Skip to content

Commit 2771ea4

Browse files
quic-santdasquic-santdas
authored
[Hexagon] Fix concat lowering for HVX for 64B vector length (#98318)
When concatenation of vector instructions is formed, as a part of it vector rotation is performed. The direction of the shift was not correctly calculated. This fixes the rotation factor.
1 parent 6d40580 commit 2771ea4

File tree

2 files changed

+136
-1
lines changed

2 files changed

+136
-1
lines changed

‎llvm/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1748,7 +1748,7 @@ HexagonTargetLowering::LowerHvxConcatVectors(SDValue Op, SelectionDAG &DAG)
17481748

17491749
unsigned InpLen = ty(Op.getOperand(0)).getVectorNumElements();
17501750
MVT ByteTy = MVT::getVectorVT(MVT::i8, HwLen);
1751-
SDValue S = DAG.getConstant(InpLen*BitBytes, dl, MVT::i32);
1751+
SDValue S = DAG.getConstant(HwLen - InpLen*BitBytes, dl, MVT::i32);
17521752
SDValue Res = getZero(dl, ByteTy, DAG);
17531753
for (unsigned i = 0, e = Prefixes.size(); i != e; ++i) {
17541754
Res = DAG.getNode(HexagonISD::VROR, dl, ByteTy, Res, S);

