MachineScheduler: Reset next cluster candidate for each node (llvm#139513)

When a node is picked, we should reset its next cluster candidate to
null before releasing its successors/predecessors.

Author: ruiling

llvm/lib/CodeGen/MachineScheduler.cpp

@@ -967,6 +967,12 @@ void ScheduleDAGMI::releaseSucc(SUnit *SU, SDep *SuccEdge) {
 
 /// releaseSuccessors - Call releaseSucc on each of SU's successors.
 void ScheduleDAGMI::releaseSuccessors(SUnit *SU) {
+  // Reset the next successor, For example, we want to cluster A B C.
+  // After A is picked, we will set B as next cluster succ, but if we pick
+  // D instead of B after A, then we need to reset the next cluster succ because
+  // we have decided to not pick the cluster candidate B during pickNode().
+  // Leaving B as the NextClusterSucc just make things messy.
+  NextClusterSucc = nullptr;
   for (SDep &Succ : SU->Succs)
     releaseSucc(SU, &Succ);
 }
@@ -1004,6 +1010,7 @@ void ScheduleDAGMI::releasePred(SUnit *SU, SDep *PredEdge) {
 
 /// releasePredecessors - Call releasePred on each of SU's predecessors.
 void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
+  NextClusterPred = nullptr;
   for (SDep &Pred : SU->Preds)
     releasePred(SU, &Pred);
 }

llvm/test/CodeGen/AArch64/expand-select.ll

@@ -8,11 +8,11 @@ define void @foo(i32 %In1, <2 x i128> %In2, <2 x i128> %In3, ptr %Out) {
 ; CHECK-NEXT:    fmov s0, wzr
 ; CHECK-NEXT:    ldr x11, [sp]
 ; CHECK-NEXT:    fmov s1, w8
-; CHECK-NEXT:    ldp x9, x10, [sp, #8]
+; CHECK-NEXT:    ldp x8, x10, [sp, #8]
 ; CHECK-NEXT:    cmeq v0.4s, v1.4s, v0.4s
-; CHECK-NEXT:    fmov w8, s0
-; CHECK-NEXT:    tst w8, #0x1
-; CHECK-NEXT:    csel x8, x5, x9, ne
+; CHECK-NEXT:    fmov w9, s0
+; CHECK-NEXT:    tst w9, #0x1
+; CHECK-NEXT:    csel x8, x5, x8, ne
 ; CHECK-NEXT:    csel x9, x4, x11, ne
 ; CHECK-NEXT:    stp x9, x8, [x10, #16]
 ; CHECK-NEXT:    csel x8, x3, x7, ne
@@ -36,14 +36,14 @@ define void @bar(i32 %In1, <2 x i96> %In2, <2 x i96> %In3, ptr %Out) {
 ; CHECK-NEXT:    ldr x10, [sp, #16]
 ; CHECK-NEXT:    fmov s1, w8
 ; CHECK-NEXT:    cmeq v0.4s, v1.4s, v0.4s
-; CHECK-NEXT:    fmov w8, s0
-; CHECK-NEXT:    tst w8, #0x1
-; CHECK-NEXT:    ldp x9, x8, [sp]
+; CHECK-NEXT:    fmov w9, s0
+; CHECK-NEXT:    tst w9, #0x1
+; CHECK-NEXT:    ldp x8, x9, [sp]
 ; CHECK-NEXT:    csel x11, x2, x6, ne
 ; CHECK-NEXT:    str x11, [x10]
-; CHECK-NEXT:    csel x9, x4, x9, ne
-; CHECK-NEXT:    csel x8, x5, x8, ne
-; CHECK-NEXT:    stur x9, [x10, #12]
+; CHECK-NEXT:    csel x8, x4, x8, ne
+; CHECK-NEXT:    stur x8, [x10, #12]
+; CHECK-NEXT:    csel x8, x5, x9, ne
 ; CHECK-NEXT:    csel x9, x3, x7, ne
 ; CHECK-NEXT:    str w8, [x10, #20]
 ; CHECK-NEXT:    str w9, [x10, #8]

