1; RUN: llc -march=hexagon -mcpu=hexagonv60 -enable-bsb-sched=0 -enable-pipeliner < %s | FileCheck %s 2; RUN: llc -march=hexagon -mcpu=hexagonv5 -enable-pipeliner < %s | FileCheck %s 3 4; From coremark. Test that we pipeline the matrix multiplication bitextract 5; function. The pipelined code should have two packets. 6 7; CHECK: loop0(.LBB0_[[LOOP:.]], 8; CHECK: .LBB0_[[LOOP]]: 9; CHECK: = extractu([[REG2:(r[0-9]+)]], 10; CHECK: = extractu([[REG2]], 11; CHECK: [[REG0:(r[0-9]+)]] = memh 12; CHECK: [[REG1:(r[0-9]+)]] = memh 13; CHECK: += mpyi 14; CHECK: [[REG2]] = mpyi([[REG0]], [[REG1]]) 15; CHECK: endloop0 16 17%union_h2_sem_t = type { i32 } 18 19@sem_i = common global [0 x %union_h2_sem_t] zeroinitializer, align 4 20 21define void @matrix_mul_matrix_bitextract(i32 %N, i32* %C, i16* %A, i16* %B) { 22entry: 23 %cmp53 = icmp eq i32 %N, 0 24 br i1 %cmp53, label %for_end27, label %for_body3_lr_ph_us 25 26for_body3_lr_ph_us: 27 %i_054_us = phi i32 [ %inc26_us, %for_cond1_for_inc25_crit_edge_us ], [ 0, %entry ] 28 %0 = mul i32 %i_054_us, %N 29 %arrayidx9_us_us_gep = getelementptr i16, i16* %A, i32 %0 30 br label %for_body3_us_us 31 32for_cond1_for_inc25_crit_edge_us: 33 %inc26_us = add i32 %i_054_us, 1 34 %exitcond89 = icmp eq i32 %inc26_us, %N 35 br i1 %exitcond89, label %for_end27, label %for_body3_lr_ph_us 36 37for_body3_us_us: 38 %j_052_us_us = phi i32 [ %inc23_us_us, %for_cond4_for_inc22_crit_edge_us_us ], [ 0, %for_body3_lr_ph_us ] 39 %add_us_us = add i32 %j_052_us_us, %0 40 %arrayidx_us_us = getelementptr inbounds i32, i32* %C, i32 %add_us_us 41 store i32 0, i32* %arrayidx_us_us, align 4 42 br label %for_body6_us_us 43 44for_cond4_for_inc22_crit_edge_us_us: 45 store i32 %add21_us_us, i32* %arrayidx_us_us, align 4 46 %inc23_us_us = add i32 %j_052_us_us, 1 47 %exitcond88 = icmp eq i32 %inc23_us_us, %N 48 br i1 %exitcond88, label %for_cond1_for_inc25_crit_edge_us, label %for_body3_us_us 49 50for_body6_us_us: 51 %1 = phi i32 [ 0, %for_body3_us_us ], [ %add21_us_us, %for_body6_us_us ] 52 %arrayidx9_us_us_phi = phi i16* [ %arrayidx9_us_us_gep, %for_body3_us_us ], [ %arrayidx9_us_us_inc, %for_body6_us_us ] 53 %k_050_us_us = phi i32 [ 0, %for_body3_us_us ], [ %inc_us_us, %for_body6_us_us ] 54 %2 = load i16, i16* %arrayidx9_us_us_phi, align 2 55 %conv_us_us = sext i16 %2 to i32 56 %mul10_us_us = mul i32 %k_050_us_us, %N 57 %add11_us_us = add i32 %mul10_us_us, %j_052_us_us 58 %arrayidx12_us_us = getelementptr inbounds i16, i16* %B, i32 %add11_us_us 59 %3 = load i16, i16* %arrayidx12_us_us, align 2 60 %conv13_us_us = sext i16 %3 to i32 61 %mul14_us_us = mul nsw i32 %conv13_us_us, %conv_us_us 62 %shr47_us_us = lshr i32 %mul14_us_us, 2 63 %and_us_us = and i32 %shr47_us_us, 15 64 %shr1548_us_us = lshr i32 %mul14_us_us, 5 65 %and16_us_us = and i32 %shr1548_us_us, 127 66 %mul17_us_us = mul i32 %and_us_us, %and16_us_us 67 %add21_us_us = add i32 %mul17_us_us, %1 68 %inc_us_us = add i32 %k_050_us_us, 1 69 %exitcond87 = icmp eq i32 %inc_us_us, %N 70 %arrayidx9_us_us_inc = getelementptr i16, i16* %arrayidx9_us_us_phi, i32 1 71 br i1 %exitcond87, label %for_cond4_for_inc22_crit_edge_us_us, label %for_body6_us_us 72 73for_end27: 74 ret void 75} 76