75 lines · plain
1; RUN: llc -mtriple=hexagon -mcpu=hexagonv60 -enable-pipeliner -terminal-rule=0 < %s | FileCheck %s2 3; From coremark. Test that we pipeline the matrix multiplication bitextract4; function. The pipelined code should have two packets.5 6; CHECK: loop0(.LBB0_[[LOOP:.]],7; CHECK: .LBB0_[[LOOP]]:8; CHECK: [[REG0:(r[0-9]+)]] = mpyi([[REG1:(r[0-9]+)]],[[REG2:(r[0-9]+)]])9; CHECK: += mpyi10; CHECK: [[REG1:(r[0-9]+)]] = memh11; CHECK: = extractu([[REG0:(r[0-9]+)]],12; CHECK: = extractu([[REG0]],13; CHECK: [[REG2:(r[0-9]+)]] = memh14; CHECK: endloop015 16%union_h2_sem_t = type { i32 }17 18@sem_i = common global [0 x %union_h2_sem_t] zeroinitializer, align 419 20define void @matrix_mul_matrix_bitextract(i32 %N, ptr %C, ptr %A, ptr %B) {21entry:22 %cmp53 = icmp eq i32 %N, 023 br i1 %cmp53, label %for_end27, label %for_body3_lr_ph_us24 25for_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, %N28 %arrayidx9_us_us_gep = getelementptr i16, ptr %A, i32 %029 br label %for_body3_us_us30 31for_cond1_for_inc25_crit_edge_us:32 %inc26_us = add i32 %i_054_us, 133 %exitcond89 = icmp eq i32 %inc26_us, %N34 br i1 %exitcond89, label %for_end27, label %for_body3_lr_ph_us35 36for_body3_us_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, %039 %arrayidx_us_us = getelementptr inbounds i32, ptr %C, i32 %add_us_us40 store i32 0, ptr %arrayidx_us_us, align 441 br label %for_body6_us_us42 43for_cond4_for_inc22_crit_edge_us_us:44 store i32 %add21_us_us, ptr %arrayidx_us_us, align 445 %inc23_us_us = add i32 %j_052_us_us, 146 %exitcond88 = icmp eq i32 %inc23_us_us, %N47 br i1 %exitcond88, label %for_cond1_for_inc25_crit_edge_us, label %for_body3_us_us48 49for_body6_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 254 %conv_us_us = sext i16 %2 to i3255 %mul10_us_us = mul i32 %k_050_us_us, %N56 %add11_us_us = add i32 %mul10_us_us, %j_052_us_us57 %arrayidx12_us_us = getelementptr inbounds i16, ptr %B, i32 %add11_us_us58 %3 = load i16, ptr %arrayidx12_us_us, align 259 %conv13_us_us = sext i16 %3 to i3260 %mul14_us_us = mul nsw i32 %conv13_us_us, %conv_us_us61 %shr47_us_us = lshr i32 %mul14_us_us, 262 %and_us_us = and i32 %shr47_us_us, 1563 %shr1548_us_us = lshr i32 %mul14_us_us, 564 %and16_us_us = and i32 %shr1548_us_us, 12765 %mul17_us_us = mul i32 %and_us_us, %and16_us_us66 %add21_us_us = add i32 %mul17_us_us, %167 %inc_us_us = add i32 %k_050_us_us, 168 %exitcond87 = icmp eq i32 %inc_us_us, %N69 %arrayidx9_us_us_inc = getelementptr i16, ptr %arrayidx9_us_us_phi, i32 170 br i1 %exitcond87, label %for_cond4_for_inc22_crit_edge_us_us, label %for_body6_us_us71 72for_end27:73 ret void74}75