| blob | d26c5f74f84de5503f3107b5b880cc359d06d622 |
1 ;RUN: llc -march=hexagon -mcpu=hexagonv66 -mhvx -filetype=obj < %s -o - | llvm-objdump --mcpu=hexagonv66 --mattr=+hvx -d - | FileCheck --check-prefix=CHECK-V66 %s
2 ;RUN: llc -march=hexagon -mcpu=hexagonv67 -mhvx -filetype=obj < %s -o - | llvm-objdump --mcpu=hexagonv67 --mattr=+hvx -d - | FileCheck --check-prefix=CHECK-V67 %s
4 ; Should not attempt to use v<even>:<odd> 'reverse' vector regpairs
5 ; on old or new arches (should not crash).
7 ; CHECK-V66: vcombine
8 ; CHECK-V67: vcombine
9 declare <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32>, <16 x i32>)
10 declare <16 x i32> @llvm.hexagon.V6.vd0()
11 declare <32 x i32> @llvm.hexagon.V6.vmpybus(<16 x i32>, i32)
12 declare <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32>, <32 x i32>, i32)
13 declare <16 x i32> @llvm.hexagon.V6.hi(<32 x i32>)
14 declare <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32>, <16 x i32>, i32 )
15 declare <16 x i32> @llvm.hexagon.V6.lo(<32 x i32>)
16 declare <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32>, <16 x i32>, i32 )
17 declare <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32>, <16 x i32>)
18 declare <16 x i32> @llvm.hexagon.V6.vmpyihb.acc(<16 x i32>, <16 x i32>, i32)
19 declare <16 x i32> @llvm.hexagon.V6.vasrhubrndsat(<16 x i32>, <16 x i32>, i32)
21 declare <32 x i32> @llvm.hexagon.V6.vaddubh(<16 x i32>, <16 x i32>)
22 declare <32 x i32> @llvm.hexagon.V6.vmpybus.acc(<32 x i32>, <16 x i32>, i32)
23 declare <16 x i32> @llvm.hexagon.V6.vmpyiwb.acc(<16 x i32>, <16 x i32>, i32)
24 declare <16 x i32> @llvm.hexagon.V6.vshuffob(<16 x i32>, <16 x i32>)
27 define void @Gaussian7x7u8PerRow(i8* %src, i32 %stride, i32 %width, i8* %dst) #0 {
28 entry:
29 %mul = mul i32 %stride, 3
30 %idx.neg = sub i32 0, %mul
31 %add.ptr = getelementptr i8, i8* %src, i32 %idx.neg
32 bitcast i8* %add.ptr to <16 x i32>*
33 %mul1 = shl i32 %stride, 1
34 %idx.neg2 = sub i32 0, %mul1
35 %add.ptr3 = getelementptr i8, i8* %src, i32 %idx.neg2
36 bitcast i8* %add.ptr3 to <16 x i32>*
37 %idx.neg5 = sub i32 0, %stride
38 %add.ptr6 = getelementptr i8, i8* %src, i32 %idx.neg5
39 bitcast i8* %add.ptr6 to <16 x i32>*
40 bitcast i8* %src to <16 x i32>*
41 %add.ptr10 = getelementptr i8, i8* %src, i32 %stride
42 bitcast i8* %add.ptr10 to <16 x i32>*
43 %add.ptr12 = getelementptr i8, i8* %src, i32 %mul1
44 bitcast i8* %add.ptr12 to <16 x i32>*
45 %add.ptr14 = getelementptr i8, i8* %src, i32 %mul
46 bitcast i8* %add.ptr14 to <16 x i32>*
47 bitcast i8* %dst to <16 x i32>*
48 load <16 x i32>, <16 x i32>* %0load <16 x i32>, <16 x i32>* %1load <16 x i32>, <16 x i32>* %2load <16 x i32>, <16 x i32>* %3load <16 x i32>, <16 x i32>* %4load <16 x i32>, <16 x i32>* %5load <16 x i32>, <16 x i32>* %6call <16 x i32> @llvm.hexagon.V6.vd0()
49 call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %15, <16 x i32> %15)
50 call <32 x i32> @llvm.hexagon.V6.vaddubh(<16 x i32> %14, <16 x i32> %8)
51 call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %13, <16 x i32> %9)
52 call <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32> %17, <32 x i32> %18, i32 101058054)
53 call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %12, <16 x i32> %10)
54 call <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32> %19, <32 x i32> %20, i32 252645135)
