246 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=-architected-sgprs -global-isel=0 < %s | FileCheck -check-prefixes=GFX9-SDAG %s3; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=-architected-sgprs -global-isel=1 < %s | FileCheck -check-prefixes=GFX9-GISEL %s4; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=+architected-sgprs -global-isel=0 < %s | FileCheck -check-prefixes=GFX9ARCH-SDAG %s5; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=+architected-sgprs -global-isel=1 < %s | FileCheck -check-prefixes=GFX9ARCH-GISEL %s6; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 -global-isel=0 < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG %s7; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 -global-isel=1 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL %s8 9define amdgpu_cs void @_amdgpu_cs_main() {10; GFX9-LABEL: _amdgpu_cs_main:11; GFX9: ; %bb.0: ; %.entry12; GFX9-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 013; GFX9-NEXT: s_endpgm14;15; GFX9-SDAG-LABEL: _amdgpu_cs_main:16; GFX9-SDAG: ; %bb.0: ; %.entry17; GFX9-SDAG-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 018; GFX9-SDAG-NEXT: s_endpgm19;20; GFX9-GISEL-LABEL: _amdgpu_cs_main:21; GFX9-GISEL: ; %bb.0: ; %.entry22; GFX9-GISEL-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 023; GFX9-GISEL-NEXT: s_endpgm24;25; GFX9ARCH-SDAG-LABEL: _amdgpu_cs_main:26; GFX9ARCH-SDAG: ; %bb.0: ; %.entry27; GFX9ARCH-SDAG-NEXT: s_lshr_b32 s0, ttmp7, 1628; GFX9ARCH-SDAG-NEXT: s_and_b32 s1, ttmp7, 0xffff29; GFX9ARCH-SDAG-NEXT: v_mov_b32_e32 v0, ttmp930; GFX9ARCH-SDAG-NEXT: v_mov_b32_e32 v1, s131; GFX9ARCH-SDAG-NEXT: v_mov_b32_e32 v2, s032; GFX9ARCH-SDAG-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 033; GFX9ARCH-SDAG-NEXT: s_endpgm34;35; GFX9ARCH-GISEL-LABEL: _amdgpu_cs_main:36; GFX9ARCH-GISEL: ; %bb.0: ; %.entry37; GFX9ARCH-GISEL-NEXT: s_mov_b32 s0, ttmp938; GFX9ARCH-GISEL-NEXT: s_and_b32 s1, ttmp7, 0xffff39; GFX9ARCH-GISEL-NEXT: s_lshr_b32 s2, ttmp7, 1640; GFX9ARCH-GISEL-NEXT: v_mov_b32_e32 v0, s041; GFX9ARCH-GISEL-NEXT: v_mov_b32_e32 v1, s142; GFX9ARCH-GISEL-NEXT: v_mov_b32_e32 v2, s243; GFX9ARCH-GISEL-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 044; GFX9ARCH-GISEL-NEXT: s_endpgm45;46; GFX12-SDAG-LABEL: _amdgpu_cs_main:47; GFX12-SDAG: ; %bb.0: ; %.entry48; GFX12-SDAG-NEXT: s_and_b32 s0, ttmp7, 0xffff49; GFX12-SDAG-NEXT: s_lshr_b32 s1, ttmp7, 1650; GFX12-SDAG-NEXT: v_dual_mov_b32 v0, ttmp9 :: v_dual_mov_b32 v1, s051; GFX12-SDAG-NEXT: v_mov_b32_e32 v2, s152; GFX12-SDAG-NEXT: buffer_store_b96 v[0:2], off, s[0:3], null53; GFX12-SDAG-NEXT: s_endpgm54;55; GFX12-GISEL-LABEL: _amdgpu_cs_main:56; GFX12-GISEL: ; %bb.0: ; %.entry57; GFX12-GISEL-NEXT: s_mov_b32 s0, ttmp958; GFX12-GISEL-NEXT: s_and_b32 s1, ttmp7, 0xffff59; GFX12-GISEL-NEXT: s_lshr_b32 s2, ttmp7, 1660; GFX12-GISEL-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s161; GFX12-GISEL-NEXT: v_mov_b32_e32 v2, s262; GFX12-GISEL-NEXT: buffer_store_b96 v[0:2], off, s[0:3], null63; GFX12-GISEL-NEXT: s_endpgm64.entry:65 %idx = call i32 @llvm.amdgcn.workgroup.id.x()66 %idy = call i32 @llvm.amdgcn.workgroup.id.y()67 %idz = call i32 @llvm.amdgcn.workgroup.id.z()68 %ielemx = insertelement <3 x i32> poison, i32 %idx, i64 069 %ielemy = insertelement <3 x i32> %ielemx, i32 %idy, i64 170 %ielemz = insertelement <3 x i32> %ielemy, i32 %idz, i64 271 call void @llvm.amdgcn.raw.ptr.buffer.store.v3i32(<3 x i32> %ielemz, ptr addrspace(8) poison, i32 0, i32 0, i32 0)72 ret void73}74 75define amdgpu_cs void @caller() {76; GFX9-SDAG-LABEL: caller:77; GFX9-SDAG: ; %bb.0:78; GFX9-SDAG-NEXT: s_getpc_b64 s[8:9]79; GFX9-SDAG-NEXT: s_mov_b32 s8, s080; GFX9-SDAG-NEXT: s_load_dwordx4 s[8:11], s[8:9], 0x1081; GFX9-SDAG-NEXT: s_mov_b32 s5, callee@abs32@hi82; GFX9-SDAG-NEXT: s_mov_b32 s4, callee@abs32@lo83; GFX9-SDAG-NEXT: s_mov_b32 s32, 084; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0)85; GFX9-SDAG-NEXT: s_add_u32 s8, s8, s086; GFX9-SDAG-NEXT: s_addc_u32 s9, s9, 087; GFX9-SDAG-NEXT: s_mov_b64 s[0:1], s[8:9]88; GFX9-SDAG-NEXT: s_mov_b64 s[2:3], s[10:11]89; GFX9-SDAG-NEXT: s_swappc_b64 s[30:31], s[4:5]90; GFX9-SDAG-NEXT: s_endpgm91;92; GFX9-GISEL-LABEL: caller:93; GFX9-GISEL: ; %bb.0:94; GFX9-GISEL-NEXT: s_getpc_b64 s[8:9]95; GFX9-GISEL-NEXT: s_mov_b32 s8, s096; GFX9-GISEL-NEXT: s_load_dwordx4 s[8:11], s[8:9], 0x1097; GFX9-GISEL-NEXT: s_mov_b32 s4, callee@abs32@lo98; GFX9-GISEL-NEXT: s_mov_b32 s5, callee@abs32@hi99; GFX9-GISEL-NEXT: s_mov_b32 s32, 0100; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0)101; GFX9-GISEL-NEXT: s_add_u32 s8, s8, s0102; GFX9-GISEL-NEXT: s_addc_u32 s9, s9, 0103; GFX9-GISEL-NEXT: s_mov_b64 s[0:1], s[8:9]104; GFX9-GISEL-NEXT: s_mov_b64 s[2:3], s[10:11]105; GFX9-GISEL-NEXT: s_swappc_b64 s[30:31], s[4:5]106; GFX9-GISEL-NEXT: s_endpgm107;108; GFX9ARCH-SDAG-LABEL: caller:109; GFX9ARCH-SDAG: ; %bb.0:110; GFX9ARCH-SDAG-NEXT: s_getpc_b64 s[8:9]111; GFX9ARCH-SDAG-NEXT: s_mov_b32 s8, s0112; GFX9ARCH-SDAG-NEXT: s_load_dwordx4 s[8:11], s[8:9], 0x10113; GFX9ARCH-SDAG-NEXT: s_mov_b32 s5, callee@abs32@hi114; GFX9ARCH-SDAG-NEXT: s_mov_b32 s4, callee@abs32@lo115; GFX9ARCH-SDAG-NEXT: v_mov_b32_e32 v0, ttmp9116; GFX9ARCH-SDAG-NEXT: s_mov_b32 s32, 0117; GFX9ARCH-SDAG-NEXT: s_waitcnt lgkmcnt(0)118; GFX9ARCH-SDAG-NEXT: s_add_u32 s8, s8, s0119; GFX9ARCH-SDAG-NEXT: s_addc_u32 s9, s9, 0120; GFX9ARCH-SDAG-NEXT: