362 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -mtriple=amdgcn -mcpu=gfx90a < %s | FileCheck -check-prefixes=GFX9 %s3; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX1010 %s4; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 < %s | FileCheck -check-prefixes=GFX10,GFX1030 %s5 6define void @uniform_br_no_metadata(i32 noundef inreg %value, ptr addrspace(8) nocapture writeonly inreg %res, i32 noundef inreg %v_offset, i32 noundef inreg %0, i32 noundef inreg %flag) {7; GFX9-LABEL: uniform_br_no_metadata:8; GFX9: ; %bb.0: ; %entry9; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)10; GFX9-NEXT: s_cmp_lt_i32 s23, 111; GFX9-NEXT: s_cbranch_scc1 .LBB0_212; GFX9-NEXT: ; %bb.1: ; %if.then13; GFX9-NEXT: s_mov_b32 s7, s2014; GFX9-NEXT: s_mov_b32 s6, s1915; GFX9-NEXT: s_mov_b32 s5, s1816; GFX9-NEXT: s_mov_b32 s4, s1717; GFX9-NEXT: v_mov_b32_e32 v1, s1618; GFX9-NEXT: v_mov_b32_e32 v0, s2119; GFX9-NEXT: buffer_store_dword v1, v0, s[4:7], 0 offen20; GFX9-NEXT: .LBB0_2: ; %if.end21; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)22; GFX9-NEXT: s_setpc_b64 s[30:31]23;24; GFX10-LABEL: uniform_br_no_metadata:25; GFX10: ; %bb.0: ; %entry26; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)27; GFX10-NEXT: s_cmp_lt_i32 s23, 128; GFX10-NEXT: s_cbranch_scc1 .LBB0_229; GFX10-NEXT: ; %bb.1: ; %if.then30; GFX10-NEXT: v_mov_b32_e32 v0, s1631; GFX10-NEXT: v_mov_b32_e32 v1, s2132; GFX10-NEXT: s_mov_b32 s7, s2033; GFX10-NEXT: s_mov_b32 s6, s1934; GFX10-NEXT: s_mov_b32 s5, s1835; GFX10-NEXT: s_mov_b32 s4, s1736; GFX10-NEXT: buffer_store_dword v0, v1, s[4:7], 0 offen37; GFX10-NEXT: .LBB0_2: ; %if.end38; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)39; GFX10-NEXT: s_setpc_b64 s[30:31]40entry:41 %cmp = icmp sgt i32 %flag, 042 br i1 %cmp, label %if.then, label %if.end43 44if.then:45 tail call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %value, ptr addrspace(8) %res, i32 %v_offset, i32 0, i32 0)46 br label %if.end47 48if.end:49 call void @llvm.amdgcn.s.waitcnt(i32 0)50 ret void51}52 53define void @uniform_br_same_weight(i32 noundef inreg %value, ptr addrspace(8) nocapture writeonly inreg %res, i32 noundef inreg %v_offset, i32 noundef inreg %0, i32 noundef inreg %flag) {54; GFX9-LABEL: uniform_br_same_weight:55; GFX9: ; %bb.0: ; %entry56; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)57; GFX9-NEXT: s_cmp_lt_i32 s23, 158; GFX9-NEXT: s_cbranch_scc1 .LBB1_259; GFX9-NEXT: ; %bb.1: ; %if.then60; GFX9-NEXT: s_mov_b32 s7, s2061; GFX9-NEXT: s_mov_b32 s6, s1962; GFX9-NEXT: s_mov_b32 s5, s1863; GFX9-NEXT: s_mov_b32 s4, s1764; GFX9-NEXT: v_mov_b32_e32 v1, s1665; GFX9-NEXT: v_mov_b32_e32 v0, s2166; GFX9-NEXT: buffer_store_dword v1, v0, s[4:7], 0 offen67; GFX9-NEXT: .LBB1_2: ; %if.end68; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)69; GFX9-NEXT: s_setpc_b64 s[30:31]70;71; GFX10-LABEL: uniform_br_same_weight:72; GFX10: ; %bb.0: ; %entry73; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)74; GFX10-NEXT: s_cmp_lt_i32 s23, 175; GFX10-NEXT: s_cbranch_scc1 .LBB1_276; GFX10-NEXT: ; %bb.1: ; %if.then77; GFX10-NEXT: v_mov_b32_e32 v0, s1678; GFX10-NEXT: v_mov_b32_e32 v1, s2179; GFX10-NEXT: s_mov_b32 s7, s2080; GFX10-NEXT: s_mov_b32 s6, s1981; GFX10-NEXT: s_mov_b32 s5, s1882; GFX10-NEXT: s_mov_b32 s4, s1783; GFX10-NEXT: buffer_store_dword v0, v1, s[4:7], 0 offen84; GFX10-NEXT: .LBB1_2: ; %if.end85; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)86; GFX10-NEXT: s_setpc_b64 s[30:31]87entry:88 %cmp = icmp sgt i32 %flag, 089 br i1 %cmp, label %if.then, label %if.end, !prof !090 91if.then:92 tail call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %value, ptr addrspace(8) %res, i32 %v_offset, i32 0, i32 0)93 br label %if.end94 95if.end:96 call void @llvm.amdgcn.s.waitcnt(i32 0)97 ret void98}99 100define void @uniform_br_then_likely(i32 noundef inreg %value, ptr addrspace(8) nocapture writeonly inreg %res, i32 noundef inreg %v_offset, i32 noundef inreg %0, i32 noundef inreg %flag) {101; GFX9-LABEL: uniform_br_then_likely:102; GFX9: ; %bb.0: ; %entry103; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)104; GFX9-NEXT: s_cmp_lt_i32 s23, 1105; GFX9-NEXT: s_cbranch_scc1 .LBB2_2106; GFX9-NEXT: ; %bb.1: ; %if.then107; GFX9-NEXT: s_mov_b32 s7, s20108; GFX9-NEXT: s_mov_b32 s6, s19109; GFX9-NEXT: s_mov_b32 s5, s18110; GFX9-NEXT: s_mov_b32 s4, s17111; GFX9-NEXT: v_mov_b32_e32 v1, s16112; GFX9-NEXT: v_mov_b32_e32 v0, s21113; GFX9-NEXT: buffer_store_dword v1, v0, s[4:7], 0 offen114; GFX9-NEXT: .LBB2_2: ; %if.end115; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)116; GFX9-NEXT: s_setpc_b64 s[30:31]117;118; GFX10-LABEL: uniform_br_then_likely:119; GFX10: ; %bb.0: ; %entry120; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)121; GFX10-NEXT: s_cmp_lt_i32 s23, 1122; GFX10-NEXT: s_cbranch_scc1 .LBB2_2123; GFX10-NEXT: ; %bb.1: ; %if.then124; GFX10-NEXT: v_mov_b32_e32 v0, s16125; GFX10-NEXT: v_mov_b32_e32 v1, s21126; GFX10-NEXT: s_mov_b32 s7, s20127; GFX10-NEXT: s_mov_b32 s6, s19128; GFX10-NEXT: s_mov_b32 s5, s18129; GFX10-NEXT: s_mov_b32 s4, s17130; GFX10-NEXT: buffer_store_dword v0, v1, s[4:7], 0 offen131; GFX10-NEXT: .LBB2_2: ; %if.end132; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)133; GFX10-NEXT: s_setpc_b64 s[30:31]134entry:135 %cmp = icmp sgt i32 %flag, 0136 br i1 %cmp, label %if.then, label %if.end, !prof !1137 138if.then:139 tail call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %value, ptr addrspace(8) %res, i32 %v_offset, i32 0, i32 0)140 br label %if.end141 142if.end:143 call void @llvm.amdgcn.s.waitcnt(i32 0)144 ret void145}146 147define void @divergent_br_no_metadata(i32 noundef inreg %value, ptr addrspace(8) nocapture writeonly inreg %res, i32 noundef inreg %v_offset, i32 noundef inreg %0, i32 noundef %flag) {148; GFX9-LABEL: divergent_br_no_metadata:149; GFX9: ; %bb.0: ; %entry150; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)151; GFX9-NEXT: v_cmp_lt_i32_e32 vcc, 0, v0152; GFX9-NEXT: s_and_saveexec_b64 s[8:9], vcc153; GFX9-NEXT: s_cbranch_execz .LBB3_2154; GFX9-NEXT: ; %bb.1: ; %if.then155; GFX9-NEXT: s_mov_b32 s7, s20156; GFX9-NEXT: s_mov_b32 s6, s19157; GFX9-NEXT: s_mov_b32 s5, s18158; GFX9-NEXT: s_mov_b32 s4, s17159; GFX9-NEXT: v_mov_b32_e32 v1, s16160; GFX9-NEXT: v_mov_b32_e32 v0, s21161; GFX9-NEXT: buffer_store_dword v1, v0, s[4:7], 0 offen162; GFX9-NEXT: .LBB3_2: ; %if.end163; GFX9-NEXT: s_or_b64 exec, exec, s[8:9]164; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)165; GFX9-NEXT: s_setpc_b64 s[30:31]166;167; GFX1010-LABEL: divergent_br_no_metadata:168; GFX1010: ; %bb.0: ; %entry169; GFX1010-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)170; GFX1010-NEXT: v_cmp_lt_i32_e32 vcc_lo, 0, v0171; GFX1010-NEXT: s_and_saveexec_b32 s8, vcc_lo172; GFX1010-NEXT: s_cbranch_execz .LBB3_2173; GFX1010-NEXT: ; %bb.1: ; %if.then174; GFX1010-NEXT: v_mov_b32_e32 v0, s16175; GFX1010-NEXT: v_mov_b32_e32 v1, s21176; GFX1010-NEXT: s_mov_b32 s7, s20177; GFX1010-NEXT: s_mov_b32 s6, s19178; GFX1010-NEXT: s_mov_b32 s5, s18179; GFX1010-NEXT: s_mov_b32 s4, s17180; GFX1010-NEXT: buffer_store_dword v0, v1, s[4:7], 0 offen181; GFX1010-NEXT: .LBB3_2: ; %if.end182; GFX1010-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)183; GFX1010-NEXT: s_or_b32 exec_lo, exec_lo, s8184; GFX1010-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)185; GFX1010-NEXT: s_setpc_b64 s[30:31]186;187; GFX1030-LABEL: divergent_br_no_metadata:188; GFX1030: ; %bb.0: ; %entry189; GFX1030-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)190; GFX1030-NEXT: s_mov_b32 s8, exec_lo191; GFX1030-NEXT: v_cmpx_lt_i32_e32 0, v0192; GFX1030-NEXT: s_cbranch_execz .LBB3_2193; GFX1030-NEXT: ; %bb.1: ; %if.then194; GFX1030-NEXT: v_mov_b32_e32 v0, s16195; GFX1030-NEXT: v_mov_b32_e32 v1, s21196; GFX1030-NEXT: s_mov_b32 s7, s20197; GFX1030-NEXT: s_mov_b32 s6, s19198; GFX1030-NEXT: s_mov_b32 s5, s18199; GFX1030-NEXT: s_mov_b32 s4, s17200; GFX1030-NEXT: buffer_store_dword v0, v1, s[4:7], 0 offen201; GFX1030-NEXT: .LBB3_2: ; %if.end202; GFX1030-NEXT: s_or_b32 exec_lo, exec_lo, s8203; GFX1030-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)204; GFX1030-NEXT: s_setpc_b64 s[30:31]205entry:206 %cmp = icmp sgt i32 %flag, 0207 br i1 %cmp, label %if.then, label %if.end208 209if.then:210 tail call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %value, ptr addrspace(8) %res, i32 %v_offset, i32 0, i32 0)211 br label %if.end212 213if.end:214 call void @llvm.amdgcn.s.waitcnt(i32 0)215 ret void216}217 218define void @divergent_br_same_weight(i32 noundef inreg %value, ptr addrspace(8) nocapture writeonly inreg %res, i32 noundef inreg %v_offset, i32 noundef inreg %0, i32 noundef %flag) {219; GFX9-LABEL: divergent_br_same_weight:220; GFX9: ; %bb.0: ; %entry221; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)222; GFX9-NEXT: v_cmp_lt_i32_e32 vcc, 0, v0223; GFX9-NEXT: s_and_saveexec_b64 s[8:9], vcc224; GFX9-NEXT: s_cbranch_execz .LBB4_2225; GFX9-NEXT: ; %bb.1: ; %if.then226; GFX9-NEXT: s_mov_b32 s7, s20227; GFX9-NEXT: s_mov_b32 s6, s19228; GFX9-NEXT: s_mov_b32 s5, s18229; GFX9-NEXT: s_mov_b32 s4, s17230; GFX9-NEXT: v_mov_b32_e32 v1, s16231; GFX9-NEXT: v_mov_b32_e32 v0, s21232; GFX9-NEXT: buffer_store_dword v1, v0, s[4:7], 0 offen233; GFX9-NEXT: .LBB4_2: ; %if.end234; GFX9-NEXT: s_or_b64 exec, exec, s[8:9]235; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)236; GFX9-NEXT: s_setpc_b64 s[30:31]237;238; GFX1010-LABEL: divergent_br_same_weight:239; GFX1010: ; %bb.0: ; %entry240; GFX1010-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)241; GFX1010-NEXT: v_cmp_lt_i32_e32 vcc_lo, 0, v0242; GFX1010-NEXT: s_and_saveexec_b32 s8, vcc_lo243; GFX1010-NEXT: s_cbranch_execz .LBB4_2244; GFX1010-NEXT: ; %bb.1: ; %if.then245; GFX1010-NEXT: v_mov_b32_e32 v0, s16246; GFX1010-NEXT: v_mov_b32_e32 v1, s21247; GFX1010-NEXT: s_mov_b32 s7, s20248; GFX1010-NEXT: s_mov_b32 s6, s19249; GFX1010-NEXT: s_mov_b32 s5, s18250; GFX1010-NEXT: s_mov_b32 s4, s17251; GFX1010-NEXT: buffer_store_dword v0, v1, s[4:7], 0 offen252; GFX1010-NEXT: .LBB4_2: ; %if.end253; GFX1010-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)254; GFX1010-NEXT: s_or_b32 exec_lo, exec_lo, s8255; GFX1010-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)256; GFX1010-NEXT: s_setpc_b64 s[30:31]257;258; GFX1030-LABEL: divergent_br_same_weight:259; GFX1030: ; %bb.0: ; %entry260; GFX1030-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)261; GFX1030-NEXT: s_mov_b32 s8, exec_lo262; GFX1030-NEXT: v_cmpx_lt_i32_e32 0, v0263; GFX1030-NEXT: s_cbranch_execz .LBB4_2264; GFX1030-NEXT: ; %bb.1: ; %if.then265; GFX1030-NEXT: v_mov_b32_e32 v0, s16266; GFX1030-NEXT: v_mov_b32_e32 v1, s21267; GFX1030-NEXT: s_mov_b32 s7, s20268; GFX1030-NEXT: s_mov_b32 s6, s19269; GFX1030-NEXT: s_mov_b32 s5, s18270; GFX1030-NEXT: s_mov_b32 s4, s17271; GFX1030-NEXT: buffer_store_dword v0, v1, s[4:7], 0 offen272; GFX1030-NEXT: .LBB4_2: ; %if.end273; GFX1030-NEXT: s_or_b32 exec_lo, exec_lo, s8274; GFX1030-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)275; GFX1030-NEXT: s_setpc_b64 s[30:31]276entry:277 %cmp = icmp sgt i32 %flag, 0278 br i1 %cmp, label %if.then, label %if.end, !prof !0279 280if.then:281 tail call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %value, ptr addrspace(8) %res, i32 %v_offset, i32 0, i32 0)282 br label %if.end283 284if.end:285 call void @llvm.amdgcn.s.waitcnt(i32 0)286 ret void287}288 289define void @divergent_br_then_likely(i32 noundef inreg %value, ptr addrspace(8) nocapture writeonly inreg %res, i32 noundef inreg %v_offset, i32 noundef inreg %0, i32 