1; RUN: opt -basic-aa -loop-idiom -use-lir-code-size-heurs=true < %s -S | FileCheck %s 2 3; When compiling for codesize we avoid idiom recognition for a 4; multi-block loop unless it is one of 5; - a loop_memset idiom, or 6; - a memset/memcpy idiom in a nested loop. 7 8declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i1) 9@APPLES = common global i32 0, align 4 10@ORANGES = common global i32 0, align 4 11 12; LIR allowed: loop_memset idiom in multi-block loop. 13; =================================================== 14; CHECK-LABEL: @LoopMemset 15; CHECK: for.body.preheader: 16; CHECK: call void @llvm.memset 17; CHECK: for.body: 18; 19define i32 @LoopMemset([2048 x i8]* noalias nocapture %DST, i32 %SIZE) local_unnamed_addr optsize { 20entry: 21 %cmp12 = icmp sgt i32 %SIZE, 0 22 br i1 %cmp12, label %for.body.preheader, label %for.end 23 24for.body.preheader: ; preds = %entry 25 br label %for.body 26 27for.body: ; preds = %for.body.preheader, %for.inc 28 %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.inc ] 29 %BASKET.013 = phi i32 [ %BASKET.1, %for.inc ], [ 0, %for.body.preheader ] 30 %arraydecay = getelementptr inbounds [2048 x i8], [2048 x i8]* %DST, i64 %indvars.iv, i64 0 31 tail call void @llvm.memset.p0i8.i64(i8* %arraydecay, i8 -1, i64 2048, i1 false) 32 %0 = trunc i64 %indvars.iv to i32 33 %rem11 = and i32 %0, 1 34 %cmp1 = icmp eq i32 %rem11, 0 35 %1 = load i32, i32* @ORANGES, align 4 36 %2 = load i32, i32* @APPLES, align 4 37 br i1 %cmp1, label %if.then, label %if.else 38 39if.else: ; preds = %for.body 40 %dec3 = add nsw i32 %2, -1 41 store i32 %dec3, i32* @APPLES, align 4 42 br label %for.inc 43 44if.then: ; preds = %for.body 45 %dec = add nsw i32 %1, -1 46 store i32 %dec, i32* @ORANGES, align 4 47 br label %for.inc 48 49for.inc: ; preds = %if.then, %if.else 50 %.pn = phi i32 [ %2, %if.then ], [ %1, %if.else ] 51 %BASKET.1 = add nsw i32 %.pn, %BASKET.013 52 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 53 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 54 %exitcond = icmp ne i32 %lftr.wideiv, %SIZE 55 br i1 %exitcond, label %for.body, label %for.end.loopexit 56 57for.end.loopexit: ; preds = %for.inc 58 %BASKET.1.lcssa = phi i32 [ %BASKET.1, %for.inc ] 59 br label %for.end 60 61for.end: ; preds = %for.end.loopexit, %entry 62 %BASKET.0.lcssa = phi i32 [ 0, %entry ], [ %BASKET.1.lcssa, %for.end.loopexit ] 63 ret i32 %BASKET.0.lcssa 64} 65 66; LIR allowed: memset idiom in multi-block nested loop, 67; which is recognized as a loop_memset in its turn. 68; ===================================================== 69; CHECK-LABEL: @NestedMemset_LoopMemset 70; CHECK: for.cond1.preheader.preheader: 71; CHECK: call void @llvm.memset 72; CHECK: for.cond1.preheader: 73; 74define i32 @NestedMemset_LoopMemset([2046 x i8]* noalias nocapture %DST, i32 %SIZE) local_unnamed_addr optsize { 75entry: 76 %cmp25 = icmp sgt i32 %SIZE, 0 77 br i1 %cmp25, label %for.cond1.preheader.preheader, label %for.end11 78 79for.cond1.preheader.preheader: ; preds = %entry 80 br label %for.cond1.preheader 81 82for.cond1.preheader: ; preds = %for.cond1.preheader.preheader, %for.inc9 83 %i.027 = phi i32 [ %inc10, %for.inc9 ], [ 0, %for.cond1.preheader.preheader ] 84 %BASKET.026 = phi i32 [ %BASKET.2.lcssa, %for.inc9 ], [ 0, %for.cond1.preheader.preheader ] 85 %idxprom4 = sext i32 %i.027 to i64 86 %rem22 = and i32 %i.027, 1 87 %cmp6 = icmp eq i32 %rem22, 0 88 br label %for.body3 89 90for.body3: ; preds = %for.cond1.preheader, %for.inc 91 %indvars.iv = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next, %for.inc ] 92 %BASKET.123 = phi i32 [ %BASKET.026, %for.cond1.preheader ], [ %BASKET.2, %for.inc ] 93 %arrayidx5 = getelementptr inbounds [2046 x i8], [2046 x i8]* %DST, i64 %idxprom4, i64 %indvars.iv 94 store i8 -1, i8* %arrayidx5, align 1 95 %0 = load i32, i32* @APPLES, align 4 96 %1 = load i32, i32* @ORANGES, align 4 97 br i1 %cmp6, label %if.then, label %if.else 98 99if.else: ; preds = %for.body3 100 %dec8 = add nsw i32 %0, -1 101 store i32 %dec8, i32* @APPLES, align 4 102 br label %for.inc 103 104if.then: ; preds = %for.body3 105 %dec = add nsw i32 %1, -1 106 store i32 %dec, i32* @ORANGES, align 4 107 br label %for.inc 108 109for.inc: ; preds = %if.then, %if.else 110 %.pn = phi i32 [ %0, %if.then ], [ %1, %if.else ] 111 %BASKET.2 = add nsw i32 %.pn, %BASKET.123 112 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 113 %exitcond = icmp ne i64 %indvars.iv.next, 2046 114 br i1 %exitcond, label %for.body3, label %for.inc9 115 116for.inc9: ; preds = %for.inc 117 %BASKET.2.lcssa = phi i32 [ %BASKET.2, %for.inc ] 118 %inc10 = add nsw i32 %i.027, 1 119 %cmp = icmp slt i32 %inc10, %SIZE 120 br i1 %cmp, label %for.cond1.preheader, label %for.end11.loopexit 121 122for.end11.loopexit: ; preds = %for.inc9 123 %BASKET.2.lcssa.lcssa = phi i32 [ %BASKET.2.lcssa, %for.inc9 ] 124 br label %for.end11 125 126for.end11: ; preds = %for.end11.loopexit, %entry 127 %BASKET.0.lcssa = phi i32 [ 0, %entry ], [ %BASKET.2.lcssa.lcssa, %for.end11.loopexit ] 128 ret i32 %BASKET.0.lcssa 129} 130 131; LIR avoided: memset idiom in multi-block top-level loop. 132; ======================================================== 133; CHECK-LABEL: @Non_NestedMemset 134; CHECK-NOT: call void @llvm.memset 135; 136define i32 @Non_NestedMemset(i8* noalias nocapture %DST, i32 %SIZE) local_unnamed_addr optsize { 137entry: 138 %cmp12 = icmp sgt i32 %SIZE, 0 139 br i1 %cmp12, label %for.body.preheader, label %for.end 140 141for.body.preheader: ; preds = %entry 142 br label %for.body 143 144for.body: ; preds = %for.body.preheader, %for.inc 145 %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.inc ] 146 %BASKET.013 = phi i32 [ %BASKET.1, %for.inc ], [ 0, %for.body.preheader ] 147 %arrayidx = getelementptr inbounds i8, i8* %DST, i64 %indvars.iv 148 store i8 -1, i8* %arrayidx, align 1 149 %0 = trunc i64 %indvars.iv to i32 150 %rem11 = and i32 %0, 1 151 %cmp1 = icmp eq i32 %rem11, 0 152 %1 = load i32, i32* @ORANGES, align 4 153 %2 = load i32, i32* @APPLES, align 4 154 br i1 %cmp1, label %if.then, label %if.else 155 156if.else: ; preds = %for.body 157 %dec3 = add nsw i32 %2, -1 158 store i32 %dec3, i32* @APPLES, align 4 159 br label %for.inc 160 161if.then: ; preds = %for.body 162 %dec = add nsw i32 %1, -1 163 store i32 %dec, i32* @ORANGES, align 4 164 br label %for.inc 165 166for.inc: ; preds = %if.then, %if.else 167 %.pn = phi i32 [ %2, %if.then ], [ %1, %if.else ] 168 %BASKET.1 = add nsw i32 %.pn, %BASKET.013 169 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 170 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 171 %exitcond = icmp ne i32 %lftr.wideiv, %SIZE 172 br i1 %exitcond, label %for.body, label %for.end.loopexit 173 174for.end.loopexit: ; preds = %for.inc 175 %BASKET.1.lcssa = phi i32 [ %BASKET.1, %for.inc ] 176 br label %for.end 177 178for.end: ; preds = %for.end.loopexit, %entry 179 %BASKET.0.lcssa = phi i32 [ 0, %entry ], [ %BASKET.1.lcssa, %for.end.loopexit ] 180 ret i32 %BASKET.0.lcssa 181} 182 183