55 call <32 x i32> @llvm.hexagon.V6.vmpybus.acc(<32 x i32> %21, <16 x i32> %11, i32 336860180)
56 %cmp155 = icmp sgt i32 %width, 64
57 br i1 %cmp155, label %for.body.preheader, label %for.end
58 for.body.preheader: %incdec.ptr20 = getelementptr i8, i8* %add.ptr14%23 = bitcast i8* %incdec.ptr20 to <16 x i32>*
59 %incdec.ptr19 = getelementptr i8, i8* %add.ptr12%24 = bitcast i8* %incdec.ptr19 to <16 x i32>*
60 %incdec.ptr18 = getelementptr i8, i8* %add.ptr10%25 = bitcast i8* %incdec.ptr18 to <16 x i32>*
61 %incdec.ptr17 = getelementptr i8, i8* %src%26 = bitcast i8* %incdec.ptr17 to <16 x i32>*
62 %incdec.ptr16 = getelementptr i8, i8* %add.ptr6%27 = bitcast i8* %incdec.ptr16 to <16 x i32>*
63 %incdec.ptr15 = getelementptr i8, i8* %add.ptr3%28 = bitcast i8* %incdec.ptr15 to <16 x i32>*
64 %incdec.ptr = getelementptr i8, i8* %add.ptr%29 = bitcast i8* %incdec.ptr to <16 x i32>*
65 br label %for.body
66 for.body: %optr.0166 = phi <16 x i32>* [ %incdec.ptr28, %for.body ], [ %7, %for.body.preheader ]
67 %iptr6.0165 = phi <16 x i32>* [ %incdec.ptr27, %for.body ], [ %23, %for.body.preheader ]
68 %iptr5.0164 = phi <16 x i32>* [ %incdec.ptr26, %for.body ], [ %24, %for.body.preheader ]
69 %iptr4.0163 = phi <16 x i32>* [ %incdec.ptr25, %for.body ], [ %25, %for.body.preheader ]
70 %iptr3.0162 = phi <16 x i32>* [ %incdec.ptr24, %for.body ], [ %26, %for.body.preheader ]
71 %iptr2.0161 = phi <16 x i32>* [ %incdec.ptr23, %for.body ], [ %27, %for.body.preheader ]
72 %iptr1.0160 = phi <16 x i32>* [ %incdec.ptr22, %for.body ], [ %28, %for.body.preheader ]
73 %iptr0.0159 = phi <16 x i32>* [ %incdec.ptr21, %for.body ], [ %29, %for.body.preheader ]
74 %dXV1.0158 = phi <32 x i32> [ %49, %for.body ], [ %22, %for.body.preheader ]
75 %dXV0.0157 = phi <32 x i32> [ %dXV1.0158, %for.body ], [ %16, %for.body.preheader ]
76 %i.0156 = phi i32 [ %sub, %for.body ], [ %width, %for.body.preheader ]
77 %incdec.ptr21 = getelementptr <16 x i32>, <16 x i32>* %iptr0.0159%30 = load <16 x i32>, <16 x i32>* %iptr0.0159%incdec.ptr22 = getelementptr <16 x i32>, <16 x i32>* %iptr1.0160%31 = load <16 x i32>, <16 x i32>* %iptr1.0160%incdec.ptr23 = getelementptr <16 x i32>, <16 x i32>* %iptr2.0161%32 = load <16 x i32>, <16 x i32>* %iptr2.0161%incdec.ptr24 = getelementptr <16 x i32>, <16 x i32>* %iptr3.0162%33 = load <16 x i32>, <16 x i32>* %iptr3.0162%incdec.ptr25 = getelementptr <16 x i32>, <16 x i32>* %iptr4.0163%34 = load <16 x i32>, <16 x i32>* %iptr4.0163%incdec.ptr26 = getelementptr <16 x i32>, <16 x i32>* %iptr5.0164%35 = load <16 x i32>, <16 x i32>* %iptr5.0164%incdec.ptr27 = getelementptr <16 x i32>, <16 x i32>* %iptr6.0165%36 = load <16 x i32>, <16 x i32>* %iptr6.0165, !tbaa !8
78 call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %dXV1.0158)
79 call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %dXV0.0157)
80 call <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32> %37, <16 x i32> %38, i32 2)
81 call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %dXV1.0158)
82 call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %dXV0.0157)
83 call <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32> %40, <16 x i32> %41, i32 2)
84 call <16 x i32> @llvm.hexagon.V6.vlalignbi(<16 x i32> %37, <16 x i32> %38, i32 4)
85 call <32 x i32> @llvm.hexagon.V6.vaddubh(<16 x i32> %36, <16 x i32> %30)
