1; REQUIRES: asserts
2; RUN: llc < %s -mtriple=arm64-linux-gnu -mcpu=cortex-a53 -pre-RA-sched=source -enable-misched -verify-misched -debug-only=machine-scheduler -disable-machine-dce -o - 2>&1 > /dev/null | FileCheck %s
3; RUN: llc < %s -mtriple=arm64-linux-gnu -mcpu=cortex-a53 -pre-RA-sched=source -enable-misched -verify-misched -debug-only=machine-scheduler -disable-machine-dce -o - -misched-limit=2 2>&1 > /dev/null | FileCheck %s
4;
5; The Cortex-A53 machine model will cause the MADD instruction to be scheduled
6; much higher than the ADD instructions in order to hide latency. When not
7; specifying a subtarget, the MADD will remain near the end of the block.
8;
9; CHECK: ********** MI Scheduling **********
10; CHECK: main
11; CHECK: *** Final schedule for %bb.2 ***
12; CHECK: MADDWrrr
13; CHECK: ADDWri
14; CHECK: ********** INTERVALS **********
15@main.x = private unnamed_addr constant [8 x i32] [i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1], align 4
16@main.y = private unnamed_addr constant [8 x i32] [i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2], align 4
17
18; Function Attrs: nounwind
19define i32 @main() #0 {
20entry:
21  %retval = alloca i32, align 4
22  %x = alloca [8 x i32], align 4
23  %y = alloca [8 x i32], align 4
24  %i = alloca i32, align 4
25  %xx = alloca i32, align 4
26  %yy = alloca i32, align 4
27  store i32 0, i32* %retval
28  %0 = bitcast [8 x i32]* %x to i8*
29  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %0, i8* align 4 bitcast ([8 x i32]* @main.x to i8*), i64 32, i1 false)
30  %1 = bitcast [8 x i32]* %y to i8*
31  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %1, i8* align 4 bitcast ([8 x i32]* @main.y to i8*), i64 32, i1 false)
32  store i32 0, i32* %xx, align 4
33  store i32 0, i32* %yy, align 4
34  store i32 0, i32* %i, align 4
35  br label %for.cond
36
37for.cond:                                         ; preds = %for.inc, %entry
38  %2 = load i32, i32* %i, align 4
39  %cmp = icmp slt i32 %2, 8
40  br i1 %cmp, label %for.body, label %for.end
41
42for.body:                                         ; preds = %for.cond
43  %3 = load i32, i32* %i, align 4
44  %idxprom = sext i32 %3 to i64
45  %arrayidx = getelementptr inbounds [8 x i32], [8 x i32]* %x, i32 0, i64 %idxprom
46  %4 = load i32, i32* %arrayidx, align 4
47  %add = add nsw i32 %4, 1
48  store i32 %add, i32* %xx, align 4
49  %5 = load i32, i32* %xx, align 4
50  %add1 = add nsw i32 %5, 12
51  store i32 %add1, i32* %xx, align 4
52  %6 = load i32, i32* %xx, align 4
53  %add2 = add nsw i32 %6, 23
54  store i32 %add2, i32* %xx, align 4
55  %7 = load i32, i32* %xx, align 4
56  %add3 = add nsw i32 %7, 34
57  store i32 %add3, i32* %xx, align 4
58  %8 = load i32, i32* %i, align 4
59  %idxprom4 = sext i32 %8 to i64
60  %arrayidx5 = getelementptr inbounds [8 x i32], [8 x i32]* %y, i32 0, i64 %idxprom4
61  %9 = load i32, i32* %arrayidx5, align 4
62  %10 = load i32, i32* %yy, align 4
63  %mul = mul nsw i32 %10, %9
64  store i32 %mul, i32* %yy, align 4
65  br label %for.inc
66
67for.inc:                                          ; preds = %for.body
68  %11 = load i32, i32* %i, align 4
69  %inc = add nsw i32 %11, 1
70  store i32 %inc, i32* %i, align 4
71  br label %for.cond
72
73for.end:                                          ; preds = %for.cond
74  %12 = load i32, i32* %xx, align 4
75  %13 = load i32, i32* %yy, align 4
76  %add6 = add nsw i32 %12, %13
77  ret i32 %add6
78}
79
80
81; The Cortex-A53 machine model will cause the FDIVvvv_42 to be raised to
82; hide latency. Whereas normally there would only be a single FADDvvv_4s
83; after it, this test checks to make sure there are more than one.