llvm/test/CodeGen/AArch64/extbinopload.ll

@@ -667,30 +667,30 @@ define <16 x i32> @extrause_load(ptr %p, ptr %q, ptr %r, ptr %s, ptr %z) {
 ; CHECK-NEXT:    add x10, x3, #12
 ; CHECK-NEXT:    bic v1.8h, #255, lsl #8
 ; CHECK-NEXT:    ld1 { v0.s }[3], [x3], #4
-; CHECK-NEXT:    ldr s3, [x0, #12]
-; CHECK-NEXT:    ldp s2, s7, [x0, #4]
+; CHECK-NEXT:    ldr s4, [x0, #12]
+; CHECK-NEXT:    ldp s5, s2, [x2, #4]
 ; CHECK-NEXT:    ldr s6, [x2, #12]
-; CHECK-NEXT:    ldp s5, s4, [x2, #4]
-; CHECK-NEXT:    ld1 { v3.s }[1], [x11]
+; CHECK-NEXT:    ldp s3, s7, [x0, #4]
+; CHECK-NEXT:    ld1 { v4.s }[1], [x11]
 ; CHECK-NEXT:    ld1 { v6.s }[1], [x10]
-; CHECK-NEXT:    ld1 { v2.s }[1], [x9]
-; CHECK-NEXT:    ld1 { v4.s }[1], [x8]
+; CHECK-NEXT:    ld1 { v2.s }[1], [x8]
 ; CHECK-NEXT:    ld1 { v5.s }[1], [x3]
 ; CHECK-NEXT:    add x8, x1, #8
+; CHECK-NEXT:    ld1 { v3.s }[1], [x9]
 ; CHECK-NEXT:    ld1 { v7.s }[1], [x8]
-; CHECK-NEXT:    uaddl v2.8h, v2.8b, v3.8b
-; CHECK-NEXT:    ushll v4.8h, v4.8b, #0
-; CHECK-NEXT:    uaddl v3.8h, v5.8b, v6.8b
+; CHECK-NEXT:    ushll v2.8h, v2.8b, #0
+; CHECK-NEXT:    uaddl v3.8h, v3.8b, v4.8b
+; CHECK-NEXT:    uaddl v4.8h, v5.8b, v6.8b
 ; CHECK-NEXT:    uaddw v1.8h, v1.8h, v7.8b
-; CHECK-NEXT:    uaddw2 v4.8h, v4.8h, v0.16b
-; CHECK-NEXT:    ushll v0.4s, v2.4h, #3
-; CHECK-NEXT:    ushll v5.4s, v3.4h, #3
+; CHECK-NEXT:    uaddw2 v2.8h, v2.8h, v0.16b
+; CHECK-NEXT:    ushll v0.4s, v3.4h, #3
+; CHECK-NEXT:    ushll v5.4s, v4.4h, #3
+; CHECK-NEXT:    ushll2 v4.4s, v4.8h, #3
 ; CHECK-NEXT:    ushll2 v3.4s, v3.8h, #3
-; CHECK-NEXT:    ushll2 v2.4s, v2.8h, #3
 ; CHECK-NEXT:    uaddw v0.4s, v0.4s, v1.4h
-; CHECK-NEXT:    uaddw2 v1.4s, v2.4s, v1.8h
-; CHECK-NEXT:    uaddw2 v3.4s, v3.4s, v4.8h
-; CHECK-NEXT:    uaddw v2.4s, v5.4s, v4.4h
+; CHECK-NEXT:    uaddw2 v1.4s, v3.4s, v1.8h
+; CHECK-NEXT:    uaddw2 v3.4s, v4.4s, v2.8h
+; CHECK-NEXT:    uaddw v2.4s, v5.4s, v2.4h
 ; CHECK-NEXT:    ret
   %lp1 = load <4 x i8>, ptr %p
   store <4 x i8> %lp1, ptr %z
@@ -1073,24 +1073,24 @@ define <16 x i32> @extrause_ext2(ptr %p, ptr %q, ptr %r, ptr %s, ptr %z) {
 ; CHECK-NEXT:    ld1 { v6.s }[1], [x10]
 ; CHECK-NEXT:    ld1 { v5.s }[1], [x9]
 ; CHECK-NEXT:    ld1 { v7.s }[1], [x8]
-; CHECK-NEXT:    uaddl v16.8h, v2.8b, v3.8b
-; CHECK-NEXT:    uaddl v3.8h, v1.8b, v6.8b
-; CHECK-NEXT:    uaddl v2.8h, v4.8b, v5.8b
+; CHECK-NEXT:    uaddl v2.8h, v2.8b, v3.8b
+; CHECK-NEXT:    uaddl v1.8h, v1.8b, v6.8b
+; CHECK-NEXT:    uaddl v3.8h, v4.8b, v5.8b
 ; CHECK-NEXT:    uaddl v4.8h, v0.8b, v7.8b
