bank-conflict.mir 5.71 KB
# RUN: llc -march=hexagon -run-pass post-RA-sched %s -o - | FileCheck %s

# Test that the Post RA scheduler does not schedule back-to-back loads
# when there is another instruction to schedule. The scheduler avoids
# the back-to-back loads to reduce potential bank conflicts.

# CHECK: = L2_loadrigp
# CHECK: = A2_tfr
# CHECK: = L2_loadrigp

# CHECK: = L4_loadri_rr
# CHECK: = S2_tstbit_i
# CHECK: = L4_loadri_rr

--- |
  %s.0 = type { [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [24 x i32], [3 x i32], [24 x i32], [8 x %s.1], [5 x i32] }
  %s.1 = type { i32, i32 }

  @g0 = global i64 0
  @g1 = global i64 0
  @g2 = global i32 0
  @g3 = global i32 0
  @g4 = global i8 0

  declare i32 @llvm.hexagon.S2.cl0(i32) #0
  declare i32 @llvm.hexagon.S2.setbit.r(i32, i32) #0
  declare i64 @llvm.hexagon.M2.vmpy2s.s0(i32, i32) #0
  declare i64 @llvm.hexagon.M2.vmac2s.s0(i64, i32, i32) #0
  declare i64 @llvm.hexagon.A2.vaddws(i64, i64) #0
  declare i32 @llvm.hexagon.A4.modwrapu(i32, i32) #0

  define void @f0(i32 %a0) {
  b0:
    %v0 = bitcast [10 x %s.0]* inttoptr (i32 -121502345 to [10 x %s.0]*) to [10 x %s.0]*
    br label %b1

  b1:                                               ; preds = %b5, %b0
    %v1 = phi i32 [ 0, %b0 ], [ %v28, %b5 ]
    %v2 = phi i32 [ 0, %b0 ], [ %v27, %b5 ]
    %v3 = load i32, i32* @g2, align 4
    %v4 = load i32, i32* @g3, align 8
    %v5 = and i32 %v4, %v3
    %v6 = getelementptr [10 x %s.0], [10 x %s.0]* %v0, i32 0, i32 %v2
    %v7 = bitcast %s.0* %v6 to %s.0*
    %v8 = getelementptr %s.0, %s.0* %v7, i32 0, i32 12
    %v9 = getelementptr %s.0, %s.0* %v7, i32 0, i32 13
    br label %b2

  b2:                                               ; preds = %b4, %b1
    %v10 = phi i64 [ %v24, %b4 ], [ 0, %b1 ]
    %v11 = phi i32 [ %v13, %b4 ], [ %v5, %b1 ]
    %v12 = tail call i32 @llvm.hexagon.S2.cl0(i32 %v11)
    %v13 = tail call i32 @llvm.hexagon.S2.setbit.r(i32 %v11, i32 %v12)
    %v14 = getelementptr [24 x i32], [24 x i32]* %v8, i32 0, i32 %v12
    %v15 = load i32, i32* %v14, align 4
    %v16 = tail call i64 @llvm.hexagon.M2.vmpy2s.s0(i32 %v15, i32 %v15)
    %v17 = getelementptr [24 x i32], [24 x i32]* %v9, i32 0, i32 %v12
    %v18 = load i32, i32* %v17, align 4
    %v19 = tail call i64 @llvm.hexagon.M2.vmac2s.s0(i64 %v16, i32 %v18, i32 %v18)
    %v20 = load i8, i8* @g4, align 1
    %v21 = and i8 %v20, 1
    %v22 = icmp eq i8 %v21, 0
    br i1 %v22, label %b3, label %b4

  b3:                                               ; preds = %b2
    %v23 = tail call i64 @llvm.hexagon.A2.vaddws(i64 %v10, i64 %v19)
    store i64 %v23, i64* @g0, align 8
    br label %b4

  b4:                                               ; preds = %b3, %b2
    %v24 = phi i64 [ %v23, %b3 ], [ %v10, %b2 ]
    %v25 = icmp eq i32 %v13, 0
    br i1 %v25, label %b5, label %b2

