop_i64(ptr addrspace(1) noalias %out, i64 %val) nounwind {18 %ctpop = call i64 @llvm.ctpop.i64(i64 %val) nounwind readnone19 %truncctpop = trunc i64 %ctpop to i3220 store i32 %truncctpop, ptr addrspace(1) %out, align 421 ret void22}23 24; GCN-LABEL: name: v_ctpop_i3225; GCN: V_BCNT_U32_B32_e6426define amdgpu_kernel void @v_ctpop_i32(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) nounwind {27 %tid = call i32 @llvm.amdgcn.workitem.id.x()28 %in.gep = getelementptr i32, ptr addrspace(1) %in, i32 %tid29 %val = load i32, ptr addrspace(1) %in.gep, align 430 %ctpop = call i32 @llvm.ctpop.i32(i32 %val) nounwind readnone31 store i32 %ctpop, ptr addrspace(1) %out, align 432 ret void33}34 35; GCN-LABEL: name: v_ctpop_i6436; GCN: %[[BCNT1:[0-9]+]]:vgpr_32 = V_BCNT_U32_B32_e64 killed %{{[0-9]+}}, 0, implicit $exec37; GCN: %[[BCNT2:[0-9]+]]:vgpr_32 = V_BCNT_U32_B32_e64 killed %{{[0-9]+}}, killed %[[BCNT1]], implicit $exec38; GCN: %[[VGPR1:[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec39; GCN: REG_SEQUENCE killed %[[BCNT2]], %subreg.sub0, killed %[[VGPR1]], %subreg.sub140define amdgpu_kernel void @v_ctpop_i64(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) nounwind {41 %tid = call i32 @llvm.amdgcn.workitem.id.x()42 %in.gep = getelementptr i64, ptr addrspace(1) %in, i32 %tid43 %val = load i64, ptr addrspace(1) %in.gep, align 844 %ctpop = call i64 @llvm.ctpop.i64(i64 %val) nounwind readnone45 %truncctpop = trunc i64 %ctpop to i3246 store i32 %truncctpop, ptr addrspace(1) %out, align 447 ret void48}49 50declare i64 @llvm.ctpop.i64(i64) nounwind readnone51 52declare i32 @llvm.ctpop.i32(i32) nounwind readnone53 54declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone55