299 lines · plain
1; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1010 --stop-after=amdgpu-isel < %s | FileCheck -check-prefixes=SDAG,SDAG-GFX10 %s2; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1010 --stop-after=instruction-select < %s | FileCheck -check-prefixes=GISEL %s3; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 --stop-after=amdgpu-isel < %s | FileCheck -check-prefixes=SDAG,SDAG-GFX11 %s4; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 --stop-after=instruction-select < %s | FileCheck -check-prefixes=GISEL %s5 6declare i32 @llvm.amdgcn.permlane16(i32, i32, i32, i32, i1, i1)7declare i32 @llvm.amdgcn.permlanex16(i32, i32, i32, i32, i1, i1)8declare i32 @llvm.amdgcn.workitem.id.x()9declare i32 @llvm.amdgcn.workitem.id.y()10 11 12define amdgpu_kernel void @v_permlane16_b32_vss(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {13 ; SDAG: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec14 ; GISEL: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec15 %v = call i32 @llvm.amdgcn.permlane16(i32 %src0, i32 %src0, i32 %src1, i32 %src2, i1 false, i1 false)16 store i32 %v, ptr addrspace(1) %out17 ret void18}19 20define amdgpu_kernel void @v_permlane16_b32_vii(ptr addrspace(1) %out, i32 %src0) {21 ; SDAG: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec22 ; GISEL: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec23 %v = call i32 @llvm.amdgcn.permlane16(i32 %src0, i32 %src0, i32 1, i32 2, i1 false, i1 false)24 store i32 %v, ptr addrspace(1) %out25 ret void26}27 28; FIXME-GFX10PLUS: It is allowed to have both immediates as literals29define amdgpu_kernel void @v_permlane16_b32_vll(ptr addrspace(1) %out, i32 %src0) {30 ; SDAG: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec31 ; GISEL: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec32 %v = call i32 @llvm.amdgcn.permlane16(i32 %src0, i32 %src0, i32 4660, i32 49617, i1 false, i1 false)33 store i32 %v, ptr addrspace(1) %out34 ret void35}36 37define amdgpu_kernel void @v_permlane16_b32_vvv(ptr addrspace(1) %out, i32 %src0) {38 ; SDAG-GFX10: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec39 ; SDAG-GFX11: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec40 ; GISEL: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec41 %tidx = call i32 @llvm.amdgcn.workitem.id.x()42 %tidy = call i32 @llvm.amdgcn.workitem.id.y()43 %v = call i32 @llvm.amdgcn.permlane16(i32 %src0, i32 %src0, i32 %tidx, i32 %tidy, i1 false, i1 false)44 store i32 %v, ptr addrspace(1) %out45 ret void46}47 48define amdgpu_kernel void @v_permlane16_b32_vvs(ptr addrspace(1) %out, i32 %src0, i32 %src2) {49; SDAG-GFX10: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec50; SDAG-GFX11: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec51; GISEL: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec52 %tidx = call i32 @llvm.amdgcn.workitem.id.x()53 %v = call i32 @llvm.amdgcn.permlane16(i32 %src0, i32 %src0, i32 %tidx, i32 %src2, i1 false, i1 false)54 store i32 %v, ptr addrspace(1) %out55 ret void56}57 58define amdgpu_kernel void @v_permlane16_b32_vsv(ptr addrspace(1) %out, i32 %src0, i32 %src1) {59 ; SDAG-GFX10: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec60 ; SDAG-GFX11: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec61 ; GISEL: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec62 %tidy = call i32 @llvm.amdgcn.workitem.id.y()63 %v = call i32 @llvm.amdgcn.permlane16(i32 %src0, i32 %src0, i32 %src1, i32 %tidy, i1 false, i1 false)64 store i32 %v, ptr addrspace(1) %out65 ret void66}67 68define amdgpu_kernel void @v_permlane16_b32_vss_fi(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {69 ; SDAG: V_PERMLANE16_B32_e64 4, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec70 ; GISEL: V_PERMLANE16_B32_e64 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec71 %v = call i32 @llvm.amdgcn.permlane16(i32 %src0, i32 %src0, i32 %src1, i32 %src2, i1 true, i1 false)72 store i32 %v, ptr addrspace(1) %out73 ret void74}75 76define amdgpu_kernel void @v_permlane16_b32_vss_bc(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {77 ; SDAG: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec78 ; GISEL: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec79 %v = call i32 @llvm.amdgcn.permlane16(i32 %src0, i32 %src0, i32 %src1, i32 %src2, i1 false, i1 true)80 store i32 %v, ptr addrspace(1) %out81 ret void82}83 84define amdgpu_kernel void @v_permlane16_b32_vss_fi_bc(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {85 ; SDAG: V_PERMLANE16_B32_e64 4, {{%[0-9]+}}, 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec86 ; GISEL: V_PERMLANE16_B32_e64 4, {{%[0-9]+}}, 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec87 %v = call i32 @llvm.amdgcn.permlane16(i32 %src0, i32 %src0, i32 %src1, i32 %src2, i1 true, i1 true)88 store i32 %v, ptr addrspace(1) %out89 ret void90}91 92define amdgpu_kernel void @v_permlanex16_b32_vss(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {93 ; SDAG: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec94 ; GISEL: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec95 %v = call i32 @llvm.amdgcn.permlanex16(i32 %src0, i32 %src0, i32 %src1, i32 %src2, i1 false, i1 false)96 store i32 %v, ptr addrspace(1) %out97 ret void98}99 100define amdgpu_kernel void @v_permlanex16_b32_vii(ptr addrspace(1) %out, i32 %src0) {101 ; SDAG: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec102 ; GISEL: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec103 %v = call i32 @llvm.amdgcn.permlanex16(i32 %src0, i32 %src0, i32 1, i32 2, i1 false, i1 false)104 store i32 %v, ptr addrspace(1) %out105 ret void106}107 108; FIXME-GFX10PLUS: It is allowed to have both immediates as literals109define amdgpu_kernel void @v_permlanex16_b32_vll(ptr addrspace(1) %out, i32 %src0) {110 ; SDAG: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec111 ; GISEL: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec112 %v = call i32 @llvm.amdgcn.permlanex16(i32 %src0, i32 %src0, i32 4660, i32 49617, i1 false, i1 false)113 store i32 %v, ptr addrspace(1) %out114 ret void115}116 117define amdgpu_kernel void @v_permlanex16_b32_vvv(ptr addrspace(1) %out, i32 %src0) {118 ; SDAG-GFX10: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec119 ; SDAG-GFX11: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec120 ; GISEL: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec121 %tidx = call i32 @llvm.amdgcn.workitem.id.x()122 %tidy = call i32 @llvm.amdgcn.workitem.id.y()123 %v = call i32 @llvm.amdgcn.permlanex16(i32 %src0, i32 %src0, i32 %tidx, i32 %tidy, i1 false, i1 false)124 store i32 %v, ptr addrspace(1) %out125 ret void126}127 128define amdgpu_kernel void @v_permlanex16_b32_vvs(ptr addrspace(1) %out, i32 %src0, i32 %src2) {129 ; SDAG-GFX10: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec130 ; SDAG-GFX11: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec131 ; GISEL: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec132 %tidx = call i32 @llvm.amdgcn.workitem.id.x()133 %v = call i32 @llvm.amdgcn.permlanex16(i32 %src0, i32 %src0, i32 %tidx, i32 %src2, i1 false, i1 false)134 store i32 %v, ptr addrspace(1) %out135 ret void136}137 138define amdgpu_kernel void @v_permlanex16_b32_vsv(ptr addrspace(1) %out, i32 %src0, i32 %src1) {139 ; SDAG-GFX10: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec140 ; SDAG-GFX11: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec141 ; GISEL: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec142 %tidy = call i32 @llvm.amdgcn.workitem.id.y()143 %v = call i32 @llvm.amdgcn.permlanex16(i32 %src0, i32 %src0, i32 %src1, i32 %tidy, i1 false, i1 false)144 store i32 %v, ptr addrspace(1) %out145 ret void146}147 148define amdgpu_kernel void @v_permlanex16_b32_vss_fi(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {149 ; SDAG: V_PERMLANEX16_B32_e64 4, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec150 ; GISEL: V_PERMLANEX16_B32_e64 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec151 %v = call i32 @llvm.amdgcn.permlanex16(i32 %src0, i32 %src0, i32 %src1, i32 %src2, i1 true, i1 false)152 store i32 %v, ptr addrspace(1) %out153 ret void154}155 156define amdgpu_kernel void @v_permlanex16_b32_vss_bc(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {157 ; SDAG: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec158 ; GISEL: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec159 %v = call i32 @llvm.amdgcn.permlanex16(i32 %src0, i32 %src0, i32 %src1, i32 %src2, i1 false, i1 