brintos

brintos / llvm-project-archived public Read only

0
0
Text · 11.1 KiB · 43752c2 Raw
241 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 22; RUN: llc -O3 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck %s3 4%"struct.__llvm_libc::rpc::Buffer" = type { [8 x i64] }5 6define void @issue63986(i64 %0, i64 %idxprom, ptr inreg %ptr) {7; CHECK-LABEL: issue63986:8; CHECK:       ; %bb.0: ; %entry9; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)10; CHECK-NEXT:    v_lshlrev_b64 v[4:5], 6, v[2:3]11; CHECK-NEXT:    v_mov_b32_e32 v6, s1712; CHECK-NEXT:    v_add_co_u32_e32 v8, vcc, s16, v413; CHECK-NEXT:    v_addc_co_u32_e32 v9, vcc, v6, v5, vcc14; CHECK-NEXT:    s_mov_b64 s[4:5], 015; CHECK-NEXT:  .LBB0_1: ; %loop-memcpy-expansion16; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=117; CHECK-NEXT:    v_mov_b32_e32 v7, s518; CHECK-NEXT:    v_mov_b32_e32 v6, s419; CHECK-NEXT:    flat_load_dwordx4 v[10:13], v[6:7]20; CHECK-NEXT:    v_add_co_u32_e32 v6, vcc, s4, v821; CHECK-NEXT:    s_add_u32 s4, s4, 1622; CHECK-NEXT:    s_addc_u32 s5, s5, 023; CHECK-NEXT:    v_cmp_ge_u64_e64 s[6:7], s[4:5], 3224; CHECK-NEXT:    v_addc_co_u32_e32 v7, vcc, v9, v7, vcc25; CHECK-NEXT:    s_and_b64 vcc, exec, s[6:7]26; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)27; CHECK-NEXT:    flat_store_dwordx4 v[6:7], v[10:13]28; CHECK-NEXT:    s_cbranch_vccz .LBB0_129; CHECK-NEXT:  ; %bb.2: ; %loop-memcpy-residual-header30; CHECK-NEXT:    s_branch .LBB0_431; CHECK-NEXT:  ; %bb.3:32; CHECK-NEXT:    ; implicit-def: $vgpr6_vgpr733; CHECK-NEXT:    s_branch .LBB0_534; CHECK-NEXT:  .LBB0_4: ; %loop-memcpy-residual-header.post-loop-memcpy-expansion_crit_edge35; CHECK-NEXT:    v_lshlrev_b64 v[6:7], 6, v[2:3]36; CHECK-NEXT:    s_cbranch_execnz .LBB0_837; CHECK-NEXT:  .LBB0_5: ; %loop-memcpy-residual.preheader38; CHECK-NEXT:    s_add_u32 s4, s16, 3239; CHECK-NEXT:    s_addc_u32 s5, s17, 040; CHECK-NEXT:    v_mov_b32_e32 v3, s541; CHECK-NEXT:    v_add_co_u32_e32 v2, vcc, s4, v442; CHECK-NEXT:    v_addc_co_u32_e32 v3, vcc, v3, v5, vcc43; CHECK-NEXT:    s_mov_b64 s[4:5], 044; CHECK-NEXT:  ; %bb.6: ; %loop-memcpy-residual45; CHECK-NEXT:    s_add_u32 s6, 32, s446; CHECK-NEXT:    s_addc_u32 s7, 0, s547; CHECK-NEXT:    v_mov_b32_e32 v6, s648; CHECK-NEXT:    v_mov_b32_e32 v7, s749; CHECK-NEXT:    flat_load_ubyte v10, v[6:7]50; CHECK-NEXT:    v_mov_b32_e32 v7, s551; CHECK-NEXT:    v_add_co_u32_e32 v6, vcc, s4, v252; CHECK-NEXT:    v_addc_co_u32_e32 v7, vcc, v3, v7, vcc53; CHECK-NEXT:    s_add_u32 s4, s4, 154; CHECK-NEXT:    s_addc_u32 s5, 0, s555; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)56; CHECK-NEXT:    flat_store_byte v[6:7], v1057; CHECK-NEXT:  ; %bb.7:58; CHECK-NEXT:    v_mov_b32_e32 v7, v559; CHECK-NEXT:    v_mov_b32_e32 v6, v460; CHECK-NEXT:  .LBB0_8: ; %post-loop-memcpy-expansion61; CHECK-NEXT:    v_and_b32_e32 v2, 15, v062; CHECK-NEXT:    v_and_b32_e32 v0, -16, v063; CHECK-NEXT:    v_add_co_u32_e32 