MacroFusion: Fix macro fusion with ExitSU failing in top-down scheduling

[MatzeB]

Summary

When fusing instructions A and B, we must add all predecessors of B as
predecessors of A to avoid instructions getting scheduling in between.

There is a special case involving ExitSU: Every other node must be
scheduled before it by design and we don't need to make this explicit in
the graph, however when fusing with a different node we need to schedule
every othere node before the fused node too and we need to make this
explicit now: This patch adds a dependency from the fused node to all
roots in the graph.

[fhahn]

LGTM, thanks

[evandro]

Thank you for this fix.

llvm/trunk/lib/CodeGen/MacroFusion.cpp

for (const SDep &SI : FirstSU.Succs) {
         continue;
       LLVM_DEBUG(dbgs() << "  Bind "; SecondSU.print(dbgs(), &DAG);
                  dbgs() << " - "; SU->print(dbgs(), &DAG); dbgs() << '\n';);
       DAG.addEdge(SU, SDep(&SecondSU, SDep::Artificial));
     }
 
   // Make the FirstSU also dependent on the dependencies of the SecondSU to
   // prevent them from being scheduled between the FirstSU and the SecondSU.
-  if (&FirstSU != &DAG.EntrySU)
+  if (&FirstSU != &DAG.EntrySU) {
     for (const SDep &SI : SecondSU.Preds) {
       SUnit *SU = SI.getSUnit();
       if (SI.isWeak() || isHazard(SI) || &FirstSU == SU || FirstSU.isSucc(SU))
         continue;
       LLVM_DEBUG(dbgs() << "  Bind "; SU->print(dbgs(), &DAG); dbgs() << " - ";
                  FirstSU.print(dbgs(), &DAG); dbgs() << '\n';);
       DAG.addEdge(&FirstSU, SDep(SU, SDep::Artificial));
     }
+    // ExitSU comes last by design, which acts like an implicit dependency
+    // between ExitSU and any bottom root in the graph. We should transfer
+    // this to FirstSU as well.
+    if (&SecondSU == &DAG.ExitSU) {
+      for (SUnit &SU : DAG.SUnits) {
+        if (SU.Succs.empty())
+          DAG.addEdge(&FirstSU, SDep(&SU, SDep::Artificial));
+      }
+    }
+  }
 
   ++NumFused;
   return true;
 }
 
 namespace {
 
 /// Post-process the DAG to create cluster edges between instrs that may

llvm/trunk/test/CodeGen/AArch64/macro-fusion-last.mir

+# RUN: llc -o - %s -mtriple=aarch64-- -mattr=+arith-bcc-fusion -run-pass postmisched | FileCheck %s --check-prefixes=CHECK,FUSION
+# RUN: llc -o - %s -mtriple=aarch64-- -mattr=-arith-bcc-fusion -run-pass postmisched | FileCheck %s --check-prefixes=CHECK,NOFUSION
+# Make sure the last instruction is correctly macro-fused when scheduling
+# top-down (post-ra).
+---
+# CHECK-LABEL: name: fuse_last
+# CHECK: $x1 = LDRXui $x0, 0
+# NOFUSION: $xzr = SUBSXri killed $x2, 0, 0, implicit-def $nzcv
+# CHECK: STRXui killed $x0, killed $x1, 0
+# FUSION: $xzr = SUBSXri killed $x2, 0, 0, implicit-def $nzcv
+# CHECK: Bcc 1, %bb.1, implicit killed $nzcv
+name: fuse_last
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $x0, $x2
+
+    $x1 = LDRXui $x0, 0
+    ; There is latency between these two instructions tempting the scheduler to
+    ; move the SUBSXri in between them. However doing so breaks macro fusion.
+    STRXui $x0, $x1, 0
+
+    $xzr = SUBSXri $x2, 0, 0, implicit-def $nzcv
+    Bcc 1, %bb.1, implicit killed $nzcv
+
+  bb.1:
+    RET_ReallyLR implicit undef $x0
+...