86 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 | FileCheck %s --check-prefix=GCN3; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 | FileCheck %s --check-prefix=GFX114 5define void @vgpr_descriptor_waterfall_loop_idom_update(ptr %arg) #0 {6; GCN-LABEL: vgpr_descriptor_waterfall_loop_idom_update:7; GCN: ; %bb.0: ; %entry8; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)9; GCN-NEXT: v_add_co_u32 v6, vcc_lo, v0, 810; GCN-NEXT: v_add_co_ci_u32_e32 v7, vcc_lo, 0, v1, vcc_lo11; GCN-NEXT: .LBB0_1: ; %bb012; GCN-NEXT: ; =>This Loop Header: Depth=113; GCN-NEXT: ; Child Loop BB0_2 Depth 214; GCN-NEXT: s_clause 0x115; GCN-NEXT: flat_load_dwordx2 v[4:5], v[6:7]16; GCN-NEXT: flat_load_dwordx2 v[2:3], v[0:1]17; GCN-NEXT: s_mov_b32 s5, exec_lo18; GCN-NEXT: .LBB0_2: ; Parent Loop BB0_1 Depth=119; GCN-NEXT: ; => This Inner Loop Header: Depth=220; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)21; GCN-NEXT: v_readfirstlane_b32 s8, v222; GCN-NEXT: v_readfirstlane_b32 s9, v323; GCN-NEXT: v_readfirstlane_b32 s10, v424; GCN-NEXT: v_readfirstlane_b32 s11, v525; GCN-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[8:9], v[2:3]26; GCN-NEXT: v_cmp_eq_u64_e64 s4, s[10:11], v[4:5]27; GCN-NEXT: s_and_b32 s4, vcc_lo, s428; GCN-NEXT: s_and_saveexec_b32 s4, s429; GCN-NEXT: buffer_store_dword v0, v0, s[8:11], 0 offen30; GCN-NEXT: ; implicit-def: $vgpr2_vgpr3_vgpr4_vgpr531; GCN-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)32; GCN-NEXT: s_xor_b32 exec_lo, exec_lo, s433; GCN-NEXT: s_cbranch_execnz .LBB0_234; GCN-NEXT: ; %bb.3: ; in Loop: Header=BB0_1 Depth=135; GCN-NEXT: s_mov_b32 exec_lo, s536; GCN-NEXT: s_mov_b32 vcc_lo, exec_lo37; GCN-NEXT: s_cbranch_vccnz .LBB0_138; GCN-NEXT: ; %bb.4: ; %DummyReturnBlock39; GCN-NEXT: s_setpc_b64 s[30:31]40;41; GFX11-LABEL: vgpr_descriptor_waterfall_loop_idom_update:42; GFX11: ; %bb.0: ; %entry43; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)44; GFX11-NEXT: .p2align 645; GFX11-NEXT: .LBB0_1: ; %bb046; GFX11-NEXT: ; =>This Loop Header: Depth=147; GFX11-NEXT: ; Child Loop BB0_2 Depth 248; GFX11-NEXT: flat_load_b128 v[2:5], v[0:1]49; GFX11-NEXT: s_mov_b32 s1, exec_lo50; GFX11-NEXT: .LBB0_2: ; Parent Loop BB0_1 Depth=151; GFX11-NEXT: ; => This Inner Loop Header: Depth=252; GFX11-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)53; GFX11-NEXT: v_readfirstlane_b32 s4, v254; GFX11-NEXT: v_readfirstlane_b32 s5, v355; GFX11-NEXT: v_readfirstlane_b32 s6, v456; GFX11-NEXT: v_readfirstlane_b32 s7, v557; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)58; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[2:3]59; GFX11-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[4:5]60; GFX11-NEXT: s_and_b32 s0, vcc_lo, s061; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)62; GFX11-NEXT: s_and_saveexec_b32 s0, s063; GFX11-NEXT: buffer_store_b32 v0, v0, s[4:7], 0 offen64; GFX11-NEXT: ; implicit-def: $vgpr2_vgpr3_vgpr4_vgpr565; GFX11-NEXT: s_xor_b32 exec_lo, exec_lo, s066; GFX11-NEXT: s_cbranch_execnz .LBB0_267; GFX11-NEXT: ; %bb.3: ; in Loop: Header=BB0_1 Depth=168; GFX11-NEXT: s_mov_b32 exec_lo, s169; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)70; GFX11-NEXT: s_mov_b32 vcc_lo, exec_lo71; GFX11-NEXT: s_cbranch_vccnz .LBB0_172; GFX11-NEXT: ; %bb.4: ; %DummyReturnBlock73; GFX11-NEXT: s_setpc_b64 s[30:31]74entry:75 br label %bb076 77bb0:78 %desc = load ptr addrspace(8), ptr %arg, align 879 tail call void @llvm.amdgcn.raw.ptr.buffer.store.f32(float poison, ptr addrspace(8) %desc, i32 poison, i32 0, i32 0)80 br label %bb081}82 83declare void @llvm.amdgcn.raw.ptr.buffer.store.f32(float, ptr addrspace(8), i32, i32, i32 immarg) #084 85attributes #0 = { nounwind memory(argmem: write) }86