v4, vcc, v6, v064; CHECK-NEXT:    v_mov_b32_e32 v3, 065; CHECK-NEXT:    v_addc_co_u32_e32 v5, vcc, v7, v1, vcc66; CHECK-NEXT:    v_cmp_ne_u64_e64 s[4:5], 0, v[0:1]67; CHECK-NEXT:    v_cmp_ne_u64_e64 s[6:7], 0, v[2:3]68; CHECK-NEXT:    v_mov_b32_e32 v6, s1769; CHECK-NEXT:    v_add_co_u32_e32 v4, vcc, s16, v470; CHECK-NEXT:    v_addc_co_u32_e32 v5, vcc, v6, v5, vcc71; CHECK-NEXT:    s_branch .LBB0_1172; CHECK-NEXT:  .LBB0_9: ; %Flow1473; CHECK-NEXT:    ; in Loop: Header=BB0_11 Depth=174; CHECK-NEXT:    s_or_b64 exec, exec, s[10:11]75; CHECK-NEXT:    s_mov_b64 s[8:9], 076; CHECK-NEXT:  .LBB0_10: ; %Flow1677; CHECK-NEXT:    ; in Loop: Header=BB0_11 Depth=178; CHECK-NEXT:    s_andn2_b64 vcc, exec, s[8:9]79; CHECK-NEXT:    s_cbranch_vccz .LBB0_1980; CHECK-NEXT:  .LBB0_11: ; %while.cond81; CHECK-NEXT:    ; =>This Loop Header: Depth=182; CHECK-NEXT:    ; Child Loop BB0_13 Depth 283; CHECK-NEXT:    ; Child Loop BB0_17 Depth 284; CHECK-NEXT:    s_and_saveexec_b64 s[8:9], s[4:5]85; CHECK-NEXT:    s_cbranch_execz .LBB0_1486; CHECK-NEXT:  ; %bb.12: ; %loop-memcpy-expansion2.preheader87; CHECK-NEXT:    ; in Loop: Header=BB0_11 Depth=188; CHECK-NEXT:    s_mov_b64 s[10:11], 089; CHECK-NEXT:    s_mov_b64 s[12:13], 090; CHECK-NEXT:  .LBB0_13: ; %loop-memcpy-expansion291; CHECK-NEXT:    ; Parent Loop BB0_11 Depth=192; CHECK-NEXT:    ; => This Inner Loop Header: Depth=293; CHECK-NEXT:    v_mov_b32_e32 v6, s1094; CHECK-NEXT:    v_mov_b32_e32 v7, s1195; CHECK-NEXT:    flat_load_dwordx4 v[10:13], v[6:7]96; CHECK-NEXT:    v_add_co_u32_e32 v6, vcc, s10, v897; CHECK-NEXT:    s_add_u32 s10, s10, 1698; CHECK-NEXT:    v_addc_co_u32_e32 v7, vcc, v9, v7, vcc99; CHECK-NEXT:    s_addc_u32 s11, s11, 0100; CHECK-NEXT:    v_cmp_ge_u64_e32 vcc, s[10:11], v[0:1]101; CHECK-NEXT:    s_or_b64 s[12:13], vcc, s[12:13]102; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)103; CHECK-NEXT:    flat_store_dwordx4 v[6:7], v[10:13]104; CHECK-NEXT:    s_andn2_b64 exec, exec, s[12:13]105; CHECK-NEXT:    s_cbranch_execnz .LBB0_13106; CHECK-NEXT:  .LBB0_14: ; %Flow15107; CHECK-NEXT:    ; in Loop: Header=BB0_11 Depth=1108; CHECK-NEXT:    s_or_b64 exec, exec, s[8:9]109; CHECK-NEXT:    s_mov_b64 s[8:9], -1110; CHECK-NEXT:    s_cbranch_execz .LBB0_10111; CHECK-NEXT:  ; %bb.15: ; %loop-memcpy-residual-header5112; CHECK-NEXT:    ; in Loop: Header=BB0_11 Depth=1113; CHECK-NEXT:    s_and_saveexec_b64 s[8:9], s[6:7]114; CHECK-NEXT:    s_xor_b64 s[10:11], exec, s[8:9]115; CHECK-NEXT:    s_cbranch_execz .LBB0_9116; CHECK-NEXT:  ; %bb.16: ; %loop-memcpy-residual4.preheader117; CHECK-NEXT:    ; in Loop: Header=BB0_11 Depth=1118; CHECK-NEXT:    s_mov_b64 s[14:15], 0119; CHECK-NEXT:    s_mov_b64 s[12:13], 0120; CHECK-NEXT:  .LBB0_17: ; %loop-memcpy-residual4121; CHECK-NEXT:    ; Parent Loop BB0_11 Depth=1122; CHECK-NEXT:    ; => This Inner Loop Header: Depth=2123; CHECK-NEXT:    v_mov_b32_e32 v10, s15124; CHECK-NEXT:    v_add_co_u32_e32 v6, vcc, s14, v0125; CHECK-NEXT:    