-; CHECK-NEXT:    ushll v0.4s, v16.4h, #3
-; CHECK-NEXT:    ushll2 v1.4s, v16.8h, #3
-; CHECK-NEXT:    ushll2 v18.4s, v16.8h, #0
-; CHECK-NEXT:    ushll v6.4s, v2.4h, #3
-; CHECK-NEXT:    ushll2 v7.4s, v2.8h, #3
-; CHECK-NEXT:    ushll2 v5.4s, v2.8h, #0
+; CHECK-NEXT:    ushll2 v0.4s, v2.8h, #0
+; CHECK-NEXT:    ushll v5.4s, v2.4h, #3
+; CHECK-NEXT:    ushll2 v16.4s, v2.8h, #3
+; CHECK-NEXT:    ushll v6.4s, v3.4h, #3
+; CHECK-NEXT:    ushll2 v7.4s, v3.8h, #3
 ; CHECK-NEXT:    ushll v17.4s, v2.4h, #0
-; CHECK-NEXT:    uaddw2 v1.4s, v1.4s, v3.8h
-; CHECK-NEXT:    uaddw v0.4s, v0.4s, v3.4h
+; CHECK-NEXT:    ushll2 v18.4s, v3.8h, #0
+; CHECK-NEXT:    ushll v19.4s, v3.4h, #0
+; CHECK-NEXT:    stp q17, q0, [x4]
+; CHECK-NEXT:    uaddw v0.4s, v5.4s, v1.4h
+; CHECK-NEXT:    uaddw2 v1.4s, v16.4s, v1.8h
 ; CHECK-NEXT:    uaddw2 v3.4s, v7.4s, v4.8h
 ; CHECK-NEXT:    uaddw v2.4s, v6.4s, v4.4h
-; CHECK-NEXT:    ushll v4.4s, v16.4h, #0
-; CHECK-NEXT:    stp q17, q5, [x4, #32]
-; CHECK-NEXT:    stp q4, q18, [x4]
+; CHECK-NEXT:    stp q19, q18, [x4, #32]
 ; CHECK-NEXT:    ret
   %lp1 = load <4 x i8>, ptr %p
   %p2 = getelementptr i8, ptr %p, i32 4
@@ -1176,19 +1176,20 @@ define <16 x i32> @extrause_shl(ptr %p, ptr %q, ptr %r, ptr %s, ptr %z) {
 ; CHECK-NEXT:    ld1 { v5.s }[1], [x9]
 ; CHECK-NEXT:    ld1 { v7.s }[1], [x8]
 ; CHECK-NEXT:    uaddl v2.8h, v2.8b, v3.8b
+; CHECK-NEXT:    uaddl v1.8h, v1.8b, v6.8b
 ; CHECK-NEXT:    uaddl v3.8h, v4.8b, v5.8b
-; CHECK-NEXT:    uaddl v4.8h, v1.8b, v6.8b
-; CHECK-NEXT:    ushll v5.4s, v2.4h, #3
-; CHECK-NEXT:    ushll2 v6.4s, v2.8h, #3
-; CHECK-NEXT:    uaddl v2.8h, v0.8b, v7.8b
-; CHECK-NEXT:    ushll v7.4s, v3.4h, #3
-; CHECK-NEXT:    ushll2 v16.4s, v3.8h, #3
-; CHECK-NEXT:    uaddw2 v1.4s, v6.4s, v4.8h
-; CHECK-NEXT:    uaddw v0.4s, v5.4s, v4.4h
-; CHECK-NEXT:    stp q5, q6, [x4]
-; CHECK-NEXT:    uaddw2 v3.4s, v16.4s, v2.8h
-; CHECK-NEXT:    uaddw v2.4s, v7.4s, v2.4h
-; CHECK-NEXT:    stp q7, q16, [x4, #32]
+; CHECK-NEXT:    uaddl v5.8h, v0.8b, v7.8b
+; CHECK-NEXT:    ushll v4.4s, v2.4h, #3
+; CHECK-NEXT:    ushll2 v2.4s, v2.8h, #3
+; CHECK-NEXT:    ushll v6.4s, v3.4h, #3
+; CHECK-NEXT:    ushll2 v7.4s, v3.8h, #3
+; CHECK-NEXT:    uaddw v0.4s, v4.4s, v1.4h
+; CHECK-NEXT:    uaddw2 v1.4s, v2.4s, v1.8h
+; CHECK-NEXT:    str q4, [x4]
+; CHECK-NEXT:    stp q2, q6, [x4, #16]
+; CHECK-NEXT:    uaddw2 v3.4s, v7.4s, v5.8h
+; CHECK-NEXT:    uaddw v2.4s, v6.4s, v5.4h
+; CHECK-NEXT:    str q7, [x4, #48]
 ; CHECK-NEXT:    ret
   %lp1 = load <4 x i8>, ptr %p
   %p2 = getelementptr i8, ptr %p, i32 4