  b5:                                               ; preds = %b4
    %v26 = add i32 %v2, 1
    %v27 = tail call i32 @llvm.hexagon.A4.modwrapu(i32 %v26, i32 10)
    %v28 = add i32 %v1, 1
    %v29 = icmp eq i32 %v28, %a0
    br i1 %v29, label %b6, label %b1

  b6:                                               ; preds = %b5
    store i64 %v19, i64* @g1, align 8
    ret void
  }

  attributes #0 = { nounwind readnone }

...
---
name:            f0
alignment:       16
tracksRegLiveness: true
registers:
liveins:
  - { reg: '$r0', virtual-reg: '' }
fixedStack:
stack:
constants:
body:             |
  bb.0:
    successors: %bb.1(0x80000000)
    liveins: $r0:0x00000001

    $r3 = A2_tfrsi 0
    $r2 = A2_tfrsi -121502345
    $r4 = A2_tfrsi 10
    J2_loop0r %bb.1, killed $r0, implicit-def $lc0, implicit-def $sa0, implicit-def $usr

  bb.1 (address-taken):
    successors: %bb.2(0x80000000)
    liveins: $lc0:0x00000004, $r2:0x00000001, $r3:0x00000001, $r4:0x00000001, $sa0:0x00000004

    $r5 = M2_mpysip $r3, 1824
    $r7 = L2_loadrigp @g2, implicit $gp :: (dereferenceable load 4 from @g2)
    $r8 = L2_loadrigp @g3, implicit killed $gp :: (dereferenceable load 4 from @g3, align 8)
    $r6 = A2_tfr $r5
    $r7 = A2_and killed $r8, killed $r7
    $r5 = M2_accii killed $r5, $r2, 1248
    $r6 = M2_accii killed $r6, $r2, 1152
    $d0 = A2_tfrpi 0

  bb.2:
    successors: %bb.3(0x04000000), %bb.2(0x7c000000)
    liveins: $lc0:0x00000004, $r0:0x00000001, $r1:0x00000001, $r2:0x00000001, $r3:0x00000001, $r4:0x00000001, $r5:0x00000001, $r6:0x00000001, $r7:0x00000001, $sa0:0x00000004

    $r8 = S2_cl0 $r7
    $r12 = L2_loadrubgp @g4, implicit $gp :: (dereferenceable load 1 from @g4)
    $r7 = S2_setbit_r killed $r7, $r8
    $r9 = L4_loadri_rr $r6, $r8, 2 :: (load 4 from %ir.v14)
    $r13 = L4_loadri_rr $r5, killed $r8, 2 :: (load 4 from %ir.v17)
    $d4 = M2_vmpy2s_s0 killed $r9, $r9, implicit-def dead $usr_ovf
    $p0 = S2_tstbit_i killed $r12, 0
    $d4 = M2_vmac2s_s0 killed $d4, killed $r13, $r13, implicit-def dead $usr_ovf
    $p1 = C2_cmpeqi $r7, 0
    $d6 = A2_vaddws $d0, $d4, implicit-def dead $usr_ovf
    $d0 = A2_tfrpt $p0, killed $d0, implicit $d0
    S4_pstorerdf_abs $p0, @g0, $d6, implicit killed $gp :: (store 8 into @g0)
    $d0 = A2_tfrpf killed $p0, killed $d6, implicit killed $d0
    J2_jumpf killed $p1, %bb.2, implicit-def dead $pc

  bb.3:
    successors: %bb.4(0x04000000), %bb.1(0x7c000000)
    liveins: $lc0:0x00000004, $r2:0x00000001, $r3:0x00000001, $r4:0x00000001, $r8:0x00000001, $r9:0x00000001, $sa0:0x00000004

    $r3 = A2_addi killed $r3, 1
    $r3 = A4_modwrapu killed $r3, $r4
    ENDLOOP0 %bb.1, implicit-def $pc, implicit-def $lc0, implicit $sa0, implicit $lc0

  bb.4:
    liveins: $r8:0x00000001, $r9:0x00000001

    S2_storerdgp @g1, killed $d4, implicit killed $gp :: (store 8 into @g1)
    PS_jmpret killed $r31, implicit-def dead $pc
...