1 ; RUN: llc -march=hexagon -mcpu=hexagonv60 -enable-pipeliner < %s | FileCheck %s
3 ; From coremark. Test that we pipeline the matrix multiplication bitextract
4 ; function. The pipelined code should have two packets.
6 ; CHECK: loop0(.LBB0_[[LOOP:.]],
7 ; CHECK: .LBB0_[[LOOP]]:
8 ; CHECK: [[REG0:(r[0-9]+)]] = mpyi([[REG1:(r[0-9]+)]],[[REG2:(r[0-9]+)]])
10 ; CHECK: [[REG1:(r[0-9]+)]] = memh
11 ; CHECK: = extractu([[REG0:(r[0-9]+)]],
12 ; CHECK: = extractu([[REG0]],
13 ; CHECK: [[REG2:(r[0-9]+)]] = memh
16 %union_h2_sem_t = type { i32 }
18 @sem_i = common global [0 x %union_h2_sem_t] zeroinitializer, align 4
20 define void @matrix_mul_matrix_bitextract(i32 %N, ptr %C, ptr %A, ptr %B) {
22 %cmp53 = icmp eq i32 %N, 0
23 br i1 %cmp53, label %for_end27, label %for_body3_lr_ph_us
26 %i_054_us = phi i32 [ %inc26_us, %for_cond1_for_inc25_crit_edge_us ], [ 0, %entry ]
27 %0 = mul i32 %i_054_us, %N
28 %arrayidx9_us_us_gep = getelementptr i16, ptr %A, i32 %0
29 br label %for_body3_us_us
31 for_cond1_for_inc25_crit_edge_us:
32 %inc26_us = add i32 %i_054_us, 1
33 %exitcond89 = icmp eq i32 %inc26_us, %N
34 br i1 %exitcond89, label %for_end27, label %for_body3_lr_ph_us
37 %j_052_us_us = phi i32 [ %inc23_us_us, %for_cond4_for_inc22_crit_edge_us_us ], [ 0, %for_body3_lr_ph_us ]
38 %add_us_us = add i32 %j_052_us_us, %0
39 %arrayidx_us_us = getelementptr inbounds i32, ptr %C, i32 %add_us_us
40 store i32 0, ptr %arrayidx_us_us, align 4
41 br label %for_body6_us_us
43 for_cond4_for_inc22_crit_edge_us_us:
44 store i32 %add21_us_us, ptr %arrayidx_us_us, align 4
45 %inc23_us_us = add i32 %j_052_us_us, 1
46 %exitcond88 = icmp eq i32 %inc23_us_us, %N
47 br i1 %exitcond88, label %for_cond1_for_inc25_crit_edge_us, label %for_body3_us_us
50 %1 = phi i32 [ 0, %for_body3_us_us ], [ %add21_us_us, %for_body6_us_us ]
51 %arrayidx9_us_us_phi = phi ptr [ %arrayidx9_us_us_gep, %for_body3_us_us ], [ %arrayidx9_us_us_inc, %for_body6_us_us ]
52 %k_050_us_us = phi i32 [ 0, %for_body3_us_us ], [ %inc_us_us, %for_body6_us_us ]
53 %2 = load i16, ptr %arrayidx9_us_us_phi, align 2
54 %conv_us_us = sext i16 %2 to i32
55 %mul10_us_us = mul i32 %k_050_us_us, %N
56 %add11_us_us = add i32 %mul10_us_us, %j_052_us_us
57 %arrayidx12_us_us = getelementptr inbounds i16, ptr %B, i32 %add11_us_us
58 %3 = load i16, ptr %arrayidx12_us_us, align 2
59 %conv13_us_us = sext i16 %3 to i32
60 %mul14_us_us = mul nsw i32 %conv13_us_us, %conv_us_us
61 %shr47_us_us = lshr i32 %mul14_us_us, 2
62 %and_us_us = and i32 %shr47_us_us, 15
63 %shr1548_us_us = lshr i32 %mul14_us_us, 5
64 %and16_us_us = and i32 %shr1548_us_us, 127
65 %mul17_us_us = mul i32 %and_us_us, %and16_us_us
66 %add21_us_us = add i32 %mul17_us_us, %1
67 %inc_us_us = add i32 %k_050_us_us, 1
68 %exitcond87 = icmp eq i32 %inc_us_us, %N
69 %arrayidx9_us_us_inc = getelementptr i16, ptr %arrayidx9_us_us_phi, i32 1
70 br i1 %exitcond87, label %for_cond4_for_inc22_crit_edge_us_us, label %for_body6_us_us