1; RUN: llc < %s -march=x86-64 -mcpu=core2 -pre-RA-sched=source -enable-misched \
2; RUN:          -misched-topdown -verify-machineinstrs \
3; RUN:     | FileCheck %s -check-prefix=TOPDOWN
4; RUN: llc < %s -march=x86-64 -mcpu=core2 -pre-RA-sched=source -enable-misched \
5; RUN:          -misched=ilpmin -verify-machineinstrs \
6; RUN:     | FileCheck %s -check-prefix=ILPMIN
7; RUN: llc < %s -march=x86-64 -mcpu=core2 -pre-RA-sched=source -enable-misched \
8; RUN:          -misched=ilpmax -verify-machineinstrs \
9; RUN:     | FileCheck %s -check-prefix=ILPMAX
10;
11; Verify that the MI scheduler minimizes register pressure for a
12; uniform set of bottom-up subtrees (unrolled matrix multiply).
13;
14; For current top-down heuristics, ensure that some folded imulls have
15; been reordered with the stores. This tests the scheduler's cheap
16; alias analysis ability (that doesn't require any AliasAnalysis pass).
17;
18; TOPDOWN-LABEL: %for.body
19; TOPDOWN: movl %{{.*}}, (
20; TOPDOWN: imull {{[0-9]*}}(
21; TOPDOWN: movl %{{.*}}, 4(
22; TOPDOWN: imull {{[0-9]*}}(
23; TOPDOWN: movl %{{.*}}, 8(
24; TOPDOWN: movl %{{.*}}, 12(
25; TOPDOWN-LABEL: %for.end
26;
27; For -misched=ilpmin, verify that each expression subtree is
28; scheduled independently, and that the imull/adds are interleaved.
29;
30; ILPMIN-LABEL: %for.body
31; ILPMIN: movl %{{.*}}, (
32; ILPMIN: imull
33; ILPMIN: imull
34; ILPMIN: addl
35; ILPMIN: imull
36; ILPMIN: addl
37; ILPMIN: imull
38; ILPMIN: addl
39; ILPMIN: movl %{{.*}}, 4(
40; ILPMIN: imull
41; ILPMIN: imull
42; ILPMIN: addl
43; ILPMIN: imull
44; ILPMIN: addl
45; ILPMIN: imull
46; ILPMIN: addl
47; ILPMIN: movl %{{.*}}, 8(
48; ILPMIN: imull
49; ILPMIN: imull
50; ILPMIN: addl
51; ILPMIN: imull
52; ILPMIN: addl
53; ILPMIN: imull
54; ILPMIN: addl
55; ILPMIN: movl %{{.*}}, 12(
56; ILPMIN-LABEL: %for.end
57;
58; For -misched=ilpmax, verify that each expression subtree is
59; scheduled independently, and that the imull/adds are clustered.
60;
61; ILPMAX-LABEL: %for.body
62; ILPMAX: movl %{{.*}}, (
63; ILPMAX: imull
64; ILPMAX: imull
65; ILPMAX: imull
66; ILPMAX: imull
67; ILPMAX: addl
68; ILPMAX: addl
69; ILPMAX: addl
70; ILPMAX: movl %{{.*}}, 4(
71; ILPMAX: imull
72; ILPMAX: imull
73; ILPMAX: imull
74; ILPMAX: imull
75; ILPMAX: addl
76; ILPMAX: addl
77; ILPMAX: addl
78; ILPMAX: movl %{{.*}}, 8(
79; ILPMAX: imull
80; ILPMAX: imull
81; ILPMAX: imull
82; ILPMAX: imull
83; ILPMAX: addl
84; ILPMAX: addl
85; ILPMAX: addl
86; ILPMAX: movl %{{.*}}, 12(
87; ILPMAX-LABEL: %for.end
88
89define void @mmult([4 x i32]* noalias nocapture %m1, [4 x i32]* noalias nocapture %m2,
90[4 x i32]* noalias nocapture %m3) nounwind uwtable ssp {
91entry:
92  br label %for.body
93
94for.body:                              ; preds = %for.body, %entry
95  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
96  %arrayidx8 = getelementptr inbounds [4 x i32], [4 x i32]* %m1, i64 %indvars.iv, i64 0
97  %tmp = load i32, i32* %arrayidx8, align 4
98  %arrayidx12 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 0, i64 0
99  %tmp1 = load i32, i32* %arrayidx12, align 4
100  %arrayidx8.1 = getelementptr inbounds [4 x i32], [4 x i32]* %m1, i64 %indvars.iv, i64 1
101  %tmp2 = load i32, i32* %arrayidx8.1, align 4
102  %arrayidx12.1 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 1, i64 0
103  %tmp3 = load i32, i32* %arrayidx12.1, align 4
104  %arrayidx8.2 = getelementptr inbounds [4 x i32], [4 x i32]* %m1, i64 %indvars.iv, i64 2
105  %tmp4 = load i32, i32* %arrayidx8.2, align 4
106  %arrayidx12.2 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 2, i64 0
107  %tmp5 = load i32, i32* %arrayidx12.2, align 4
108  %arrayidx8.3 = getelementptr inbounds [4 x i32], [4 x i32]* %m1, i64 %indvars.iv, i64 3
109  %tmp6 = load i32, i32* %arrayidx8.3, align 4
110  %arrayidx12.3 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 3, i64 0
111  %tmp8 = load i32, i32* %arrayidx8, align 4
112  %arrayidx12.137 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 0, i64 1
113  %tmp9 = load i32, i32* %arrayidx12.137, align 4
114  %tmp10 = load i32, i32* %arrayidx8.1, align 4
115  %arrayidx12.1.1 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 1, i64 1
116  %tmp11 = load i32, i32* %arrayidx12.1.1, align 4
117  %tmp12 = load i32, i32* %arrayidx8.2, align 4
118  %arrayidx12.2.1 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 2, i64 1
119  %tmp13 = load i32, i32* %arrayidx12.2.1, align 4
120  %tmp14 = load i32, i32* %arrayidx8.3, align 4
121  %arrayidx12.3.1 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 3, i64 1
122  %tmp15 = load i32, i32* %arrayidx12.3.1, align 4
123  %tmp16 = load i32, i32* %arrayidx8, align 4
124  %arrayidx12.239 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 0, i64 2
125  %tmp17 = load i32, i32* %arrayidx12.239, align 4
126  %tmp18 = load i32, i32* %arrayidx8.1, align 4
127  %arrayidx12.1.2 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 1, i64 2
128  %tmp19 = load i32, i32* %arrayidx12.1.2, align 4
129  %tmp20 = load i32, i32* %arrayidx8.2, align 4
130  %arrayidx12.2.2 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 2, i64 2
131  %tmp21 = load i32, i32* %arrayidx12.2.2, align 4
132  %tmp22 = load i32, i32* %arrayidx8.3, align 4
133  %arrayidx12.3.2 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 3, i64 2
134  %tmp23 = load i32, i32* %arrayidx12.3.2, align 4
135  %tmp24 = load i32, i32* %arrayidx8, align 4
136  %arrayidx12.341 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 0, i64 3
137  %tmp25 = load i32, i32* %arrayidx12.341, align 4
138  %tmp26 = load i32, i32* %arrayidx8.1, align 4
139  %arrayidx12.1.3 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 1, i64 3
140  %tmp27 = load i32, i32* %arrayidx12.1.3, align 4
141  %tmp28 = load i32, i32* %arrayidx8.2, align 4
142  %arrayidx12.2.3 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 2, i64 3
143  %tmp29 = load i32, i32* %arrayidx12.2.3, align 4
144  %tmp30 = load i32, i32* %arrayidx8.3, align 4
145  %arrayidx12.3.3 = getelementptr inbounds [4 x i32], [4 x i32]* %m2, i64 3, i64 3
146  %tmp31 = load i32, i32* %arrayidx12.3.3, align 4
147  %tmp7 = load i32, i32* %arrayidx12.3, align 4
148  %mul = mul nsw i32 %tmp1, %tmp
149  %mul.1 = mul nsw i32 %tmp3, %tmp2
150  %mul.2 = mul nsw i32 %tmp5, %tmp4
151  %mul.3 = mul nsw i32 %tmp7, %tmp6
152  %mul.138 = mul nsw i32 %tmp9, %tmp8
153  %mul.1.1 = mul nsw i32 %tmp11, %tmp10
154  %mul.2.1 = mul nsw i32 %tmp13, %tmp12
155  %mul.3.1 = mul nsw i32 %tmp15, %tmp14
156  %mul.240 = mul nsw i32 %tmp17, %tmp16
157  %mul.1.2 = mul nsw i32 %tmp19, %tmp18
158  %mul.2.2 = mul nsw i32 %tmp21, %tmp20
159  %mul.3.2 = mul nsw i32 %tmp23, %tmp22
160  %mul.342 = mul nsw i32 %tmp25, %tmp24
161  %mul.1.3 = mul nsw i32 %tmp27, %tmp26
162  %mul.2.3 = mul nsw i32 %tmp29, %tmp28
163  %mul.3.3 = mul nsw i32 %tmp31, %tmp30
164  %add.1 = add nsw i32 %mul.1, %mul
165  %add.2 = add nsw i32 %mul.2, %add.1
166  %add.3 = add nsw i32 %mul.3, %add.2
167  %add.1.1 = add nsw i32 %mul.1.1, %mul.138
168  %add.2.1 = add nsw i32 %mul.2.1, %add.1.1
169  %add.3.1 = add nsw i32 %mul.3.1, %add.2.1
170  %add.1.2 = add nsw i32 %mul.1.2, %mul.240
171  %add.2.2 = add nsw i32 %mul.2.2, %add.1.2
172  %add.3.2 = add nsw i32 %mul.3.2, %add.2.2
173  %add.1.3 = add nsw i32 %mul.1.3, %mul.342
174  %add.2.3 = add nsw i32 %mul.2.3, %add.1.3
175  %add.3.3 = add nsw i32 %mul.3.3, %add.2.3
176  %arrayidx16 = getelementptr inbounds [4 x i32], [4 x i32]* %m3, i64 %indvars.iv, i64 0
177  store i32 %add.3, i32* %arrayidx16, align 4
178  %arrayidx16.1 = getelementptr inbounds [4 x i32], [4 x i32]* %m3, i64 %indvars.iv, i64 1
179  store i32 %add.3.1, i32* %arrayidx16.1, align 4
180  %arrayidx16.2 = getelementptr inbounds [4 x i32], [4 x i32]* %m3, i64 %indvars.iv, i64 2
181  store i32 %add.3.2, i32* %arrayidx16.2, align 4
182  %arrayidx16.3 = getelementptr inbounds [4 x i32], [4 x i32]* %m3, i64 %indvars.iv, i64 3
183  store i32 %add.3.3, i32* %arrayidx16.3, align 4
184  %indvars.iv.next = add i64 %indvars.iv, 1
185  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
186  %exitcond = icmp eq i32 %lftr.wideiv, 4
187  br i1 %exitcond, label %for.end, label %for.body
188
189for.end:                                        ; preds = %for.body
190  ret void
191}
192