noundef %flag) {290; GFX9-LABEL: divergent_br_then_likely:291; GFX9: ; %bb.0: ; %entry292; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)293; GFX9-NEXT: v_cmp_lt_i32_e32 vcc, 0, v0294; GFX9-NEXT: s_and_saveexec_b64 s[8:9], vcc295; GFX9-NEXT: ; %bb.1: ; %if.then296; GFX9-NEXT: s_mov_b32 s7, s20297; GFX9-NEXT: s_mov_b32 s6, s19298; GFX9-NEXT: s_mov_b32 s5, s18299; GFX9-NEXT: s_mov_b32 s4, s17300; GFX9-NEXT: v_mov_b32_e32 v1, s16301; GFX9-NEXT: v_mov_b32_e32 v0, s21302; GFX9-NEXT: buffer_store_dword v1, v0, s[4:7], 0 offen303; GFX9-NEXT: ; %bb.2: ; %if.end304; GFX9-NEXT: s_or_b64 exec, exec, s[8:9]305; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)306; GFX9-NEXT: s_setpc_b64 s[30:31]307;308; GFX1010-LABEL: divergent_br_then_likely:309; GFX1010: ; %bb.0: ; %entry310; GFX1010-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)311; GFX1010-NEXT: v_cmp_lt_i32_e32 vcc_lo, 0, v0312; GFX1010-NEXT: s_and_saveexec_b32 s8, vcc_lo313; GFX1010-NEXT: ; %bb.1: ; %if.then314; GFX1010-NEXT: v_mov_b32_e32 v0, s16315; GFX1010-NEXT: v_mov_b32_e32 v1, s21316; GFX1010-NEXT: s_mov_b32 s7, s20317; GFX1010-NEXT: s_mov_b32 s6, s19318; GFX1010-NEXT: s_mov_b32 s5, s18319; GFX1010-NEXT: s_mov_b32 s4, s17320; GFX1010-NEXT: buffer_store_dword v0, v1, s[4:7], 0 offen321; GFX1010-NEXT: ; %bb.2: ; %if.end322; GFX1010-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)323; GFX1010-NEXT: s_or_b32 exec_lo, exec_lo, s8324; GFX1010-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)325; GFX1010-NEXT: s_setpc_b64 s[30:31]326;327; GFX1030-LABEL: divergent_br_then_likely:328; GFX1030: ; %bb.0: ; %entry329; GFX1030-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)330; GFX1030-NEXT: s_mov_b32 s8, exec_lo331; GFX1030-NEXT: v_cmpx_lt_i32_e32 0, v0332; GFX1030-NEXT: ; %bb.1: ; %if.then333; GFX1030-NEXT: v_mov_b32_e32 v0, s16334; GFX1030-NEXT: v_mov_b32_e32 v1, s21335; GFX1030-NEXT: s_mov_b32 s7, s20336; GFX1030-NEXT: s_mov_b32 s6, s19337; GFX1030-NEXT: s_mov_b32 s5, s18338; GFX1030-NEXT: s_mov_b32 s4, s17339; GFX1030-NEXT: buffer_store_dword v0, v1, s[4:7], 0 offen340; GFX1030-NEXT: ; %bb.2: ; %if.end341; GFX1030-NEXT: s_or_b32 exec_lo, exec_lo, s8342; GFX1030-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)343; GFX1030-NEXT: s_setpc_b64 s[30:31]344entry:345 %cmp = icmp sgt i32 %flag, 0346 br i1 %cmp, label %if.then, label %if.end, !prof !1347 348if.then:349 tail call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %value, ptr addrspace(8) %res, i32 %v_offset, i32 0, i32 0)350 br label %if.end351 352if.end:353 call void @llvm.amdgcn.s.waitcnt(i32 0)354 ret void355}356 357declare void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32, ptr addrspace(8) nocapture writeonly, i32, i32, i32 immarg)358declare void @llvm.amdgcn.s.waitcnt(i32)359 360!0 = !{!"branch_weights", i32 1000, i32 1000}361!1 = !{!"branch_weights", i32 2000, i32 1}362