v_addc_co_u32_e32 v7, vcc, v1, v10, vcc126; CHECK-NEXT:    flat_load_ubyte v11, v[6:7]127; CHECK-NEXT:    v_add_co_u32_e32 v6, vcc, s14, v4128; CHECK-NEXT:    s_add_u32 s14, s14, 1129; CHECK-NEXT:    s_addc_u32 s15, s15, 0130; CHECK-NEXT:    v_cmp_ge_u64_e64 s[8:9], s[14:15], v[2:3]131; CHECK-NEXT:    v_addc_co_u32_e32 v7, vcc, v5, v10, vcc132; CHECK-NEXT:    s_or_b64 s[12:13], s[8:9], s[12:13]133; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)134; CHECK-NEXT:    flat_store_byte v[6:7], v11135; CHECK-NEXT:    s_andn2_b64 exec, exec, s[12:13]136; CHECK-NEXT:    s_cbranch_execnz .LBB0_17137; CHECK-NEXT:  ; %bb.18: ; %Flow138; CHECK-NEXT:    ; in Loop: Header=BB0_11 Depth=1139; CHECK-NEXT:    s_or_b64 exec, exec, s[12:13]140; CHECK-NEXT:    s_branch .LBB0_9141; CHECK-NEXT:  .LBB0_19: ; %DummyReturnBlock142; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)143; CHECK-NEXT:    s_setpc_b64 s[30:31]144entry:145  %arrayidx = getelementptr [32 x %"struct.__llvm_libc::rpc::Buffer"], ptr %ptr, i64 0, i64 %idxprom146  %spec.select = tail call i64 @llvm.umin.i64(i64 sub (i64 ptrtoint (ptr addrspacecast (ptr addrspace(4) inttoptr (i64 32 to ptr addrspace(4)) to ptr) to i64), i64 ptrtoint (ptr addrspacecast (ptr addrspace(4) null to ptr) to i64)), i64 56)147  tail call void @llvm.memcpy.p0.p0.i64(ptr %arrayidx, ptr null, i64 %spec.select, i1 false)148  br label %while.cond149 150while.cond:                         ; preds = %while.cond151  tail call void @llvm.memcpy.p0.p0.i64(ptr %arrayidx, ptr null, i64 %0, i1 false)152  br label %while.cond153}154 155define void @issue63986_reduced_expanded(i64 %idxprom) {156; CHECK-LABEL: issue63986_reduced_expanded:157; CHECK:       ; %bb.0: ; %entry158; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)159; CHECK-NEXT:  ; %bb.1: ; %loop-memcpy-expansion.preheader160; CHECK-NEXT:    s_setpc_b64 s[30:31]161; CHECK-NEXT:  ; %bb.2: ; %loop-memcpy-residual-header162; CHECK-NEXT:    s_and_b32 s4, 32, 15163; CHECK-NEXT:    s_mov_b32 s5, 0164; CHECK-NEXT:    s_cbranch_scc0 .LBB1_4165; CHECK-NEXT:  ; %bb.3:166; CHECK-NEXT:    ; implicit-def: $vgpr0_vgpr1167; CHECK-NEXT:    s_branch .LBB1_5168; CHECK-NEXT:  .LBB1_4: ; %loop-memcpy-residual-header.post-loop-memcpy-expansion_crit_edge169; CHECK-NEXT:    v_lshlrev_b64 v[0:1], 1, v[0:1]170; CHECK-NEXT:    s_cbranch_execnz .LBB1_8171; CHECK-NEXT:  .LBB1_5: ; %loop-memcpy-residual.preheader172; CHECK-NEXT:    v_mov_b32_e32 v0, s4173; CHECK-NEXT:    s_mov_b64 s[8:9], 0174; CHECK-NEXT:    s_mov_b32 s7, 0175; CHECK-NEXT:    v_mov_b32_e32 v1, s5176; CHECK-NEXT:  .LBB1_6: ; %loop-memcpy-residual177; CHECK-NEXT:    s_add_i32 s6, s8, 1178; CHECK-NEXT:    v_cmp_lt_u64_e32 vcc, s[6:7], v[0:1]179; CHECK-NEXT:    s_mov_b64 s[8:9], 1180; CHECK-NEXT:    s_cbranch_vccnz .LBB1_6181; CHECK-NEXT:  ; %bb.7: ; %Flow182; CHECK-NEXT:    v_mov_b32_e32 v0, 0183; CHECK-NEXT:    v_mov_b32_e32 v1, 0184; CHECK-NEXT:  .LBB1_8: ; %post-loop-memcpy-expansion185; CHECK-NEXT:    