84;
85; CHECK: ********** MI Scheduling **********
86; CHECK: neon4xfloat:%bb.0
87; CHECK: *** Final schedule for %bb.0 ***
88; CHECK: FDIVv4f32
89; CHECK: FADDv4f32
90; CHECK: FADDv4f32
91; CHECK: ********** INTERVALS **********
92define <4 x float> @neon4xfloat(<4 x float> %A, <4 x float> %B) {
93        %tmp1 = fadd <4 x float> %A, %B;
94        %tmp2 = fadd <4 x float> %A, %tmp1;
95        %tmp3 = fadd <4 x float> %A, %tmp2;
96        %tmp4 = fadd <4 x float> %A, %tmp3;
97        %tmp5 = fadd <4 x float> %A, %tmp4;
98        %tmp6 = fadd <4 x float> %A, %tmp5;
99        %tmp7 = fadd <4 x float> %A, %tmp6;
100        %tmp8 = fadd <4 x float> %A, %tmp7;
101        %tmp9 = fdiv <4 x float> %A, %B;
102        %tmp10 = fadd <4 x float> %tmp8, %tmp9;
103
104        ret <4 x float> %tmp10
105}
106
107; Function Attrs: nounwind
108declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture readonly, i64, i1) #1
109
110attributes #0 = { nounwind "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
111attributes #1 = { nounwind }
112
113
114; Regression Test for PR19761
115;   [ARM64] Cortex-a53 schedule mode can't handle NEON post-increment load
116;
117; Nothing explicit to check other than llc not crashing.
118define { <16 x i8>, <16 x i8> } @test_v16i8_post_imm_ld2(i8* %A, i8** %ptr) {
119  %ld2 = tail call { <16 x i8>, <16 x i8> } @llvm.aarch64.neon.ld2.v16i8.p0i8(i8* %A)
120  %tmp = getelementptr i8, i8* %A, i32 32
121  store i8* %tmp, i8** %ptr
122  ret { <16 x i8>, <16 x i8> } %ld2
123}
124
125declare { <16 x i8>, <16 x i8> } @llvm.aarch64.neon.ld2.v16i8.p0i8(i8*)
126
127; Regression Test for PR20057.
128;
129; Cortex-A53 machine model stalls on A53UnitFPMDS contention. Instructions that
130; are otherwise ready are jammed in the pending queue.
131; CHECK: ********** MI Scheduling **********
132; CHECK: testResourceConflict
133; CHECK: *** Final schedule for %bb.0 ***
134; CHECK: BRK
135; CHECK: ********** INTERVALS **********
136define void @testResourceConflict(float* %ptr) {
137entry:
138  %add1 = fadd float undef, undef
139  %mul2 = fmul float undef, undef
140  %add3 = fadd float %mul2, undef
141  %mul4 = fmul float undef, %add3
142  %add5 = fadd float %mul4, undef
143  %sub6 = fsub float 0.000000e+00, undef
144  %sub7 = fsub float %add5, undef
145  %div8 = fdiv float 1.000000e+00, undef
146  %mul9 = fmul float %div8, %sub7
147  %mul14 = fmul float %sub6, %div8
148  %mul10 = fsub float -0.000000e+00, %mul14
149  %mul15 = fmul float undef, %div8
150  %mul11 = fsub float -0.000000e+00, %mul15
151  %mul12 = fmul float 0.000000e+00, %div8
152  %mul13 = fmul float %add1, %mul9
153  %mul21 = fmul float %add5, %mul11
154  %add22 = fadd float %mul13, %mul21
155  store float %add22, float* %ptr, align 4
156  %mul28 = fmul float %add1, %mul10
157  %mul33 = fmul float %add5, %mul12
158  %add34 = fadd float %mul33, %mul28
159  store float %add34, float* %ptr, align 4
160  %mul240 = fmul float undef, %mul9
161  %add246 = fadd float %mul240, undef
162  store float %add246, float* %ptr, align 4
163  %mul52 = fmul float undef, %mul10
164  %mul57 = fmul float undef, %mul12
165  %add58 = fadd float %mul57, %mul52
166  store float %add58, float* %ptr, align 4
167  %mul27 = fmul float 0.000000e+00, %mul9
168  %mul81 = fmul float undef, %mul10
169  %add82 = fadd float %mul27, %mul81
170  store float %add82, float* %ptr, align 4
171  call void @llvm.trap()
172  unreachable
173}
174
175declare void @llvm.trap()
176
177; Regression test for PR20057: "permanent hazard"'
178; Resource contention on LDST.
179; CHECK: ********** MI Scheduling **********
180; CHECK: testLdStConflict
181; CHECK: *** Final schedule for %bb.1 ***
182; CHECK: LD4Fourv2d
183; CHECK: STRQui
184; CHECK: ********** INTERVALS **********
185define void @testLdStConflict(<2 x i64> %v) {
186entry:
187  br label %loop
188
189loop:
190  %0 = call { <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.aarch64.neon.ld4.v2i64.p0i8(i8* null)
191  %ptr = bitcast i8* undef to <2 x i64>*
192  store <2 x i64> %v, <2 x i64>* %ptr, align 4
193  %ptr1 = bitcast i8* undef to <2 x i64>*
194  store <2 x i64> %v, <2 x i64>* %ptr1, align 4
195  %ptr2 = bitcast i8* undef to <2 x i64>*
196  store <2 x i64> %v, <2 x i64>* %ptr2, align 4
197  %ptr3 = bitcast i8* undef to <2 x i64>*
198  store <2 x i64> %v, <2 x i64>* %ptr3, align 4
199  %ptr4 = bitcast i8* undef to <2 x i64>*
200  store <2 x i64> %v, <2 x i64>* %ptr4, align 4
201  br label %loop
202}
203
204declare { <2 x i64>, <2 x i64>, <2 x i64>, <2 x i64> } @llvm.aarch64.neon.ld4.v2i64.p0i8(i8*)
205