‎llvm/test/CodeGen/Hexagon/hvx-concat-lower.ll

Lines changed: 135 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,135 @@
1+
; During lowering of HVX instruction for 64B vector, the rotation
2+
; direction for VROR (as part of concat of vectors lowering) is fixed.
3+
4+
; RUN: llc -march=hexagon -O2 %s -o - | FileCheck %s
5+
6+
; CHECK: vec.epilog.ph
7+
; CHECK: r{{.*}} = {{.*}}#48
8+
; CHECK: vec.epilog.vector.body
9+
10+
11+
%struct.str = type { i8, i8, i8 }
12+
13+
define dso_local void @foo(i16* nocapture noundef writeonly %pOut, i16* nocapture noundef readonly %Coefs, %struct.str* nocapture noundef readonly %pQ, i32 noundef %Intra) local_unnamed_addr #0 {
14+
entry:
15+
%Coefs13 = ptrtoint i16* %Coefs to i32
16+
%pOut12 = ptrtoint i16* %pOut to i32
17+
%cmp10 = icmp slt i32 %Intra, 16
18+
br i1 %cmp10, label %iter.check, label %for.end
19+
20+
iter.check: ; preds = %entry
21+
%Q = getelementptr inbounds %struct.str, %struct.str* %pQ, i32 0, i32 0
22+
%0 = load i8, i8* %Q, align 1
23+
%conv3 = zext i8 %0 to i32
24+
%1 = sub nsw i32 0, %conv3
25+
%2 = sub i32 16, %Intra
26+
%min.iters.check = icmp ult i32 %2, 8
27+
br i1 %min.iters.check, label %for.body.preheader, label %vector.memcheck
28+
29+
vector.memcheck: ; preds = %iter.check
30+
%3 = shl i32 %Intra, 1
31+
%4 = add i32 %3, %pOut12
32+
%5 = add i32 %3, %Coefs13
33+
%6 = sub i32 %4, %5
34+
%diff.check = icmp ult i32 %6, 128
35+
br i1 %diff.check, label %for.body.preheader, label %vector.main.loop.iter.check
36+
37+
vector.main.loop.iter.check: ; preds = %vector.memcheck
38+
%min.iters.check14 = icmp ult i32 %2, 64
39+
br i1 %min.iters.check14, label %vec.epilog.ph, label %vector.ph
40+
41+
vector.ph: ; preds = %vector.main.loop.iter.check
42+
%n.vec = and i32 %2, -64
43+
%bd.spinsert = insertelement <32 x i32> poison, i32 %1, i64 0
44+
%bd.sp = shufflevector <32 x i32> %bd.spinsert, <32 x i32> poison, <32 x i32> zeroinitializer
45+
%bd.spinsert16 = insertelement <32 x i32> poison, i32 %conv3, i64 0
46+
%bd.sp17 = shufflevector <32 x i32> %bd.spinsert16, <32 x i32> poison, <32 x i32> zeroinitializer
47+
br label %vector.body
48+
49+
vector.body: ; preds = %vector.body, %vector.ph
50+
%index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
51+
%offset.idx = add i32 %index, %Intra
52+
%7 = getelementptr inbounds i16, i16* %Coefs, i32 %offset.idx
53+
%8 = bitcast i16* %7 to <32 x i16>*
54+
%wide.load = load <32 x i16>, <32 x i16>* %8, align 2
55+
%9 = getelementptr inbounds i16, i16* %7, i32 32
56+
%10 = bitcast i16* %9 to <32 x i16>*
57+
%wide.load15 = load <32 x i16>, <32 x i16>* %10, align 2
58+
%11 = icmp slt <32 x i16> %wide.load, zeroinitializer
59+
%12 = icmp slt <32 x i16> %wide.load15, zeroinitializer
60+
%13 = select <32 x i1> %11, <32 x i32> %bd.sp, <32 x i32> %bd.sp17
61+
%14 = select <32 x i1> %12, <32 x i32> %bd.sp, <32 x i32> %bd.sp17
62+
%15 = trunc <32 x i32> %13 to <32 x i16>
63+
%16 = trunc <32 x i32> %14 to <32 x i16>
64+
%17 = getelementptr inbounds i16, i16* %pOut, i32 %offset.idx
65+
%18 = bitcast i16* %17 to <32 x i16>*
66+
store <32 x i16> %15, <32 x i16>* %18, align 2
67+
%19 = getelementptr inbounds i16, i16* %17, i32 32
68+
%20 = bitcast i16* %19 to <32 x i16>*
69+
store <32 x i16> %16, <32 x i16>* %20, align 2
70+
%index.next = add nuw i32 %index, 64
71+
%21 = icmp eq i32 %index.next, %n.vec
72+
br i1 %21, label %middle.block, label %vector.body
73+
74+
middle.block: ; preds = %vector.body
75+
%cmp.n = icmp eq i32 %2, %n.vec
76+
br i1 %cmp.n, label %for.end, label %vec.epilog.iter.check
77+
78+
vec.epilog.iter.check: ; preds = %middle.block
79+
%ind.end24 = add i32 %n.vec, %Intra
80+
%n.vec.remaining = and i32 %2, 56
81+
%min.epilog.iters.check = icmp eq i32 %n.vec.remaining, 0
82+
br i1 %min.epilog.iters.check, label %for.body.preheader, label %vec.epilog.ph
83+
84+
vec.epilog.ph: ; preds = %vector.main.loop.iter.check, %vec.epilog.iter.check
85+
%vec.epilog.resume.val = phi i32 [ %n.vec, %vec.epilog.iter.check ], [ 0, %vector.main.loop.iter.check ]
86+
%n.vec23 = and i32 %2, -8
87+
%ind.end = add i32 %n.vec23, %Intra
88+
%bd.spinsert29 = insertelement <8 x i32> poison, i32 %1, i64 0
89+
%bd.sp30 = shufflevector <8 x i32> %bd.spinsert29, <8 x i32> poison, <8 x i32> zeroinitializer
90+
%bd.spinsert31 = insertelement <8 x i32> poison, i32 %conv3, i64 0
91+
%bd.sp32 = shufflevector <8 x i32> %bd.spinsert31, <8 x i32> poison, <8 x i32> zeroinitializer
92+
br label %vec.epilog.vector.body
93+
94+
vec.epilog.vector.body: ; preds = %vec.epilog.vector.body, %vec.epilog.ph
95+
%index26 = phi i32 [ %vec.epilog.resume.val, %vec.epilog.ph ], [ %index.next33, %vec.epilog.vector.body ]
96+
%offset.idx27 = add i32 %index26, %Intra
97+
%22 = getelementptr inbounds i16, i16* %Coefs, i32 %offset.idx27
98+
%23 = bitcast i16* %22 to <8 x i16>*
99+
%wide.load28 = load <8 x i16>, <8 x i16>* %23, align 2
100+
%24 = icmp slt <8 x i16> %wide.load28, zeroinitializer
101+
%25 = select <8 x i1> %24, <8 x i32> %bd.sp30, <8 x i32> %bd.sp32
102+
%26 = trunc <8 x i32> %25 to <8 x i16>
103+
%27 = getelementptr inbounds i16, i16* %pOut, i32 %offset.idx27
104+
%28 = bitcast i16* %27 to <8 x i16>*
105+
store <8 x i16> %26, <8 x i16>* %28, align 2
106+
%index.next33 = add nuw i32 %index26, 8
107+
%29 = icmp eq i32 %index.next33, %n.vec23
108+
br i1 %29, label %vec.epilog.middle.block, label %vec.epilog.vector.body
109+
110+
vec.epilog.middle.block: ; preds = %vec.epilog.vector.body
111+
%cmp.n25 = icmp eq i32 %2, %n.vec23
112+
br i1 %cmp.n25, label %for.end, label %for.body.preheader
113+
114+
for.body.preheader: ; preds = %vector.memcheck, %iter.check, %vec.epilog.iter.check, %vec.epilog.middle.block
115+
%i.011.ph = phi i32 [ %Intra, %iter.check ], [ %Intra, %vector.memcheck ], [ %ind.end24, %vec.epilog.iter.check ], [ %ind.end, %vec.epilog.middle.block ]
116+
br label %for.body
117+
118+
for.body: ; preds = %for.body.preheader, %for.body
119+
%i.011 = phi i32 [ %inc, %for.body ], [ %i.011.ph, %for.body.preheader ]
120+
%arrayidx = getelementptr inbounds i16, i16* %Coefs, i32 %i.011
121+
%30 = load i16, i16* %arrayidx, align 2
122+
%cmp1 = icmp slt i16 %30, 0
123+
%31 = select i1 %cmp1, i32 %1, i32 %conv3
124+
%conv4 = trunc i32 %31 to i16
125+
%arrayidx5 = getelementptr inbounds i16, i16* %pOut, i32 %i.011
126+
store i16 %conv4, i16* %arrayidx5, align 2
127+
%inc = add i32 %i.011, 1
128+
%exitcond.not = icmp eq i32 %inc, 16
129+
br i1 %exitcond.not, label %for.end, label %for.body
130+
131+
for.end: ; preds = %for.body, %middle.block, %vec.epilog.middle.block, %entry
132+
ret void
133+
}
134+
135+
attributes #0 = { argmemonly nofree norecurse nosync nounwind "target-cpu"="hexagonv66" "target-features"="+hvx-length64b,+hvxv66,+v66" }

0 commit comments

Comments
 (0)