226 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 22; RUN: llc -O2 -o - %s | FileCheck %s3target datalayout = "e-m:e-i64:64-n32:64"4target triple = "powerpc64le-grtev4-linux-gnu"5 6; Intended layout:7; The code for tail-duplication during layout will produce the layout:8; test19; test210; body1 (with copy of test2)11; body212; exit13 14define void @tail_dup_break_cfg(i32 %tag) {15; CHECK-LABEL: tail_dup_break_cfg:16; CHECK: # %bb.0: # %entry17; CHECK-NEXT: mflr 018; CHECK-NEXT: .cfi_def_cfa_offset 4819; CHECK-NEXT: .cfi_offset lr, 1620; CHECK-NEXT: .cfi_offset r30, -1621; CHECK-NEXT: std 30, -16(1) # 8-byte Folded Spill22; CHECK-NEXT: stdu 1, -48(1)23; CHECK-NEXT: mr 30, 324; CHECK-NEXT: std 0, 64(1)25; CHECK-NEXT: andi. 3, 30, 126; CHECK-NEXT: bc 12, 1, .LBB0_327; CHECK-NEXT: # %bb.1: # %test228; CHECK-NEXT: andi. 3, 30, 229; CHECK-NEXT: bne 0, .LBB0_430; CHECK-NEXT: .LBB0_2: # %exit31; CHECK-NEXT: addi 1, 1, 4832; CHECK-NEXT: ld 0, 16(1)33; CHECK-NEXT: ld 30, -16(1) # 8-byte Folded Reload34; CHECK-NEXT: mtlr 035; CHECK-NEXT: blr36; CHECK-NEXT: .LBB0_3: # %body137; CHECK-NEXT: bl a38; CHECK-NEXT: nop39; CHECK-NEXT: bl a40; CHECK-NEXT: nop41; CHECK-NEXT: bl a42; CHECK-NEXT: nop43; CHECK-NEXT: bl a44; CHECK-NEXT: nop45; CHECK-NEXT: andi. 3, 30, 246; CHECK-NEXT: beq 0, .LBB0_247; CHECK-NEXT: .LBB0_4: # %body248; CHECK-NEXT: bl b49; CHECK-NEXT: nop50; CHECK-NEXT: bl b51; CHECK-NEXT: nop52; CHECK-NEXT: bl b53; CHECK-NEXT: nop54; CHECK-NEXT: bl b55; CHECK-NEXT: nop56; CHECK-NEXT: b .LBB0_257entry:58 br label %test159test1:60 %tagbit1 = and i32 %tag, 161 %tagbit1eq0 = icmp eq i32 %tagbit1, 062 br i1 %tagbit1eq0, label %test2, label %body1, !prof !1 ; %test2 more likely63body1:64 call void @a()65 call void @a()66 call void @a()67 call void @a()68 br label %test269test2:70 %tagbit2 = and i32 %tag, 271 %tagbit2eq0 = icmp eq i32 %tagbit2, 072 br i1 %tagbit2eq0, label %exit, label %body2, !prof !1 ; %exit more likely73body2:74 call void @b()75 call void @b()76 call void @b()77 call void @b()78 br label %exit79exit:80 ret void81}82 83; The branch weights here hint that we shouldn't tail duplicate in this case.84define void @tail_dup_dont_break_cfg(i32 %tag) {85; CHECK-LABEL: tail_dup_dont_break_cfg:86; CHECK: # %bb.0: # %entry87; CHECK-NEXT: mflr 088; CHECK-NEXT: .cfi_def_cfa_offset 4889; CHECK-NEXT: .cfi_offset lr, 1690; CHECK-NEXT: .cfi_offset r30, -1691; CHECK-NEXT: std 30, -16(1) # 8-byte Folded Spill92; CHECK-NEXT: stdu 1, -48(1)93; CHECK-NEXT: mr 30, 394; CHECK-NEXT: std 0, 64(1)95; CHECK-NEXT: andi. 3, 30, 196; CHECK-NEXT: bc 4, 1, .LBB1_297; CHECK-NEXT: # %bb.1: # %body198; CHECK-NEXT: bl a99; CHECK-NEXT: nop100; CHECK-NEXT: bl a101; CHECK-NEXT: nop102; CHECK-NEXT: bl a103; CHECK-NEXT: nop104; CHECK-NEXT: bl a105; CHECK-NEXT: nop106; CHECK-NEXT: .LBB1_2: # %test2107; CHECK-NEXT: andi. 3, 30, 2108; CHECK-NEXT: beq 0, .LBB1_4109; CHECK-NEXT: # %bb.3: # %body2110; CHECK-NEXT: bl b111; CHECK-NEXT: nop112; CHECK-NEXT: bl b113; CHECK-NEXT: nop114; CHECK-NEXT: bl b115; CHECK-NEXT: nop116; CHECK-NEXT: bl b117; CHECK-NEXT: nop118; CHECK-NEXT: .LBB1_4: # %exit119; CHECK-NEXT: addi 1, 1, 48120; CHECK-NEXT: ld 0, 16(1)121; CHECK-NEXT: ld 30, -16(1) # 8-byte Folded Reload122; CHECK-NEXT: mtlr 0123; CHECK-NEXT: blr124entry:125 br label %test1126test1:127 %tagbit1 = and i32 %tag, 1128 %tagbit1eq0 = icmp eq i32 %tagbit1, 0129 br i1 %tagbit1eq0, label %test2, label %body1, !prof !1 ; %test2 more likely130body1:131 call void @a()132 call void @a()133 call void @a()134 call void @a()135 br label %test2136test2:137 %tagbit2 = and i32 %tag, 2138 %tagbit2eq0 = icmp ne i32 %tagbit2, 0139 br i1 %tagbit2eq0, label %body2, label %exit, !prof !3 ; %body2 more likely140body2:141 call void @b()142 call void @b()143 call void @b()144 call void @b()145 br label %exit146exit:147 ret void148}149 150declare void @a()151declare void @b()152declare void @c()153declare void @d()154 155; This function arranges for the successors of %succ to have already been laid156; out. When we consider whether to lay out succ after bb and to tail-duplicate157; it, v and ret have already been placed, so we tail-duplicate as it removes a158; branch and strictly increases fallthrough159define void @tail_dup_no_succ(i32 %tag) {160; CHECK-LABEL: tail_dup_no_succ:161; CHECK: # %bb.0: # %entry162; CHECK-NEXT: mflr 0163; CHECK-NEXT: .cfi_def_cfa_offset 48164; CHECK-NEXT: .cfi_offset lr, 16165; CHECK-NEXT: .cfi_offset r30, -16166; CHECK-NEXT: std 30, -16(1) # 8-byte Folded Spill167; CHECK-NEXT: stdu 1, -48(1)168; CHECK-NEXT: andi. 4, 3, 1169; CHECK-NEXT: std 0, 64(1)170; CHECK-NEXT: bc 12, 1, .LBB2_3171; CHECK-NEXT: .LBB2_1: # %v172; CHECK-NEXT: bl d173; CHECK-NEXT: nop174; CHECK-NEXT: bl d175; CHECK-NEXT: nop176; CHECK-NEXT: .LBB2_2: # %ret177; CHECK-NEXT: addi 1, 1, 48178; CHECK-NEXT: ld 0, 16(1)179; CHECK-NEXT: ld 30, -16(1) # 8-byte Folded Reload180; CHECK-NEXT: mtlr 0181; CHECK-NEXT: blr182; CHECK-NEXT: .LBB2_3: # %bb183; CHECK-NEXT: andi. 4, 3, 2184; CHECK-NEXT: bne 0, .LBB2_5185; CHECK-NEXT: # %bb.4: # %succ186; CHECK-NEXT: andi. 3, 3, 4187; CHECK-NEXT: beq 0, .LBB2_2188; CHECK-NEXT: b .LBB2_1189; CHECK-NEXT: .LBB2_5: # %c190; CHECK-NEXT: mr 30, 3191; CHECK-NEXT: bl c192; CHECK-NEXT: nop193; CHECK-NEXT: bl c194; CHECK-NEXT: nop195; CHECK-NEXT: mr 3, 30196; CHECK-NEXT: andi. 3, 3, 4197; CHECK-NEXT: beq 0, .LBB2_2198; CHECK-NEXT: b .LBB2_1199entry:200 %tagbit1 = and i32 %tag, 1201 %tagbit1eq0 = icmp eq i32 %tagbit1, 0202 br i1 %tagbit1eq0, label %v, label %bb, !prof !2 ; %v very much more likely203bb:204 %tagbit2 = and i32 %tag, 2205 %tagbit2eq0 = icmp eq i32 %tagbit2, 0206 br i1 %tagbit2eq0, label %succ, label %c, !prof !3 ; %succ more likely207c:208 call void @c()209 call void @c()210 br label %succ211succ:212 %tagbit3 = and i32 %tag, 4213 %tagbit3eq0 = icmp eq i32 %tagbit3, 0214 br i1 %tagbit3eq0, label %ret, label %v, !prof !1 ; %u more likely215v:216 call void @d()217 call void @d()218 br label %ret219ret:220 ret void221}222 223!1 = !{!"branch_weights", i32 5, i32 3}224!2 = !{!"branch_weights", i32 95, i32 5}225!3 = !{!"branch_weights", i32 8, i32 3}226