v_mov_b32_e32 v2, 0186; CHECK-NEXT:    v_mov_b32_e32 v3, v2187; CHECK-NEXT:    v_mov_b32_e32 v4, v2188; CHECK-NEXT:    v_mov_b32_e32 v5, v2189; CHECK-NEXT:    s_and_b64 vcc, exec, 0190; CHECK-NEXT:    flat_store_dwordx4 v[0:1], v[2:5]191; CHECK-NEXT:  .LBB1_9: ; %loop-memcpy-expansion2192; CHECK-NEXT:    s_mov_b64 vcc, vcc193; CHECK-NEXT:    s_cbranch_vccz .LBB1_9194; CHECK-NEXT:  ; %bb.10: ; %DummyReturnBlock195; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)196; CHECK-NEXT:    s_setpc_b64 s[30:31]197entry:198  %spec.select = tail call i64 @llvm.umin.i64(i64 sub (i64 ptrtoint (ptr addrspacecast (ptr addrspace(4) inttoptr (i64 32 to ptr addrspace(4)) to ptr) to i64), i64 ptrtoint (ptr addrspacecast (ptr addrspace(4) null to ptr) to i64)), i64 56)199  %i = trunc i64 %spec.select to i32200  %i1 = urem i32 %i, 16201  %i2 = zext i32 %i to i64202  %i3 = zext i32 %i1 to i64203  %i4 = icmp ne i64 %i2, 0204  br i1 %i4, label %loop-memcpy-expansion.preheader, label %loop-memcpy-residual-header205 206loop-memcpy-expansion.preheader:                  ; preds = %entry207  ret void208 209loop-memcpy-residual:                             ; preds = %loop-memcpy-residual.preheader, %loop-memcpy-residual210  %residual-loop-index1 = phi i64 [ 1, %loop-memcpy-residual ], [ 0, %loop-memcpy-residual.preheader ]211  %i5 = add i64 %residual-loop-index1, 1212  %i6 = icmp ult i64 %i5, %i3213  br i1 %i6, label %loop-memcpy-residual, label %post-loop-memcpy-expansion214 215post-loop-memcpy-expansion:                       ; preds = %loop-memcpy-residual-header.post-loop-memcpy-expansion_crit_edge, %loop-memcpy-residual216  %.pre-phi = phi i64 [ %.pre, %loop-memcpy-residual-header.post-loop-memcpy-expansion_crit_edge ], [ 0, %loop-memcpy-residual ]217  br label %loop-memcpy-expansion2218 219loop-memcpy-expansion2:                           ; preds = %loop-memcpy-expansion2, %post-loop-memcpy-expansion220  %scevgep7 = getelementptr i8, ptr null, i64 %.pre-phi221  store <4 x i32> zeroinitializer, ptr %scevgep7, align 1222  br label %loop-memcpy-expansion2223 224loop-memcpy-residual-header:                      ; preds = %entry225  %i7 = icmp ne i64 %i3, 0226  br i1 %i7, label %loop-memcpy-residual.preheader, label %loop-memcpy-residual-header.post-loop-memcpy-expansion_crit_edge227 228loop-memcpy-residual-header.post-loop-memcpy-expansion_crit_edge: ; preds = %loop-memcpy-residual-header229  %.pre = shl i64 %idxprom, 1230  br label %post-loop-memcpy-expansion231 232loop-memcpy-residual.preheader:                   ; preds = %loop-memcpy-residual-header233  br label %loop-memcpy-residual234}235 236declare void @llvm.memcpy.p0.p0.i64(ptr noalias nocapture writeonly, ptr noalias nocapture readonly, i64, i1 immarg) #0237declare i64 @llvm.umin.i64(i64, i64) #1238 239attributes #0 = { nocallback nofree nounwind willreturn memory(argmem: readwrite) }240attributes #1 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }241