1; RUN: llc -march=hexagon < %s | FileCheck %s 2 3; Test that the pipeliner correctly fixes up the pipelined CFG when the loop 4; has a constant trip count, and the trip count is less than the number of 5; prolog blocks. Prior to the bug, the pipeliner deleted one extra prolog and 6; epilog stage. We check this by counting the number of sxth instructions. 7 8; CHECK: r{{[0-9]+}} = sxth(r{{[0-9]+}}) 9; CHECK: r{{[0-9]+}} = sxth(r{{[0-9]+}}) 10; CHECK: r{{[0-9]+}} = sxth(r{{[0-9]+}}) 11; CHECK: r{{[0-9]+}} = sxth(r{{[0-9]+}}) 12 13; Function Attrs: nounwind readonly 14define signext i16 @f0(i16* nocapture readonly %a0, i16* nocapture readnone %a1, i16* nocapture readonly %a2, i16* nocapture readonly %a3, i16 signext %a4, i16 signext %a5, i16 signext %a6) #0 { 15b0: 16 %v0 = icmp sgt i16 %a5, 0 17 br i1 %v0, label %b1, label %b7 18 19b1: ; preds = %b0 20 %v1 = load i16, i16* %a0, align 2 21 %v2 = sext i16 %v1 to i32 22 %v3 = load i16, i16* %a3, align 2 23 %v4 = sext i16 %v3 to i32 24 br label %b2 25 26b2: ; preds = %b6, %b1 27 %v5 = phi i32 [ 2147483647, %b1 ], [ %v44, %b6 ] 28 %v6 = phi i16 [ 0, %b1 ], [ %v45, %b6 ] 29 %v7 = phi i16 [ 0, %b1 ], [ %v43, %b6 ] 30 %v8 = phi i16* [ %a2, %b1 ], [ %v38, %b6 ] 31 %v9 = load i16, i16* %v8, align 2 32 %v10 = sext i16 %v9 to i32 33 %v11 = tail call i32 @llvm.hexagon.A2.subh.l16.sat.ll(i32 %v2, i32 %v10) 34 %v12 = shl i32 %v11, 16 35 %v13 = ashr exact i32 %v12, 16 36 %v14 = tail call i32 @llvm.hexagon.M2.mpy.sat.ll.s1(i32 %v13, i32 %v4) 37 %v15 = tail call i32 @llvm.hexagon.M2.hmmpyl.s1(i32 %v14, i32 %v13) 38 %v16 = tail call i32 @llvm.hexagon.S2.asr.r.r.sat(i32 %v15, i32 10) 39 %v17 = getelementptr inbounds i16, i16* %v8, i32 1 40 br label %b3 41 42b3: ; preds = %b3, %b2 43 %v18 = phi i16* [ %v8, %b2 ], [ %v19, %b3 ] 44 %v19 = phi i16* [ %v17, %b2 ], [ %v38, %b3 ] 45 %v20 = phi i32 [ %v16, %b2 ], [ %v36, %b3 ] 46 %v21 = phi i32 [ 1, %b2 ], [ %v37, %b3 ] 47 %v22 = getelementptr inbounds i16, i16* %a0, i32 %v21 48 %v23 = load i16, i16* %v22, align 2 49 %v24 = sext i16 %v23 to i32 50 %v25 = load i16, i16* %v19, align 2 51 %v26 = sext i16 %v25 to i32 52 %v27 = tail call i32 @llvm.hexagon.A2.subh.l16.sat.ll(i32 %v24, i32 %v26) 53 %v28 = shl i32 %v27, 16 54 %v29 = ashr exact i32 %v28, 16 55 %v30 = getelementptr inbounds i16, i16* %a3, i32 %v21 56 %v31 = load i16, i16* %v30, align 2 57 %v32 = sext i16 %v31 to i32 58 %v33 = tail call i32 @llvm.hexagon.M2.mpy.sat.ll.s1(i32 %v29, i32 %v32) 59 %v34 = tail call i32 @llvm.hexagon.M2.hmmpyl.s1(i32 %v33, i32 %v29) 60 %v35 = tail call i32 @llvm.hexagon.S2.asr.r.r.sat(i32 %v34, i32 10) 61 %v36 = tail call i32 @llvm.hexagon.A2.addsat(i32 %v20, i32 %v35) 62 %v37 = add i32 %v21, 1 63 %v38 = getelementptr inbounds i16, i16* %v18, i32 2 64 %v39 = icmp eq i32 %v37, 3 65 br i1 %v39, label %b4, label %b3 66 67b4: ; preds = %b3 68 %v40 = tail call i32 @llvm.hexagon.A2.subsat(i32 %v36, i32 %v5) 69 %v41 = icmp slt i32 %v40, 0 70 br i1 %v41, label %b5, label %b6 71 72b5: ; preds = %b4 73 %v42 = tail call i32 @llvm.hexagon.A2.addsat(i32 %v36, i32 0) 74 br label %b6 75 76b6: ; preds = %b5, %b4 77 %v43 = phi i16 [ %v6, %b5 ], [ %v7, %b4 ] 78 %v44 = phi i32 [ %v42, %b5 ], [ %v5, %b4 ] 79 %v45 = add i16 %v6, 1 80 %v46 = icmp eq i16 %v45, %a5 81 br i1 %v46, label %b7, label %b2 82 83b7: ; preds = %b6, %b0 84 %v47 = phi i16 [ 0, %b0 ], [ %v43, %b6 ] 85 ret i16 %v47 86} 87 88; Function Attrs: nounwind readnone 89declare i32 @llvm.hexagon.A2.subh.l16.sat.ll(i32, i32) #1 90 91; Function Attrs: nounwind readnone 92declare i32 @llvm.hexagon.M2.mpy.sat.ll.s1(i32, i32) #1 93 94; Function Attrs: nounwind readnone 95declare i32 @llvm.hexagon.M2.hmmpyl.s1(i32, i32) #1 96 97; Function Attrs: nounwind readnone 98declare i32 @llvm.hexagon.S2.asr.r.r.sat(i32, i32) #1 99 100; Function Attrs: nounwind readnone 101declare i32 @llvm.hexagon.A2.addsat(i32, i32) #1 102 103; Function Attrs: nounwind readnone 104declare i32 @llvm.hexagon.A2.subsat(i32, i32) #1 105 106attributes #0 = { nounwind readonly "target-cpu"="hexagonv60" } 107attributes #1 = { nounwind readnone } 108