1; 2; Here all the branches we unswitch are exiting from the inner loop. 3; That means we should not be getting exponential behavior on inner-loop 4; unswitch. In fact there should be just a single version of inner-loop, 5; with possibly some outer loop copies. 6; 7; There should be just a single copy of each loop when strictest mutiplier 8; candidates formula (unscaled candidates == 0) is enforced: 9 10; RUN: opt < %s -enable-unswitch-cost-multiplier=true \ 11; RUN: -unswitch-num-initial-unscaled-candidates=0 -unswitch-siblings-toplevel-div=1 \ 12; RUN: -passes='loop(simple-loop-unswitch<nontrivial>),print<loops>' -disable-output 2>&1 | FileCheck %s --check-prefixes=LOOP1 13; 14; RUN: opt < %s -enable-unswitch-cost-multiplier=true \ 15; RUN: -unswitch-num-initial-unscaled-candidates=0 -unswitch-siblings-toplevel-div=16 \ 16; RUN: -passes='loop(simple-loop-unswitch<nontrivial>),print<loops>' -disable-output 2>&1 | FileCheck %s --check-prefixes=LOOP1 17; 18; RUN: opt < %s -enable-unswitch-cost-multiplier=true \ 19; RUN: -unswitch-num-initial-unscaled-candidates=0 -unswitch-siblings-toplevel-div=1 \ 20; RUN: -passes='loop-mssa(simple-loop-unswitch<nontrivial>),print<loops>' -disable-output 2>&1 | FileCheck %s --check-prefixes=LOOP1 21; 22; RUN: opt < %s -enable-unswitch-cost-multiplier=true \ 23; RUN: -unswitch-num-initial-unscaled-candidates=0 -unswitch-siblings-toplevel-div=16 \ 24; RUN: -passes='loop-mssa(simple-loop-unswitch<nontrivial>),print<loops>' -disable-output 2>&1 | FileCheck %s --check-prefixes=LOOP1 25; 26; When we relax the candidates part of a multiplier formula 27; (unscaled candidates == 2) we start getting some unswitches in outer loops, 28; which leads to siblings multiplier kicking in. 29; 30; The tests below also run licm, because it is needed to hoist out 31; loop-invariant freeze instructions, which otherwise may block further 32; unswitching. 33; 34; RUN: opt < %s -enable-unswitch-cost-multiplier=true \ 35; RUN: -unswitch-num-initial-unscaled-candidates=3 -unswitch-siblings-toplevel-div=1 \ 36; RUN: -passes='loop-mssa(licm,simple-loop-unswitch<nontrivial>),print<loops>' -disable-output 2>&1 | \ 37; RUN: sort -b -k 1 | FileCheck %s --check-prefixes=LOOP-UNSCALE3-DIV1 38; 39; NB: sort -b is essential here and below, otherwise blanks might lead to different 40; order depending on locale. 41; 42; RUN: opt < %s -enable-unswitch-cost-multiplier=true \ 43; RUN: -unswitch-num-initial-unscaled-candidates=3 -unswitch-siblings-toplevel-div=2 \ 44; RUN: -passes='loop-mssa(licm,simple-loop-unswitch<nontrivial>),print<loops>' -disable-output 2>&1 | \ 45; RUN: sort -b -k 1 | FileCheck %s --check-prefixes=LOOP-UNSCALE3-DIV2 46; 47; With disabled cost-multiplier we get maximal possible amount of unswitches. 48; 49; RUN: opt < %s -enable-unswitch-cost-multiplier=false \ 50; RUN: -passes='loop-mssa(licm,simple-loop-unswitch<nontrivial>),print<loops>' -disable-output 2>&1 | \ 51; RUN: sort -b -k 1 | FileCheck %s --check-prefixes=LOOP-MAX 52; 53; Single loop nest, not unswitched 54; LOOP1: Loop at depth 1 containing: 55; LOOP1-NOT: Loop at depth 1 containing: 56; LOOP1: Loop at depth 2 containing: 57; LOOP1-NOT: Loop at depth 2 containing: 58; LOOP1: Loop at depth 3 containing: 59; LOOP1-NOT: Loop at depth 3 containing: 60; 61; Half unswitched