Fix bug in Hexagon's AnalyzeBranch function

[djasper]

Summary

This was surfaced by r231064 which introduces more critical edge splitting.

This fixes llvm.org/PR22767.

[kparzysz]

This fix is incorrect. Please do not commit.

[djasper]

Could you elaborate a bit?

[kparzysz]

I'm actually preparing a better one right now.

The problem was caused by incompleteness of Hexagon's InsertBranch and RemoveBranch. AnalyzeBranch handled the combination of ENDLOOP0+J2_jump, which is a loop-end branch (conditional) followed by an unconditional branch, but none of the other two functions did.

The instruction ENDLOOP0 has only one explicit operand (the loop header), the other operands are implicit uses/defs of various reserved registers. Your patch picked up PC (program counter), and the branch insertion code interpreted a "Cond" vector with a register in it as a conditional branch, where the register is a predicate register. As a result, the instruction was "J2_jumpt %PC", which is not valid.

[djasper]

I see. Thank you!

[kparzysz]

I couldn't upload a new diff into your revision, so I created a new one:

http://reviews.llvm.org/D8418

[kparzysz]

Fix was committed in r232643.

lib/Target/Hexagon/HexagonInstrInfo.cpp

if (AllowModify)
       I->eraseFromParent();
     return false;
   }
 
   // If the block ends with an ENDLOOP, and JMP, handle it.
   if (SecLastOpcode == Hexagon::ENDLOOP0 &&
       LastOpcode == Hexagon::J2_jump) {
     TBB = SecondLastInst->getOperand(0).getMBB();
-    Cond.push_back(SecondLastInst->getOperand(0));
+    Cond.push_back(SecondLastInst->getOperand(1));
     FBB = LastInst->getOperand(0).getMBB();
     return false;
   }
 
   // Otherwise, can't handle this.
   return true;
 }
 

test/CodeGen/Hexagon/hwloop-cleanup.ll

-; RUN: llc -march=hexagon -mcpu=hexagonv4 < %s | FileCheck %s
+; RUN: llc -march=hexagon -mcpu=hexagonv4 -no-phi-elim-live-out-early-exit < %s | FileCheck %s
 ; Check that we remove the compare and induction variable instructions
 ; after generating hardware loops.
 ; Bug 6685.
 
 ; CHECK: loop0
 ; CHECK-NOT: r{{[0-9]+}}{{.}}={{.}}add(r{{[0-9]+}},{{.}}#-1)
 ; CHECK-NOT: cmp.eq
 ; CHECK: endloop0