[PeepholeOptimizer] Fix infinite recusion in CopyRewriter.

[efriedma]

Summary

CopyRewriter::getNewSource calls itself, with no obvious limit to the amount of allowed recursion. Limit it to recursing once, which should catch the obvious cases without the potential infinite recursion.

Arbitrarily limiting the recursion to one isn't ideal, but I'm not sure what the intended limit is, and there aren't any regression tests for more complicated cases.

[efriedma]

Err, sorry, there's something wrong with my testcase... used the wrong version or something? Looking into it.

[efriedma]

Nevermind, we actually don't need this, I think. (I only managed to reproduce the issue on the 6.0 branch; I think the underlying problem might have been fixed in https://reviews.llvm.org/D40763 .)

lib/CodeGen/PeepholeOptimizer.cpp

 /// multiple sources for a given \p Def are found along the way, we found a
 /// PHI instructions that needs to be rewritten.
 /// TODO: HandleMultipleSources should be removed once we test PHI handling
 /// with coalescable copies.
 static RegSubRegPair
 getNewSource(MachineRegisterInfo *MRI, const TargetInstrInfo *TII,
              RegSubRegPair Def,
              const PeepholeOptimizer::RewriteMapTy &RewriteMap,
-             bool HandleMultipleSources = true) {
+             bool HandleMultipleSources = true,
+             bool Recursing = false) {
   RegSubRegPair LookupSrc(Def.Reg, Def.SubReg);
   while (true) {
     ValueTrackerResult Res = RewriteMap.lookup(LookupSrc);
     // If there are no entries on the map, LookupSrc is the new source.
     if (!Res.isValid())
       return LookupSrc;
 
     // There's only one source for this definition, keep searching...
     unsigned NumSrcs = Res.getNumSources();
     if (NumSrcs == 1) {
       LookupSrc.Reg = Res.getSrcReg(0);
       LookupSrc.SubReg = Res.getSrcSubReg(0);
       continue;
     }
 
     // TODO: Remove once multiple srcs w/ coalescable copies are supported.
     if (!HandleMultipleSources)
       break;
 
+    if (Recursing)
+      return LookupSrc;
+
     // Multiple sources, recurse into each source to find a new source
     // for it. Then, rewrite the PHI accordingly to its new edges.
     SmallVector<RegSubRegPair, 4> NewPHISrcs;
     for (unsigned i = 0; i < NumSrcs; ++i) {
       RegSubRegPair PHISrc(Res.getSrcReg(i), Res.getSrcSubReg(i));
       NewPHISrcs.push_back(
-          getNewSource(MRI, TII, PHISrc, RewriteMap, HandleMultipleSources));
+          getNewSource(MRI, TII, PHISrc, RewriteMap,
+                       HandleMultipleSources, true));
     }
 
     // Build the new PHI node and return its def register as the new source.
     MachineInstr &OrigPHI = const_cast<MachineInstr &>(*Res.getInst());
     MachineInstr &NewPHI = insertPHI(*MRI, *TII, NewPHISrcs, OrigPHI);
     DEBUG(dbgs() << "-- getNewSource\n");
     DEBUG(dbgs() << "   Replacing: " << OrigPHI);
     DEBUG(dbgs() << "        With: " << NewPHI);

test/CodeGen/AArch64/peephole-recursion.mir

+# RUN: llc -o - %s -run-pass=peephole-opt | FileCheck %s
+# Make sure this doesn't crash
+# CHECK: COPY %r0
+
+--- |
+  target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
+  target triple = "thumbv8--linux-gnueabi"
+  
+  define i32 @test() {
+  entry:
+    ret i32 0
+  }
+
+...
+---
+name:            test
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    successors: %bb.2(0x80000000)
+    liveins: $r0
+
+    $0:rgpr = COPY $r0
+    t2B %bb.2, 14, $noreg
+  
+  bb.1:
+    successors: %bb.18(0x30000000), %bb.12(0x50000000)
+  
+    %7:rgpr = PHI $0, %bb.4, $10, %bb.2
+    t2CMPrr $0, $0, 14, $noreg, implicit-def $cpsr
+    t2Bcc %bb.18, 0, $cpsr
+    t2B %bb.12, 14, $noreg
+  
+  bb.2:
+    successors: %bb.1(0x30000000)
+  
+    $10:rgpr = PHI $0, %bb.0, $291, %bb.58
+    t2B %bb.1, 14, $noreg
+  
+  bb.4:
+    successors: %bb.1(0x7c000000)
+    t2B %bb.1, 14, $noreg
+  
+  bb.9:
+    %54:gpr = PHI $0, %bb.38, $271, %bb.59
+    %563:spr = VMOVSR %54, 14, $noreg
+    tBX_RET 14, $noreg
+  
+  bb.12:
+    successors: %bb.18(0x7c000000)
+  
+  bb.18:
+    successors: %bb.25(0x50000000)
+  
+    %91:rgpr = PHI $0, %bb.12, %7, %bb.1
+    t2B %bb.25, 14, $noreg
+  
+  bb.24:
+    successors: %bb.25(0x7c000000)
+    t2B %bb.25, 14, $noreg
+  
+  bb.25:
+    successors: %bb.26(0x50000000)
+    $127:rgpr = PHI $0, %bb.18, %91, %bb.24
+  
+  bb.26:
+    successors: %bb.32(0x7c000000)
+  
+    t2B %bb.32, 14, $noreg
+  
+  bb.31:
+    successors: %bb.32(0x7c000000)
+  
+  bb.32:
+    successors: %bb.39(0x50000000)
+    $163:rgpr = PHI $0, %bb.26, $127, %bb.31
+    t2B %bb.39, 14, $noreg
+  
+  bb.38:
+    successors: %bb.39(0x7c000000), %bb.9(0x04000000)
+  
+    t2CMPrr $0, $0, 14, $noreg, implicit-def $cpsr
+    t2Bcc %bb.9, 1, $cpsr
+  
+  bb.39:
+    successors: %bb.46(0x50000000)
+  
+    $199:rgpr = PHI $0, %bb.32, $163, %bb.38
+    t2B %bb.46, 14, $noreg
+  
+  bb.45:
+    successors: %bb.46(0x7c000000)
+  
+  bb.46:
+    successors: %bb.47(0x50000000)
+  
+    $235:rgpr = PHI $0, %bb.39, $199, %bb.45
+  
+  bb.47:
+    successors: %bb.53(0x7c000000)
+  
+    t2B %bb.53, 14, $noreg
+  
+  bb.52:
+    successors: %bb.53(0x7c000000)
+  
+  bb.53:
+    successors: %bb.59(0x30000000), %bb.54(0x50000000)
+  
+    $271:rgpr = PHI $0, %bb.47, $235, %bb.52
+    t2CMPrr $0, $0, 14, $noreg, implicit-def $cpsr
+    t2Bcc %bb.59, 0, $cpsr
+  
+  bb.54:
+    successors: %bb.57(0x40000000), %bb.58(0x40000000)
+  
+    t2CMPrr $0, $0, 14, $noreg, implicit-def $cpsr
+    t2Bcc %bb.58, 9, $cpsr
+  
+  bb.57:
+    successors: %bb.58(0x80000000)
+  
+  bb.58:
+    successors: %bb.2(0x7c000000)
+  
+    $291:gpr = PHI $271, %bb.54, $0, %bb.57
+    t2B %bb.2, 14, $noreg
+  
+  bb.59:
+    successors: %bb.9(0x04000000)
+  
+    t2B %bb.9, 14, $noreg
+
+...