xref: /llvm-project/llvm/test/CodeGen/ARM/ParallelDSP/smlad8.ll (revision 9bd7b149c2f577086a716ccd0363d057caa3d98a)
1; RUN: opt -mtriple=armv8m.main-none-none-eabi -mattr=+dsp < %s -arm-parallel-dsp -S | FileCheck %s
2;
3; Mul with operands that are not simple load and sext/zext chains: this is not
4; yet supported so the rewrite shouldn't trigger (but we do want to support this
5; soon).
6;
7; CHECK-NOT:  call i32 @llvm.arm.smlad
8;
9define dso_local i32 @test(i32 %arg, ptr nocapture readnone %arg1, ptr nocapture readonly %arg2, ptr nocapture readonly %arg3, ptr %arg4) {
10entry:
11  %cmp24 = icmp sgt i32 %arg, 0
12  br i1 %cmp24, label %for.body.preheader, label %for.cond.cleanup
13
14for.body.preheader:
15  %.pre = load i16, ptr %arg3, align 2
16  %.pre27 = load i16, ptr %arg2, align 2
17  %gep1 = getelementptr inbounds i16, ptr %arg4, i32 1
18  %.add4 = load i16, ptr %arg4, align 2
19  %.add5 = load i16, ptr %gep1, align 2
20  %.zext4 = zext i16 %.add4 to i32
21  %.zext5 = zext i16 %.add5 to i32
22  br label %for.body
23
24for.cond.cleanup:
25  %mac1.0.lcssa = phi i32 [ 0, %entry ], [ %add11, %for.body ]
26  ret i32 %mac1.0.lcssa
27
28for.body:
29  %mac1.026 = phi i32 [ %add11, %for.body ], [ 0, %for.body.preheader ]
30  %i.025 = phi i32 [ %add, %for.body ], [ 0, %for.body.preheader ]
31  %arrayidx = getelementptr inbounds i16, ptr %arg3, i32 %i.025
32  %0 = load i16, ptr %arrayidx, align 2
33  %add = add nuw nsw i32 %i.025, 1
34  %arrayidx1 = getelementptr inbounds i16, ptr %arg3, i32 %add
35  %1 = load i16, ptr %arrayidx1, align 2
36  %arrayidx3 = getelementptr inbounds i16, ptr %arg2, i32 %i.025
37  %2 = load i16, ptr %arrayidx3, align 2
38  %conv = sext i16 %2 to i32
39  %conv4 = sext i16 %0 to i32
40  %add1 = add i32 %conv, %.zext4
41
42; This mul has a more complicated pattern as an operand, %add1
43; is another add and load, which we don't support for now.
44  %mul = mul nsw i32 %add1, %conv4
45  %arrayidx6 = getelementptr inbounds i16, ptr %arg2, i32 %add
46  %3 = load i16, ptr %arrayidx6, align 2
47  %conv7 = sext i16 %3 to i32
48  %conv8 = sext i16 %1 to i32
49  %add2 = add i32 %conv7, %.zext5
50
51; Same here
52  %mul9 = mul nsw i32 %add2, %conv8
53  %add10 = add i32 %mul, %mac1.026
54
55  %add11 = add i32 %mul9, %add10
56  %exitcond = icmp ne i32 %add, %arg
57  br i1 %exitcond, label %for.body, label %for.cond.cleanup
58}
59