115 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 32; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1031 < %s | FileCheck %s3 4; A VGPR loop variable was incorrectly sunk into a flow block, past5; the si_end_cf reconvergence point.6 7define void @machinesink_loop_variable_out_of_divergent_loop(i32 %arg, i1 %cmp49280.not, i32 %arg1, i1 %cmp108) {8; CHECK-LABEL: machinesink_loop_variable_out_of_divergent_loop:9; CHECK: ; %bb.0: ; %entry10; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)11; CHECK-NEXT: v_and_b32_e32 v1, 1, v112; CHECK-NEXT: v_and_b32_e32 v3, 1, v313; CHECK-NEXT: s_mov_b32 s6, 014; CHECK-NEXT: v_cmp_ne_u32_e64 s4, 1, v115; CHECK-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v316; CHECK-NEXT: s_inst_prefetch 0x117; CHECK-NEXT: s_branch .LBB0_318; CHECK-NEXT: .p2align 619; CHECK-NEXT: .LBB0_1: ; %Flow20; CHECK-NEXT: ; in Loop: Header=BB0_3 Depth=121; CHECK-NEXT: s_or_b32 exec_lo, exec_lo, s822; CHECK-NEXT: v_add_nc_u32_e32 v3, -4, v323; CHECK-NEXT: .LBB0_2: ; %Flow124; CHECK-NEXT: ; in Loop: Header=BB0_3 Depth=125; CHECK-NEXT: s_or_b32 exec_lo, exec_lo, s726; CHECK-NEXT: v_cmp_ne_u32_e64 s5, 0, v127; CHECK-NEXT: ;;#ASMSTART28; CHECK-NEXT: ; j lastloop entry29; CHECK-NEXT: ;;#ASMEND30; CHECK-NEXT: s_or_b32 s6, s5, s631; CHECK-NEXT: s_andn2_b32 exec_lo, exec_lo, s632; CHECK-NEXT: s_cbranch_execz .LBB0_833; CHECK-NEXT: .LBB0_3: ; %for.body3334; CHECK-NEXT: ; =>This Loop Header: Depth=135; CHECK-NEXT: ; Child Loop BB0_6 Depth 236; CHECK-NEXT: v_mov_b32_e32 v3, 037; CHECK-NEXT: v_mov_b32_e32 v1, 038; CHECK-NEXT: s_and_saveexec_b32 s7, s439; CHECK-NEXT: s_cbranch_execz .LBB0_240; CHECK-NEXT: ; %bb.4: ; %for.body51.preheader41; CHECK-NEXT: ; in Loop: Header=BB0_3 Depth=142; CHECK-NEXT: s_mov_b32 s8, 043; CHECK-NEXT: s_mov_b32 s9, 044; CHECK-NEXT: s_branch .LBB0_645; CHECK-NEXT: .p2align 646; CHECK-NEXT: .LBB0_5: ; %if.end11847; CHECK-NEXT: ; in Loop: Header=BB0_6 Depth=248; CHECK-NEXT: s_or_b32 exec_lo, exec_lo, s549; CHECK-NEXT: s_add_i32 s9, s9, 450; CHECK-NEXT: ;;#ASMSTART51; CHECK-NEXT: ; backedge52; CHECK-NEXT: ;;#ASMEND53; CHECK-NEXT: v_add_nc_u32_e32 v3, s9, v254; CHECK-NEXT: v_cmp_ge_u32_e64 s5, v3, v055; CHECK-NEXT: s_or_b32 s8, s5, s856; CHECK-NEXT: s_andn2_b32 exec_lo, exec_lo, s857; CHECK-NEXT: s_cbranch_execz .LBB0_158; CHECK-NEXT: .LBB0_6: ; %for.body5159; CHECK-NEXT: ; Parent Loop BB0_3 Depth=160; CHECK-NEXT: ; => This Inner Loop Header: Depth=261; CHECK-NEXT: v_mov_b32_e32 v1, 162; CHECK-NEXT: s_and_saveexec_b32 s5, vcc_lo63; CHECK-NEXT: s_cbranch_execz .LBB0_564; CHECK-NEXT: ; %bb.7: ; %if.then11265; CHECK-NEXT: ; in Loop: Header=BB0_6 Depth=266; CHECK-NEXT: s_add_i32 s10, s9, 467; CHECK-NEXT: v_mov_b32_e32 v1, 068; CHECK-NEXT: v_mov_b32_e32 v3, s1069; CHECK-NEXT: ds_write_b32 v1, v370; CHECK-NEXT: s_branch .LBB0_571; CHECK-NEXT: .LBB0_8: ; %for.body159.preheader72; CHECK-NEXT: s_inst_prefetch 0x273; CHECK-NEXT: s_or_b32 exec_lo, exec_lo, s674; CHECK-NEXT: s_mov_b32 vcc_lo, exec_lo75; CHECK-NEXT: .LBB0_9: ; %for.body15976; CHECK-NEXT: ; =>This Inner Loop Header: Depth=177; CHECK-NEXT: s_cbranch_vccnz .LBB0_978; CHECK-NEXT: ; %bb.10: ; %DummyReturnBlock79; CHECK-NEXT: s_waitcnt lgkmcnt(0)80; CHECK-NEXT: s_setpc_b64 s[30:31]81entry:82 br label %for.body3383 84for.body33: ; preds = %for.end121, %entry85 br i1 %cmp49280.not, label %for.end121, label %for.body5186 87for.body51: ; preds = %if.end118, %for.body3388 %add48284 = phi i32 [ %add48, %if.end118 ], [ %arg1, %for.body33 ]89 %collision.0281 = phi i32 [ %inc119, %if.end118 ], [ 1, %for.body33 ]90 br i1 %cmp108, label %if.then112, label %if.end11891 92if.then112: ; preds = %for.body5193 %inc101 = add i32 %collision.0281, 394 store i32 %inc101, ptr addrspace(3) null, align 214748364895 br label %if.end11896 97if.end118: ; preds = %if.then112, %for.body5198 %thCollNum.5 = phi i32 [ 0, %if.then112 ], [ 1, %for.body51 ]99 %inc119 = add i32 %collision.0281, 4100 tail call void asm sideeffect "; backedge", ""()101 %add48 = add i32 %add48284, 4102 %cmp49 = icmp ult i32 %add48, %arg103 br i1 %cmp49, label %for.body51, label %for.end121104 105for.end121: ; preds = %if.end118, %for.body33106 %thCollNum.1.lcssa = phi i32 [ 0, %for.body33 ], [ %thCollNum.5, %if.end118 ]107 %j.0.lcssa = phi i32 [ 0, %for.body33 ], [ %add48284, %if.end118 ]108 %i5 = tail call i32 asm sideeffect "; j lastloop entry", "=v,0"(i32 %j.0.lcssa)109 %cmp31 = icmp eq i32 %thCollNum.1.lcssa, 0110 br i1 %cmp31, label %for.body33, label %for.body159111 112for.body159: ; preds = %for.body159, %for.end121113 br label %for.body159114}115