1; RUN: opt < %s -S -loop-unroll -unroll-max-iteration-count-to-analyze=100 -unroll-dynamic-cost-savings-discount=1000 -unroll-threshold=10 -unroll-percent-dynamic-cost-saved-threshold=50 | FileCheck %s 2target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128" 3 4@known_constant = internal unnamed_addr constant [10 x i32] [i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1], align 16 5 6; We should be able to propagate constant data through comparisons. 7; For example, in this test we have a load, which becomes constant after 8; unrolling, making comparison with 0 also known to be 0 (false) - and that 9; will trigger further simplifications. 10; 11; We expect this loop to be unrolled, because in this case load would become 12; constant, which is always 1, and which, in its turn, helps to simplify 13; following comparison, zero-extension, and addition. In total, unrolling should help to 14; optimize more than 50% of all instructions in this case. 15; 16; CHECK-LABEL: @const_compare 17; CHECK-NOT: br i1 % 18; CHECK: ret i32 19define i32 @const_compare(i32* noalias nocapture readonly %b) { 20entry: 21 br label %for.body 22 23for.body: ; preds = %for.inc, %entry 24 %iv.0 = phi i64 [ 0, %entry ], [ %iv.1, %for.body ] 25 %r.0 = phi i32 [ 0, %entry ], [ %r.1, %for.body ] 26 %arrayidx1 = getelementptr inbounds [10 x i32], [10 x i32]* @known_constant, i64 0, i64 %iv.0 27 %x1 = load i32, i32* %arrayidx1, align 4 28 %cmp = icmp eq i32 %x1, 0 29 %cast = zext i1 %cmp to i32 30 %iv.1 = add nuw nsw i64 %iv.0, 1 31 %r.1 = add i32 %r.0, %cast 32 %exitcond = icmp eq i64 %iv.1, 10 33 br i1 %exitcond, label %for.end, label %for.body 34 35for.end: ; preds = %for.inc 36 ret i32 %r.1 37} 38 39; If we can figure out result of comparison on each iteration, we can resolve 40; the depending branch. That means, that the unrolled version of the loop would 41; have less code, because we don't need not-taken basic blocks there. 42; This test checks that this is taken into consideration. 43; We expect this loop to be unrolled, because the most complicated part of its 44; body (if.then block) is never actually executed. 45; CHECK-LABEL: @branch_folded 46; CHECK-NOT: br i1 % 47; CHECK: ret i32 48define i32 @branch_folded(i32* noalias nocapture readonly %b) { 49entry: 50 br label %for.body 51 52for.body: ; preds = %for.inc, %entry 53 %iv.0 = phi i64 [ 0, %entry ], [ %iv.1, %for.inc ] 54 %r.0 = phi i32 [ 0, %entry ], [ %r.1, %for.inc ] 55 %arrayidx1 = getelementptr inbounds [10 x i32], [10 x i32]* @known_constant, i64 0, i64 %iv.0 56 %x1 = load i32, i32* %arrayidx1, align 4 57 %cmp = icmp eq i32 %x1, 0 58 %iv.1 = add nuw nsw i64 %iv.0, 1 59 br i1 %cmp, label %if.then, label %for.inc 60 61if.then: ; preds = %for.body 62 %arrayidx2 = getelementptr inbounds i32, i32* %b, i64 %iv.0 63 %x2 = load i32, i32* %arrayidx2, align 4 64 %add = add nsw i32 %x2, %r.0 65 br label %for.inc 66 67for.inc: ; preds = %for.body, %if.then 68 %r.1 = phi i32 [ %add, %if.then ], [ %x1, %for.body ] 69 %exitcond = icmp eq i64 %iv.1, 10 70 br i1 %exitcond, label %for.end, label %for.body 71 72for.end: ; preds = %for.inc 73 ret i32 %r.1 74} 75 76; This test is similar to the previous one, but in this we use IV in comparison 77; (not a loaded value as we did there). 78; CHECK-LABEL: @branch_iv 79; CHECK-NOT: br i1 % 80; CHECK: ret i64 81define i64 @branch_iv(i64* noalias nocapture readonly %b) { 82entry: 83 br label %for.body 84 85for.body: ; preds = %for.inc, %entry 86 %indvars.iv = phi i64 [ 0, %entry ], [ %tmp3, %for.inc ] 87 %r.030 = phi i64 [ 0, %entry ], [ %r.1, %for.inc ] 88 %cmp3 = icmp eq i64 %indvars.iv, 5 89 %tmp3 = add nuw nsw i64 %indvars.iv, 1 90 br i1 %cmp3, label %if.then, label %for.inc 91 92if.then: ; preds = %for.body 93 %arrayidx2 = getelementptr inbounds i64, i64* %b, i64 %tmp3 94 %tmp1 = load i64, i64* %arrayidx2, align 