true)160 store i32 %v, ptr addrspace(1) %out161 ret void162}163 164define amdgpu_kernel void @v_permlanex16_b32_vss_fi_bc(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {165 ; SDAG: V_PERMLANEX16_B32_e64 4, {{%[0-9]+}}, 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec166 ; GISEL: V_PERMLANEX16_B32_e64 4, {{%[0-9]+}}, 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec167 %v = call i32 @llvm.amdgcn.permlanex16(i32 %src0, i32 %src0, i32 %src1, i32 %src2, i1 true, i1 true)168 store i32 %v, ptr addrspace(1) %out169 ret void170}171 172define amdgpu_kernel void @v_permlane16_b32_tid_tid(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {173 ; SDAG-GFX10: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}(s32), 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}(s32), 0, implicit $exec174 ; SDAG-GFX11: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec175 ; GISEL: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec176 %tidx = call i32 @llvm.amdgcn.workitem.id.x()177 %v = call i32 @llvm.amdgcn.permlane16(i32 %tidx, i32 %tidx, i32 %src1, i32 %src2, i1 false, i1 false)178 store i32 %v, ptr addrspace(1) %out179 ret void180}181 182define amdgpu_kernel void @v_permlane16_b32_undef_tid(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {183 ; SDAG-GFX10: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}(s32), 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec184 ; SDAG-GFX11: V_PERMLANE16_B32_e64 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec185 ; GISEL: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec186 %tidx = call i32 @llvm.amdgcn.workitem.id.x()187 %undef = freeze i32 poison188 %v = call i32 @llvm.amdgcn.permlane16(i32 %undef, i32 %tidx, i32 %src1, i32 %src2, i1 false, i1 false)189 store i32 %v, ptr addrspace(1) %out190 ret void191}192 193define amdgpu_kernel void @v_permlane16_b32_i_tid(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {194 ; SDAG-GFX10: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}(s32), 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec195 ; SDAG-GFX11: V_PERMLANE16_B32_e64 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec196 ; GISEL: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec197 %tidx = call i32 @llvm.amdgcn.workitem.id.x()198 %v = call i32 @llvm.amdgcn.permlane16(i32 12345, i32 %tidx, i32 %src1, i32 %src2, i1 false, i1 false)199 store i32 %v, ptr addrspace(1) %out200 ret void201}202 203define amdgpu_kernel void @v_permlane16_b32_i_tid_fi(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {204 ; SDAG-GFX10: V_PERMLANE16_B32_e64 4, {{%[0-9]+}}(s32), 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec205 ; SDAG-GFX11: V_PERMLANE16_B32_e64 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec206 ; GISEL: V_PERMLANE16_B32_e64 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec207 %tidx = call i32 @llvm.amdgcn.workitem.id.x()208 %undef = freeze i32 poison209 %v = call i32 @llvm.amdgcn.permlane16(i32 %undef, i32 %tidx, i32 %src1, i32 %src2, i1 true, i1 false)210 store i32 %v, ptr addrspace(1) %out211 ret void212}213 214define amdgpu_kernel void @v_permlane16_b32_i_tid_bc(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {215 ; SDAG-GFX10: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}(s32), 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec216 ; SDAG-GFX11: V_PERMLANE16_B32_e64 0, killed {{%[0-9]+}}, 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec217 ; GISEL: V_PERMLANE16_B32_e64 0, {{%[0-9]+}}, 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec218 %tidx = call i32 @llvm.amdgcn.workitem.id.x()219 %undef = freeze i32 poison220 %v = call i32 @llvm.amdgcn.permlane16(i32 %undef, i32 %tidx, i32 %src1, i32 %src2, i1 false, i1 true)221 store i32 %v, ptr addrspace(1) %out222 ret void223}224 225define amdgpu_kernel void @v_permlane16_b32_i_tid_fi_bc(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {226 ; SDAG-GFX10: V_PERMLANE16_B32_e64 4, {{%[0-9]+}}(s32), 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec227 ; SDAG-GFX11: V_PERMLANE16_B32_e64 4, killed {{%[0-9]+}}, 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec228 ; GISEL: V_PERMLANE16_B32_e64 4, {{%[0-9]+}}, 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec229 %tidx = call i32 @llvm.amdgcn.workitem.id.x()230 %undef = freeze i32 poison231 %v = call i32 @llvm.amdgcn.permlane16(i32 %undef, i32 %tidx, i32 %src1, i32 %src2, i1 true, i1 true)232 store i32 %v, ptr addrspace(1) %out233 ret void234}235 236define amdgpu_kernel void @v_permlanex16_b32_tid_tid(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {237 ; SDAG-GFX10: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}(s32), 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}(s32), 0, implicit $exec238 ; SDAG-GFX11: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec239 ; GISEL: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec240 %tidx = call i32 @llvm.amdgcn.workitem.id.x()241 %v = call i32 @llvm.amdgcn.permlanex16(i32 %tidx, i32 %tidx, i32 %src1, i32 %src2, i1 false, i1 false)242 store i32 %v, ptr addrspace(1) %out243 ret void244}245 246define amdgpu_kernel void @v_permlanex16_b32_undef_tid(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {247 ; SDAG-GFX10: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}(s32), 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec248 ; SDAG-GFX11: V_PERMLANEX16_B32_e64 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec249 ; GISEL: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec250 %tidx = call i32 @llvm.amdgcn.workitem.id.x()251 %undef = freeze i32 poison252 %v = call i32 @llvm.amdgcn.permlanex16(i32 %undef, i32 %tidx, i32 %src1, i32 %src2, i1 false, i1 false)253 store i32 %v, ptr addrspace(1) %out254 ret void255}256 257define amdgpu_kernel void @v_permlanex16_b32_i_tid(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {258 ; SDAG-GFX10: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}(s32), 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec259 ; SDAG-GFX11: V_PERMLANEX16_B32_e64 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec260 ; GISEL: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec261 %tidx = call i32 @llvm.amdgcn.workitem.id.x()262 %v = call i32 @llvm.amdgcn.permlanex16(i32 12345, i32 %tidx, i32 %src1, i32 %src2, i1 false, i1 false)263 store i32 %v, ptr addrspace(1) %out264 ret void265}266 267define amdgpu_kernel void @v_permlanex16_b32_i_tid_fi(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {268 ; SDAG-GFX10: V_PERMLANEX16_B32_e64 4, {{%[0-9]+}}(s32), 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec269 ; SDAG-GFX11: V_PERMLANEX16_B32_e64 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec270 ; GISEL: V_PERMLANEX16_B32_e64 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec271 %tidx = call i32 @llvm.amdgcn.workitem.id.x()272 %undef = freeze i32 poison273 %v = call i32 @llvm.amdgcn.permlanex16(i32 %undef, i32 %tidx, i32 %src1, i32 %src2, i1 true, i1 false)274 store i32 %v, ptr addrspace(1) %out275 ret void276}277 278define amdgpu_kernel void @v_permlanex16_b32_i_tid_bc(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {279 ; SDAG-GFX10: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}(s32), 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec280 ; SDAG-GFX11: V_PERMLANEX16_B32_e64 0, killed {{%[0-9]+}}, 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec281 ; GISEL: V_PERMLANEX16_B32_e64 0, {{%[0-9]+}}, 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec282 %tidx = call i32 @llvm.amdgcn.workitem.id.x()283 %undef = freeze i32 poison284 %v = call i32 @llvm.amdgcn.permlanex16(i32 %undef, i32 %tidx, i32 %src1, i32 %src2, i1 false, i1 true)285 store i32 %v, ptr addrspace(1) %out286 ret void287}288 289define amdgpu_kernel void @v_permlanex16_b32_i_tid_fi_bc(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) {290 ; SDAG-GFX10: V_PERMLANEX16_B32_e64 4, {{%[0-9]+}}(s32), 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec291 ; SDAG-GFX11: V_PERMLANEX16_B32_e64 4, killed {{%[0-9]+}}, 4, killed {{%[0-9]+}}, 0, killed {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec292 ; GISEL: V_PERMLANEX16_B32_e64 4, {{%[0-9]+}}, 4, {{%[0-9]+}}, 0, {{%[0-9]+}}, {{%[0-9]+}}, 0, implicit $exec293 %tidx = call i32 @llvm.amdgcn.workitem.id.x()294 %undef = freeze i32 poison295 %v = call i32 @llvm.amdgcn.permlanex16(i32 %undef, i32 %tidx, i32 %src1, i32 %src2, i1 true, i1 true)296 store i32 %v, ptr addrspace(1) %out297 ret void298}299