loop nests, with unscaled3 and div1 it gets less depth1 loops unswitched 62; since they have more cost. 63; LOOP-UNSCALE3-DIV1-COUNT-4: Loop at depth 1 containing: 64; LOOP-UNSCALE3-DIV1-NOT: Loop at depth 1 containing: 65; LOOP-UNSCALE3-DIV1-COUNT-1: Loop at depth 2 containing: 66; LOOP-UNSCALE3-DIV1-NOT: Loop at depth 2 containing: 67; LOOP-UNSCALE3-DIV1-COUNT-1: Loop at depth 3 containing: 68; LOOP-UNSCALE3-DIV1-NOT: Loop at depth 3 containing: 69; 70; Half unswitched loop nests, with unscaled3 and div2 it gets more depth1 loops unswitched 71; as div2 kicks in. 72; LOOP-UNSCALE3-DIV2-COUNT-6: Loop at depth 1 containing: 73; LOOP-UNSCALE3-DIV2-NOT: Loop at depth 1 containing: 74; LOOP-UNSCALE3-DIV2-COUNT-1: Loop at depth 2 containing: 75; LOOP-UNSCALE3-DIV2-NOT: Loop at depth 2 containing: 76; LOOP-UNSCALE3-DIV2-COUNT-1: Loop at depth 3 containing: 77; LOOP-UNSCALE3-DIV2-NOT: Loop at depth 3 containing: 78; 79; Maximally unswitched (copy of the outer loop per each condition) 80; LOOP-MAX-COUNT-6: Loop at depth 1 containing: 81; LOOP-MAX-NOT: Loop at depth 1 containing: 82; LOOP-MAX-COUNT-1: Loop at depth 2 containing: 83; LOOP-MAX-NOT: Loop at depth 2 containing: 84; LOOP-MAX-COUNT-1: Loop at depth 3 containing: 85; LOOP-MAX-NOT: Loop at depth 3 containing: 86 87declare void @bar() 88 89define void @loop_nested3_conds5(ptr %addr, i1 %c1, i1 %c2, i1 %c3, i1 %c4, i1 %c5) { 90entry: 91 %addr2 = getelementptr i32, ptr %addr, i64 1 92 %addr3 = getelementptr i32, ptr %addr, i64 2 93 br label %outer 94outer: 95 %iv1 = phi i32 [0, %entry], [%iv1.next, %outer_latch] 96 %iv1.next = add i32 %iv1, 1 97 ;; skip nontrivial unswitch 98 call void @bar() 99 br label %middle 100middle: 101 %iv2 = phi i32 [0, %outer], [%iv2.next, %middle_latch] 102 %iv2.next = add i32 %iv2, 1 103 ;; skip nontrivial unswitch 104 call void @bar() 105 br label %loop 106loop: 107 %iv3 = phi i32 [0, %middle], [%iv3.next, %loop_latch] 108 %iv3.next = add i32 %iv3, 1 109 ;; skip nontrivial unswitch 110 call void @bar() 111 br i1 %c1, label %loop_next1_left, label %outer_latch 112loop_next1_left: 113 br label %loop_next1 114loop_next1_right: 115 br label %loop_next1 116 117loop_next1: 118 br i1 %c2, label %loop_next2_left, label %outer_latch 119loop_next2_left: 120 br label %loop_next2 121loop_next2_right: 122 br label %loop_next2 123 124loop_next2: 125 br i1 %c3, label %loop_next3_left, label %outer_latch 126loop_next3_left: 127 br label %loop_next3 128loop_next3_right: 129 br label %loop_next3 130 131loop_next3: 132 br i1 %c4, label %loop_next4_left, label %outer_latch 133loop_next4_left: 134 br label %loop_next4 135loop_next4_right: 136 br label %loop_next4 137 138loop_next4: 139 br i1 %c5, label %loop_latch_left, label %outer_latch 140loop_latch_left: 141 br label %loop_latch 142loop_latch_right: 143 br label %loop_latch 144 145loop_latch: 146 store volatile i32 0, ptr %addr 147 %test_loop = icmp slt i32 %iv3, 50 148 br i1 %test_loop, label %loop, label %middle_latch 149middle_latch: 150 store volatile i32 0, ptr %addr2 151 %test_middle = icmp slt i32 %iv2, 50 152 br i1 %test_middle, label %middle, label %outer_latch 153outer_latch: 154 store volatile i32 0, ptr %addr3 155 %test_outer = icmp slt i32 %iv1, 50 156 br i1 %test_outer, label %outer, label %exit 157exit: 158 ret void 159} 160