86 call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %35, <16 x i32> %31)
87 call <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32> %44, <32 x i32> %45, i32 101058054)
88 call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %34, <16 x i32> %32)
89 call <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32> %46, <32 x i32> %47, i32 252645135)
90 call <32 x i32> @llvm.hexagon.V6.vmpybus.acc(<32 x i32> %48, <16 x i32> %33, i32 336860180)
91 call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %49)
92 call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %50, <16 x i32> %40, i32 2)
93 call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %49)
94 call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %52, <16 x i32> %37, i32 2)
95 call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %50, <16 x i32> %40, i32 4)
96 call <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32> %37, <16 x i32> %39)
97 call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %55, <16 x i32> %40)
98 call <32 x i32> @llvm.hexagon.V6.vmpahb(<32 x i32> %56, i32 252972820)
99 call <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32> %51, <16 x i32> %40)
100 call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %58, <16 x i32> %37)
101 call <32 x i32> @llvm.hexagon.V6.vmpahb(<32 x i32> %59, i32 252972820)
102 call <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32> %53, <16 x i32> %43)
103 call <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32> %51, <16 x i32> %42)
104 call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %61, <16 x i32> %62)
105 call <32 x i32> @llvm.hexagon.V6.vmpahb.acc(<32 x i32> %57, <32 x i32> %63, i32 17170694)
106 call <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32> %54, <16 x i32> %42)
107 call <16 x i32> @llvm.hexagon.V6.vaddh(<16 x i32> %53, <16 x i32> %39)
108 call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %65, <16 x i32> %66)
109 call <32 x i32> @llvm.hexagon.V6.vmpahb.acc(<32 x i32> %60, <32 x i32> %67, i32 17170694)
110 call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %64)
111 call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %64)
112 call <16 x i32> @llvm.hexagon.V6.vasrwh(<16 x i32> %69, <16 x i32> %70, i32 12)
113 call <16 x i32> @llvm.hexagon.V6.hi(<32 x i32> %68)
114 call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %68)
115 call <16 x i32> @llvm.hexagon.V6.vasrwh(<16 x i32> %72, <16 x i32> %73, i32 12)
116 call <16 x i32> @llvm.hexagon.V6.vshuffeb(<16 x i32> %74, <16 x i32> %71)
117 %incdec.ptr28 = getelementptr <16 x i32>, <16 x i32>* %1
118 store <16 x i32> %75, <16 x i32>* %optr.0166%sub = add i32 %i.0156, -64
119 %cmp = icmp sgt i32 %sub, 64
120 br i1 %cmp, label %for.body, label %for.end
121 for.end: ret void
122 }
123 declare <32 x i32> @llvm.hexagon.V6.vmpahb(<32 x i32>, i32)
124 declare <32 x i32> @llvm.hexagon.V6.vmpahb.acc(<32 x i32>, <32 x i32>, i32)
125 declare <16 x i32> @llvm.hexagon.V6.vasrwh(<16 x i32>, <16 x i32>, i32)
126 declare <16 x i32> @llvm.hexagon.V6.vshuffeb(<16 x i32>, <16 x i32>)
128 attributes #0 = { "correctly-rounded-divide-sqrt-fp-math""target-cpu"="hexagonv65" "target-features"="+hvx-length64b,+hvxv65,+v65,-long-calls" "unsafe-fp-math"}
129 !8 = !{!9, !9, i64 0}
130 !9 = !{!"omnipotent char", !10}
131 !10 = !{}
132 !14 = !{}
133 !19 = !{}
134 !24 = !{}