[llvm][TailDuplicator] permit taildup'ing pred MBB containing INLINEASM_BR

[nickdesaulniers]

Summary

If we have:
bb0:

INLINEASM_BR ...
b bb2

bb1:

...

bb2:

ret

it would be nice to taildup that ret into bb0.

[efriedma]

llvm/lib/CodeGen/TailDuplicator.cpp

794

This doesn't seem right... the reasons we want to abort taildup are:

• There's a conditional branch, and we can't branch in the middle of a block.
• We can't merge two blocks that have different unwind destinations.
• We don't want multiple INLINEASM_BRs in a basic block, I think?

Checking !PredBB->mayHaveInlineAsmBr() shortcuts all those potential reasons.

---

Might be worth considering implementing a form of taildup that doesn't merge the duplicated block with the predecessor; that would work in more cases without tripping over weird edge cases.

llvm/lib/CodeGen/TailDuplicator.cpp

for (MachineBasicBlock *PredBB : Preds) {
     TDBBs.push_back(PredBB);
   }
   return Changed;
 }
 
 bool TailDuplicator::canTailDuplicate(MachineBasicBlock *TailBB,
                                       MachineBasicBlock *PredBB) {
   // EH edges are ignored by analyzeBranch.
-  if (PredBB->succ_size() > 1)
+  if (PredBB->succ_size() > 1 && !PredBB->mayHaveInlineAsmBr())
     return false;
 
   MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
   SmallVector<MachineOperand, 4> PredCond;
   if (TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond))
     return false;
   if (!PredCond.empty())
     return false;
for (MachineInstr &MI : llvm::make_early_inc_range(*TailBB)) {
       }
     }
     appendCopies(PredBB, CopyInfos, Copies);
 
     NumTailDupAdded += TailBB->size() - 1; // subtract one for removed branch
 
     // Update the CFG.
     PredBB->removeSuccessor(PredBB->succ_begin());
-    assert(PredBB->succ_empty() &&
+    assert((PredBB->succ_empty() || PredBB->mayHaveInlineAsmBr()) &&
            "TailDuplicate called on block with multiple successors!");
     for (MachineBasicBlock *Succ : TailBB->successors())
       PredBB->addSuccessor(Succ, MBPI->getEdgeProbability(TailBB, Succ));
 
     // Update branches in pred to jump to tail's layout successor if needed.
     if (ShouldUpdateTerminators)
       PredBB->updateTerminator(TailBB->getNextNode());
 

llvm/test/CodeGen/X86/callbr-asm-sink.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu | FileCheck %s
 
 ;; Verify that the machine instructions generated from the first
 ;; getelementptr don't get sunk below the callbr. (Reduced from a bug
 ;; report.)
 
 %struct1 = type { i8*, i32 }
 
 define void @klist_dec_and_del(%struct1*) {
 ; CHECK-LABEL: klist_dec_and_del:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    leaq 8(%rdi), %rax
 ; CHECK-NEXT:    #APP
 ; CHECK-NEXT:    # 8(%rdi) .LBB0_1
 ; CHECK-NEXT:    #NO_APP
-; CHECK-NEXT:  # %bb.2:
 ; CHECK-NEXT:    retq
 ; CHECK-NEXT:  .LBB0_1: # Block address taken
 ; CHECK-NEXT:    # Label of block must be emitted
 ; CHECK-NEXT:    movq $0, -8(%rax)
 ; CHECK-NEXT:    retq
   %2 = getelementptr inbounds %struct1, %struct1* %0, i64 0, i32 1
   callbr void asm sideeffect "# $0 $1", "*m,!i,~{memory},~{dirflag},~{fpsr},~{flags}"(i32* elementtype(i32) %2)
           to label %6 [label %3]

llvm/test/CodeGen/X86/callbr-asm.ll

normal1:                                          ; preds = %normal0
   ret i32 %1
 }
 
 ; Test 4 - asm-goto referenced with the 'l' (ell) modifier and not.
 define void @test4() {
 ; CHECK-LABEL: test4:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    #APP
-; CHECK-NEXT:    ja .LBB4_3
+; CHECK-NEXT:    ja .LBB4_2
 ; CHECK-NEXT:    #NO_APP
 ; CHECK-NEXT:  # %bb.1: # %asm.fallthrough
 ; CHECK-NEXT:    #APP
-; CHECK-NEXT:    ja .LBB4_3
+; CHECK-NEXT:    ja .LBB4_2
 ; CHECK-NEXT:    #NO_APP
-; CHECK-NEXT:  .LBB4_3: # Block address taken
+; CHECK-NEXT:    retl
+; CHECK-NEXT:  .LBB4_2: # Block address taken
 ; CHECK-NEXT:    # %quux
 ; CHECK-NEXT:    # Label of block must be emitted
 ; CHECK-NEXT:    retl
 entry:
   callbr void asm sideeffect "ja $0", "!i,~{dirflag},~{fpsr},~{flags}"()
           to label %asm.fallthrough [label %quux]
 
 asm.fallthrough:                                  ; preds = %entry

llvm/test/CodeGen/X86/inline-asm-pic.ll

 
 define void @x() {
 ; CHECK-LABEL: x:
 ; CHECK:       ## %bb.0:
 ; CHECK-NEXT:    ## InlineAsm Start
 ; CHECK-NEXT:    ## LBB1_1
 ; CHECK-EMPTY:
 ; CHECK-NEXT:    ## InlineAsm End
-; CHECK-NEXT:  ## %bb.2: ## %return
 ; CHECK-NEXT:    retl
 ; CHECK-NEXT:  Ltmp0: ## Block address taken
 ; CHECK-NEXT:  LBB1_1: ## %overflow
 ; CHECK-NEXT:    ## Label of block must be emitted
 ; CHECK-NEXT:    retl
   callbr void asm "#  ${0:l}\0A", "!i"()
           to label %return [label %overflow]