1 ; RUN: llc -O3 -march=hexagon -mcpu=hexagonv5 < %s | FileCheck %s
3 ; Generate hardware loop when unknown trip count loop is vectorized.
5 ; CHECK: loop0(.LBB{{[0-9]*}}_{{[0-9]*}},r{{[0-9]+}})
7 ; CHECK: loop0(.LBB{{[0-9]*}}_{{[0-9]*}},r{{[0-9]+}})
10 @A = common global [1000 x i32] zeroinitializer, align 8
11 @B = common global [1000 x i32] zeroinitializer, align 8
13 define i32 @dotprod2(i32 %count) #0 {
15 %cmp6 = icmp sgt i32 %count, 0
16 br i1 %cmp6, label %polly.cond, label %for.end
22 %sum.0.lcssa.reg2mem.0.load37 = phi i32 [ 0, %entry.split ], [ %p_add34, %polly.loop_if13 ], [ %p_add, %for.end.loopexit ]
23 ret i32 %sum.0.lcssa.reg2mem.0.load37
26 %0 = icmp sgt i32 %count, 1
27 br i1 %0, label %polly.loop_if, label %polly.loop_if13
29 polly.loop_exit.loopexit:
30 br label %polly.loop_exit
33 %1 = phi <2 x i32> [ zeroinitializer, %polly.loop_if ], [ %addp_vec, %polly.loop_exit.loopexit ]
34 %2 = extractelement <2 x i32> %1, i32 0
35 %3 = extractelement <2 x i32> %1, i32 1
36 %add_sum = add i32 %2, %3
37 br label %polly.loop_if13
40 %4 = add i32 %count, -1
41 %leftover_lb = and i32 %4, -2
42 %polly.loop_guard = icmp eq i32 %leftover_lb, 0
43 br i1 %polly.loop_guard, label %polly.loop_exit, label %polly.loop_preheader
46 %addp_vec28 = phi <2 x i32> [ zeroinitializer, %polly.loop_preheader ], [ %addp_vec, %polly.stmt.for.body ]
47 %scevgep.phi = phi i32* [ getelementptr inbounds ([1000 x i32], [1000 x i32]* @A, i32 0, i32 0), %polly.loop_preheader ], [ %scevgep.inc, %polly.stmt.for.body ]
48 %scevgep9.phi = phi i32* [ getelementptr inbounds ([1000 x i32], [1000 x i32]* @B, i32 0, i32 0), %polly.loop_preheader ], [ %scevgep9.inc, %polly.stmt.for.body ]
49 %polly.indvar = phi i32 [ 0, %polly.loop_preheader ], [ %polly.indvar_next, %polly.stmt.for.body ]
50 %vector_ptr = bitcast i32* %scevgep.phi to <2 x i32>*
51 %_p_vec_full = load <2 x i32>, <2 x i32>* %vector_ptr, align 8
52 %vector_ptr10 = bitcast i32* %scevgep9.phi to <2 x i32>*
53 %_p_vec_full11 = load <2 x i32>, <2 x i32>* %vector_ptr10, align 8
54 %mulp_vec = mul <2 x i32> %_p_vec_full11, %_p_vec_full
55 %addp_vec = add <2 x i32> %mulp_vec, %addp_vec28
56 %polly.indvar_next = add nsw i32 %polly.indvar, 2
57 %polly.loop_cond = icmp eq i32 %polly.indvar, %polly.adjust_ub
58 %scevgep.inc = getelementptr i32, i32* %scevgep.phi, i32 2
59 %scevgep9.inc = getelementptr i32, i32* %scevgep9.phi, i32 2
60 br i1 %polly.loop_cond, label %polly.loop_exit.loopexit, label %polly.stmt.for.body
63 %polly.adjust_ub = add i32 %leftover_lb, -2
64 br label %polly.stmt.for.body
67 %p_add34 = phi i32 [ 0, %polly.cond ], [ %add_sum, %polly.loop_exit ]
68 %merge.lb = phi i32 [ 0, %polly.cond ], [ %leftover_lb, %polly.loop_exit ]
69 %polly.loop_guard17 = icmp slt i32 %merge.lb, %count
70 br i1 %polly.loop_guard17, label %polly.loop_preheader15, label %for.end
72 polly.stmt.for.body22:
73 %p_add30 = phi i32 [ %p_add34, %polly.loop_preheader15 ], [ %p_add, %polly.stmt.for.body22 ]
74 %polly.indvar18 = phi i32 [ %merge.lb, %polly.loop_preheader15 ], [ %polly.indvar_next19, %polly.stmt.for.body22 ]
75 %5 = tail call i32 @llvm.annotation.i32(i32 %polly.indvar18, i8* null, i8* null, i32 0), !polly.loop.smallTripCount !0
76 %scevgep23 = getelementptr [1000 x i32], [1000 x i32]* @A, i32 0, i32 %polly.indvar18
77 %_p_scalar_ = load i32, i32* %scevgep23, align 4
78 %scevgep24 = getelementptr [1000 x i32], [1000 x i32]* @B, i32 0, i32 %polly.indvar18
79 %_p_scalar_25 = load i32, i32* %scevgep24, align 4
80 %p_mul = mul nsw i32 %_p_scalar_25, %_p_scalar_
81 %p_add = add nsw i32 %p_mul, %p_add30
82 %polly.indvar_next19 = add nsw i32 %polly.indvar18, 1
83 %polly.loop_cond21 = icmp slt i32 %polly.indvar18, %polly.adjust_ub20
84 br i1 %polly.loop_cond21, label %polly.stmt.for.body22, label %for.end.loopexit
86 polly.loop_preheader15:
87 %polly.adjust_ub20 = add i32 %count, -1
88 br label %polly.stmt.for.body22
91 declare i32 @llvm.annotation.i32(i32, i8*, i8*, i32) #1