4 95 %add = add nsw i64 %tmp1, %r.030 96 br label %for.inc 97 98for.inc: ; preds = %if.then, %for.body 99 %r.1 = phi i64 [ %add, %if.then ], [ %r.030, %for.body ] 100 %exitcond = icmp eq i64 %tmp3, 20 101 br i1 %exitcond, label %for.end, label %for.body 102 103for.end: ; preds = %for.inc 104 ret i64 %r.1 105} 106 107; Induction variables are often casted to another type, and that shouldn't 108; prevent us from folding branches. Tthis test specifically checks if we can 109; handle this. Other than thatm it's similar to the previous test. 110; CHECK-LABEL: @branch_iv_trunc 111; CHECK-NOT: br i1 % 112; CHECK: ret i32 113define i32 @branch_iv_trunc(i32* noalias nocapture readonly %b) { 114entry: 115 br label %for.body 116 117for.body: ; preds = %for.inc, %entry 118 %indvars.iv = phi i64 [ 0, %entry ], [ %tmp3, %for.inc ] 119 %r.030 = phi i32 [ 0, %entry ], [ %r.1, %for.inc ] 120 %tmp2 = trunc i64 %indvars.iv to i32 121 %cmp3 = icmp eq i32 %tmp2, 5 122 %tmp3 = add nuw nsw i64 %indvars.iv, 1 123 br i1 %cmp3, label %if.then, label %for.inc 124 125if.then: ; preds = %for.body 126 %arrayidx2 = getelementptr inbounds i32, i32* %b, i64 %tmp3 127 %tmp1 = load i32, i32* %arrayidx2, align 4 128 %add = add nsw i32 %tmp1, %r.030 129 br label %for.inc 130 131for.inc: ; preds = %if.then, %for.body 132 %r.1 = phi i32 [ %add, %if.then ], [ %r.030, %for.body ] 133 %exitcond = icmp eq i64 %tmp3, 10 134 br i1 %exitcond, label %for.end, label %for.body 135 136for.end: ; preds = %for.inc 137 ret i32 %r.1 138} 139 140; Check that we don't crash when we analyze icmp with pointer-typed IV and a 141; pointer. 142; CHECK-LABEL: @ptr_cmp_crash 143; CHECK: ret void 144define void @ptr_cmp_crash() { 145entry: 146 br label %while.body 147 148while.body: 149 %iv.0 = phi i32* [ getelementptr inbounds ([10 x i32], [10 x i32]* @known_constant, i64 0, i64 0), %entry ], [ %iv.1, %while.body ] 150 %iv.1 = getelementptr inbounds i32, i32* %iv.0, i64 1 151 %exitcond = icmp eq i32* %iv.1, getelementptr inbounds ([10 x i32], [10 x i32]* @known_constant, i64 0, i64 9) 152 br i1 %exitcond, label %loop.exit, label %while.body 153 154loop.exit: 155 ret void 156} 157 158; Check that we don't crash when we analyze ptrtoint cast. 159; CHECK-LABEL: @ptrtoint_cast_crash 160; CHECK: ret void 161define void @ptrtoint_cast_crash(i8 * %a) { 162entry: 163 %limit = getelementptr i8, i8* %a, i64 512 164 br label %loop.body 165 166loop.body: 167 %iv.0 = phi i8* [ %a, %entry ], [ %iv.1, %loop.body ] 168 %cast = ptrtoint i8* %iv.0 to i64 169 %iv.1 = getelementptr inbounds i8, i8* %iv.0, i64 1 170 %exitcond = icmp ne i8* %iv.1, %limit 171 br i1 %exitcond, label %loop.body, label %loop.exit 172 173loop.exit: 174 ret void 175} 176 177; Loop unroller should be able to predict that a comparison would become 178; constant if the operands are pointers with the same base and constant 179; offsets. 180; We expect this loop to be unrolled, since most of its instructions would 181; become constant after it. 182; CHECK-LABEL: @ptr_cmp 183; CHECK-NOT: br i1 % 184; CHECK: ret i64 185define i64 @ptr_cmp(i8 * %a) { 186entry: 187 %limit = getelementptr i8, i8* %a, i64 40 188 %start.iv2 = getelementptr i8, i8* %a, i64 7 189 br label %loop.body 190 191loop.body: 192 %iv.0 = phi i8* [ %a, %entry ], [ %iv.1, %loop.body ] 193 %iv2.0 = phi i8* [ %start.iv2, %entry ], [ %iv2.1, %loop.body ] 194 %r.0 = phi i64 [ 0, %entry ], [ %r.1, %loop.body ] 195 %cast = ptrtoint i8* %iv.0 to i64 196 %cmp = icmp eq i8* %iv2.0, %iv.0 197 %sub = sext i1 %cmp to i64 198 %mul = mul i64 %sub, %cast 199 %r.1 = add i64 %r.0, %mul 200 %iv.1 = getelementptr inbounds i8, i8* %iv.0, i64 1 201 %iv2.1 = getelementptr inbounds i8, i8* %iv2.0, i64 1 202 %exitcond = icmp ne i8* %iv.1, %limit 203 br i1 %exitcond, label %loop.body, label %loop.exit 204 205loop.exit: 206 ret i64 %r.1 207} 208