1 ; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s 2 3 declare i64 @llvm.ctpop.i64(i64) nounwind readnone 4 declare <2 x i64> @llvm.ctpop.v2i64(<2 x i64>) nounwind readnone 5 declare <4 x i64> @llvm.ctpop.v4i64(<4 x i64>) nounwind readnone 6 declare <8 x i64> @llvm.ctpop.v8i64(<8 x i64>) nounwind readnone 7 declare <16 x i64> @llvm.ctpop.v16i64(<16 x i64>) nounwind readnone 8 9 ; FUNC-LABEL: @s_ctpop_i64: 10 ; SI: S_LOAD_DWORDX2 [[SVAL:s\[[0-9]+:[0-9]+\]]], 11 ; SI: S_BCNT1_I32_B64 [[SRESULT:s[0-9]+]], [[SVAL]] 12 ; SI: V_MOV_B32_e32 [[VRESULT:v[0-9]+]], [[SRESULT]] 13 ; SI: BUFFER_STORE_DWORD [[VRESULT]], 14 ; SI: S_ENDPGM 15 define void @s_ctpop_i64(i32 addrspace(1)* noalias %out, i64 %val) nounwind { 16 %ctpop = call i64 @llvm.ctpop.i64(i64 %val) nounwind readnone 17 %truncctpop = trunc i64 %ctpop to i32 18 store i32 %truncctpop, i32 addrspace(1)* %out, align 4 19 ret void 20 } 21 22 ; FUNC-LABEL: @v_ctpop_i64: 23 ; SI: BUFFER_LOAD_DWORDX2 v{{\[}}[[LOVAL:[0-9]+]]:[[HIVAL:[0-9]+]]{{\]}}, 24 ; SI: V_MOV_B32_e32 [[VZERO:v[0-9]+]], 0 25 ; SI: V_BCNT_U32_B32_e32 [[MIDRESULT:v[0-9]+]], v[[LOVAL]], [[VZERO]] 26 ; SI-NEXT: V_BCNT_U32_B32_e32 [[RESULT:v[0-9]+]], v[[HIVAL]], [[MIDRESULT]] 27 ; SI: BUFFER_STORE_DWORD [[RESULT]], 28 ; SI: S_ENDPGM 29 define void @v_ctpop_i64(i32 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in) nounwind { 30 %val = load i64 addrspace(1)* %in, align 8 31 %ctpop = call i64 @llvm.ctpop.i64(i64 %val) nounwind readnone 32 %truncctpop = trunc i64 %ctpop to i32 33 store i32 %truncctpop, i32 addrspace(1)* %out, align 4 34 ret void 35 } 36 37 ; FUNC-LABEL: @s_ctpop_v2i64: 38 ; SI: S_BCNT1_I32_B64 39 ; SI: S_BCNT1_I32_B64 40 ; SI: S_ENDPGM 41 define void @s_ctpop_v2i64(<2 x i32> addrspace(1)* noalias %out, <2 x i64> %val) nounwind { 42 %ctpop = call <2 x i64> @llvm.ctpop.v2i64(<2 x i64> %val) nounwind readnone 43 %truncctpop = trunc <2 x i64> %ctpop to <2 x i32> 44 store <2 x i32> %truncctpop, <2 x i32> addrspace(1)* %out, align 8 45 ret void 46 } 47 48 ; FUNC-LABEL: @s_ctpop_v4i64: 49 ; SI: S_BCNT1_I32_B64 50 ; SI: S_BCNT1_I32_B64 51 ; SI: S_BCNT1_I32_B64 52 ; SI: S_BCNT1_I32_B64 53 ; SI: S_ENDPGM 54 define void @s_ctpop_v4i64(<4 x i32> addrspace(1)* noalias %out, <4 x i64> %val) nounwind { 55 %ctpop = call <4 x i64> @llvm.ctpop.v4i64(<4 x i64> %val) nounwind readnone 56 %truncctpop = trunc <4 x i64> %ctpop to <4 x i32> 57 store <4 x i32> %truncctpop, <4 x i32> addrspace(1)* %out, align 16 58 ret void 59 } 60 61 ; FUNC-LABEL: @v_ctpop_v2i64: 62 ; SI: V_BCNT_U32_B32 63 ; SI: V_BCNT_U32_B32 64 ; SI: V_BCNT_U32_B32 65 ; SI: V_BCNT_U32_B32 66 ; SI: S_ENDPGM 67 define void @v_ctpop_v2i64(<2 x i32> addrspace(1)* noalias %out, <2 x i64> addrspace(1)* noalias %in) nounwind { 68 %val = load <2 x i64> addrspace(1)* %in, align 16 69 %ctpop = call <2 x i64> @llvm.ctpop.v2i64(<2 x i64> %val) nounwind readnone 70 %truncctpop = trunc <2 x i64> %ctpop to <2 x i32> 71 store <2 x i32> %truncctpop, <2 x i32> addrspace(1)* %out, align 8 72 ret void 73 } 74 75 ; FUNC-LABEL: @v_ctpop_v4i64: 76 ; SI: V_BCNT_U32_B32 77 ; SI: V_BCNT_U32_B32 78 ; SI: V_BCNT_U32_B32 79 ; SI: V_BCNT_U32_B32 80 ; SI: V_BCNT_U32_B32 81 ; SI: V_BCNT_U32_B32 82 ; SI: V_BCNT_U32_B32 83 ; SI: V_BCNT_U32_B32 84 ; SI: S_ENDPGM 85 define void @v_ctpop_v4i64(<4 x i32> addrspace(1)* noalias %out, <4 x i64> addrspace(1)* noalias %in) nounwind { 86 %val = load <4 x i64> addrspace(1)* %in, align 32 87 %ctpop = call <4 x i64> @llvm.ctpop.v4i64(<4 x i64> %val) nounwind readnone 88 %truncctpop = trunc <4 x i64> %ctpop to <4 x i32> 89 store <4 x i32> %truncctpop, <4 x i32> addrspace(1)* %out, align 16 90 ret void 91 } 92 93 ; FIXME: We currently disallow SALU instructions in all branches, 94 ; but there are some cases when the should be allowed. 95 96 ; FUNC-LABEL: @ctpop_i64_in_br 97 ; SI: V_BCNT_U32_B32_e64 [[BCNT_LO:v[0-9]+]], v{{[0-9]+}}, 0 98 ; SI: V_BCNT_U32_B32_e32 v[[BCNT:[0-9]+]], v{{[0-9]+}}, [[BCNT_LO]] 99 ; SI: V_MOV_B32_e32 v[[ZERO:[0-9]+]], 0 100 ; SI: BUFFER_STORE_DWORDX2 v[ 101 ; SI: [[BCNT]]:[[ZERO]]] 102 ; SI: S_ENDPGM 103 define void @ctpop_i64_in_br(i64 addrspace(1)* %out, i64 addrspace(1)* %in, i32 %cond) { 104 entry: 105 %0 = icmp eq i32 %cond, 0 106 br i1 %0, label %if, label %else 107 108 if: 109 %1 = load i64 addrspace(1)* %in 110 %2 = call i64 @llvm.ctpop.i64(i64 %1) 111 br label %endif 112 113 else: 114 %3 = getelementptr i64 addrspace(1)* %in, i32 1 115 %4 = load i64 addrspace(1)* %3 116 br label %endif 117 118 endif: 119 %5 = phi i64 [%2, %if], [%4, %else] 120 store i64 %5, i64 addrspace(1)* %out 121 ret void 122 } 123