s_mov_b64 s[0:1], s[8:9]121; GFX9ARCH-SDAG-NEXT: s_mov_b64 s[2:3], s[10:11]122; GFX9ARCH-SDAG-NEXT: s_swappc_b64 s[30:31], s[4:5]123; GFX9ARCH-SDAG-NEXT: s_endpgm124;125; GFX9ARCH-GISEL-LABEL: caller:126; GFX9ARCH-GISEL: ; %bb.0:127; GFX9ARCH-GISEL-NEXT: s_getpc_b64 s[8:9]128; GFX9ARCH-GISEL-NEXT: s_mov_b32 s8, s0129; GFX9ARCH-GISEL-NEXT: s_load_dwordx4 s[8:11], s[8:9], 0x10130; GFX9ARCH-GISEL-NEXT: s_mov_b32 s4, callee@abs32@lo131; GFX9ARCH-GISEL-NEXT: s_mov_b32 s5, callee@abs32@hi132; GFX9ARCH-GISEL-NEXT: v_mov_b32_e32 v0, ttmp9133; GFX9ARCH-GISEL-NEXT: s_mov_b32 s32, 0134; GFX9ARCH-GISEL-NEXT: s_waitcnt lgkmcnt(0)135; GFX9ARCH-GISEL-NEXT: s_add_u32 s8, s8, s0136; GFX9ARCH-GISEL-NEXT: s_addc_u32 s9, s9, 0137; GFX9ARCH-GISEL-NEXT: s_mov_b64 s[0:1], s[8:9]138; GFX9ARCH-GISEL-NEXT: s_mov_b64 s[2:3], s[10:11]139; GFX9ARCH-GISEL-NEXT: s_swappc_b64 s[30:31], s[4:5]140; GFX9ARCH-GISEL-NEXT: s_endpgm141;142; GFX12-SDAG-LABEL: caller:143; GFX12-SDAG: ; %bb.0:144; GFX12-SDAG-NEXT: v_mov_b32_e32 v0, ttmp9145; GFX12-SDAG-NEXT: s_mov_b32 s1, callee@abs32@hi146; GFX12-SDAG-NEXT: s_mov_b32 s0, callee@abs32@lo147; GFX12-SDAG-NEXT: s_mov_b32 s32, 0148; GFX12-SDAG-NEXT: s_wait_alu depctr_sa_sdst(0)149; GFX12-SDAG-NEXT: s_swappc_b64 s[30:31], s[0:1]150; GFX12-SDAG-NEXT: s_endpgm151;152; GFX12-GISEL-LABEL: caller:153; GFX12-GISEL: ; %bb.0:154; GFX12-GISEL-NEXT: v_mov_b32_e32 v0, ttmp9155; GFX12-GISEL-NEXT: s_mov_b32 s0, callee@abs32@lo156; GFX12-GISEL-NEXT: s_mov_b32 s1, callee@abs32@hi157; GFX12-GISEL-NEXT: s_mov_b32 s32, 0158; GFX12-GISEL-NEXT: s_wait_alu depctr_sa_sdst(0)159; GFX12-GISEL-NEXT: s_swappc_b64 s[30:31], s[0:1]160; GFX12-GISEL-NEXT: s_endpgm161 %idx = call i32 @llvm.amdgcn.workgroup.id.x()162 call amdgpu_gfx void @callee(i32 %idx)163 ret void164}165 166declare amdgpu_gfx void @callee(i32)167 168define amdgpu_gfx void @workgroup_ids_gfx(ptr addrspace(1) %outx, ptr addrspace(1) %outy, ptr addrspace(1) %outz) {169; GFX9-SDAG-LABEL: workgroup_ids_gfx:170; GFX9-SDAG: ; %bb.0:171; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)172; GFX9-SDAG-NEXT: global_store_dword v[0:1], v0, off173; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)174; GFX9-SDAG-NEXT: global_store_dword v[2:3], v0, off175; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)176; GFX9-SDAG-NEXT: global_store_dword v[4:5], v0, off177; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)178; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]179;180; GFX9-GISEL-LABEL: workgroup_ids_gfx:181; GFX9-GISEL: ; %bb.0:182; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)183; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]184;185; GFX9ARCH-SDAG-LABEL: workgroup_ids_gfx:186; GFX9ARCH-SDAG: ; %bb.0:187; GFX9ARCH-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)188; GFX9ARCH-SDAG-NEXT: v_mov_b32_e32 v6, ttmp9189; GFX9ARCH-SDAG-NEXT: s_and_b32 s34, ttmp7, 0xffff190; GFX9ARCH-SDAG-NEXT: global_store_dword v[0:1], v6, off191; GFX9ARCH-SDAG-NEXT: s_waitcnt vmcnt(0)192; GFX9ARCH-SDAG-NEXT: v_mov_b32_e32 v0, s34193; GFX9ARCH-SDAG-NEXT: s_lshr_b32 s34, ttmp7, 16194; GFX9ARCH-SDAG-NEXT: global_store_dword v[2:3], v0, off195; GFX9ARCH-SDAG-NEXT: s_waitcnt vmcnt(0)196; GFX9ARCH-SDAG-NEXT: v_mov_b32_e32 v0, s34197; GFX9ARCH-SDAG-NEXT: global_store_dword v[4:5], v0, off198; GFX9ARCH-SDAG-NEXT: s_waitcnt vmcnt(0)199; GFX9ARCH-SDAG-NEXT: s_setpc_b64 s[30:31]200;201; GFX9ARCH-GISEL-LABEL: workgroup_ids_gfx:202; GFX9ARCH-GISEL: ; %bb.0:203; GFX9ARCH-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)204; GFX9ARCH-GISEL-NEXT: v_mov_b32_e32 v6, ttmp9205; GFX9ARCH-GISEL-NEXT: s_and_b32 s34, ttmp7, 0xffff206; GFX9ARCH-GISEL-NEXT: s_lshr_b32 s35, ttmp7, 16207; GFX9ARCH-GISEL-NEXT: global_store_dword v[0:1], v6, off208; GFX9ARCH-GISEL-NEXT: s_waitcnt vmcnt(0)209; GFX9ARCH-GISEL-NEXT: v_mov_b32_e32 v0, s34210; GFX9ARCH-GISEL-NEXT: global_store_dword v[2:3], v0, off211; GFX9ARCH-GISEL-NEXT: s_waitcnt vmcnt(0)212; GFX9ARCH-GISEL-NEXT: v_mov_b32_e32 v0, s35213; GFX9ARCH-GISEL-NEXT: global_store_dword v[4:5], v0, off214; GFX9ARCH-GISEL-NEXT: s_waitcnt vmcnt(0)215; GFX9ARCH-GISEL-NEXT: s_setpc_b64 s[30:31]216;217; GFX12-LABEL: workgroup_ids_gfx:218; GFX12: ; %bb.0:219; GFX12-NEXT: s_wait_loadcnt_dscnt 0x0220; GFX12-NEXT: s_wait_expcnt 0x0221; GFX12-NEXT: s_wait_samplecnt 0x0222; GFX12-NEXT: s_wait_bvhcnt 0x0223; GFX12-NEXT: s_wait_kmcnt 0x0224; GFX12-NEXT: s_and_b32 s0, ttmp7, 0xffff225; GFX12-NEXT: s_wait_alu depctr_sa_sdst(0)226; GFX12-NEXT: v_dual_mov_b32 v6, ttmp9 :: v_dual_mov_b32 v7, s0227; GFX12-NEXT: s_lshr_b32 s1, ttmp7, 16228; GFX12-NEXT: s_wait_alu depctr_sa_sdst(0)229; GFX12-NEXT: v_mov_b32_e32 v8, s1230; GFX12-NEXT: s_wait_storecnt 0x0231; GFX12-NEXT: global_store_b32 v[0:1], v6, off scope:SCOPE_SYS232; GFX12-NEXT: s_wait_storecnt 0x0233; GFX12-NEXT: global_store_b32 v[2:3], v7, off scope:SCOPE_SYS234; GFX12-NEXT: s_wait_storecnt 0x0235; GFX12-NEXT: global_store_b32 v[4:5], v8, off scope:SCOPE_SYS236; GFX12-NEXT: s_wait_storecnt 0x0237; GFX12-NEXT: s_setpc_b64 s[30:31]238 %id.x = call i32 @llvm.amdgcn.workgroup.id.x()239 %id.y = call i32 @llvm.amdgcn.workgroup.id.y()240 %id.z = call i32 @llvm.amdgcn.workgroup.id.z()241 store volatile i32 %id.x, ptr addrspace(1) %outx242 store volatile i32 %id.y, ptr addrspace(1) %outy243 store volatile i32 %id.z, ptr addrspace(1) %outz244 ret void245}246