Teach CodeGenPrep to look past bitcast when it's duplicating return instruction

into predecessor blocks to enable tail call optimization.

rdar://11958338

llvm-svn: 160894
This commit is contained in:
Evan Cheng 2012-07-27 21:21:26 +00:00
parent 97e14e02f1
commit 249716e8ae
3 changed files with 121 additions and 7 deletions

View File

@ -645,10 +645,18 @@ bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) {
if (!TLI)
return false;
PHINode *PN = 0;
BitCastInst *BCI = 0;
Value *V = RI->getReturnValue();
PHINode *PN = V ? dyn_cast<PHINode>(V) : NULL;
if (V && !PN)
if (V) {
BCI = dyn_cast<BitCastInst>(V);
if (BCI)
V = BCI->getOperand(0);
PN = dyn_cast<PHINode>(V);
if (!PN)
return false;
}
BasicBlock *BB = RI->getParent();
if (PN && PN->getParent() != BB)
@ -666,6 +674,9 @@ bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) {
if (PN) {
BasicBlock::iterator BI = BB->begin();
do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
if (&*BI == BCI)
// Also skip over the bitcast.
++BI;
if (&*BI != RI)
return false;
} else {

View File

@ -659,10 +659,26 @@ ReturnInst *llvm::FoldReturnIntoUncondBranch(ReturnInst *RI, BasicBlock *BB,
// If the return instruction returns a value, and if the value was a
// PHI node in "BB", propagate the right value into the return.
for (User::op_iterator i = NewRet->op_begin(), e = NewRet->op_end();
i != e; ++i)
if (PHINode *PN = dyn_cast<PHINode>(*i))
if (PN->getParent() == BB)
i != e; ++i) {
Value *V = *i;
Instruction *NewBC = 0;
if (BitCastInst *BCI = dyn_cast<BitCastInst>(V)) {
// Return value might be bitcasted. Clone and insert it before the
// return instruction.
V = BCI->getOperand(0);
NewBC = BCI->clone();
Pred->getInstList().insert(NewRet, NewBC);
*i = NewBC;
}
if (PHINode *PN = dyn_cast<PHINode>(V)) {
if (PN->getParent() == BB) {
if (NewBC)
NewBC->setOperand(0, PN->getIncomingValueForBlock(Pred));
else
*i = PN->getIncomingValueForBlock(Pred);
}
}
}
// Update any PHI nodes in the returning block to realize that we no
// longer branch to them.

View File

@ -0,0 +1,87 @@
; RUN: llc < %s -mtriple=x86_64-apple-darwin | FileCheck %s
; Teach CGP to dup returns to enable tail call optimization.
; rdar://9147433
define i32 @foo(i32 %x) nounwind ssp {
; CHECK: foo:
entry:
switch i32 %x, label %return [
i32 1, label %sw.bb
i32 2, label %sw.bb1
i32 3, label %sw.bb3
i32 4, label %sw.bb5
i32 5, label %sw.bb7
i32 6, label %sw.bb9
]
sw.bb: ; preds = %entry
; CHECK: jmp _f1
%call = tail call i32 @f1() nounwind
br label %return
sw.bb1: ; preds = %entry
; CHECK: jmp _f2
%call2 = tail call i32 @f2() nounwind
br label %return
sw.bb3: ; preds = %entry
; CHECK: jmp _f3
%call4 = tail call i32 @f3() nounwind
br label %return
sw.bb5: ; preds = %entry
; CHECK: jmp _f4
%call6 = tail call i32 @f4() nounwind
br label %return
sw.bb7: ; preds = %entry
; CHECK: jmp _f5
%call8 = tail call i32 @f5() nounwind
br label %return
sw.bb9: ; preds = %entry
; CHECK: jmp _f6
%call10 = tail call i32 @f6() nounwind
br label %return
return: ; preds = %entry, %sw.bb9, %sw.bb7, %sw.bb5, %sw.bb3, %sw.bb1, %sw.bb
%retval.0 = phi i32 [ %call10, %sw.bb9 ], [ %call8, %sw.bb7 ], [ %call6, %sw.bb5 ], [ %call4, %sw.bb3 ], [ %call2, %sw.bb1 ], [ %call, %sw.bb ], [ 0, %entry ]
ret i32 %retval.0
}
declare i32 @f1()
declare i32 @f2()
declare i32 @f3()
declare i32 @f4()
declare i32 @f5()
declare i32 @f6()
; rdar://11958338
%0 = type opaque
declare i8* @bar(i8*) uwtable optsize noinline ssp
define hidden %0* @thingWithValue(i8* %self) uwtable ssp {
entry:
; CHECK: thingWithValue:
; CHECK: jmp _bar
br i1 undef, label %if.then.i, label %if.else.i
if.then.i: ; preds = %entry
br label %someThingWithValue.exit
if.else.i: ; preds = %entry
%call4.i = tail call i8* @bar(i8* undef) optsize
br label %someThingWithValue.exit
someThingWithValue.exit: ; preds = %if.else.i, %if.then.i
%retval.0.in.i = phi i8* [ undef, %if.then.i ], [ %call4.i, %if.else.i ]
%retval.0.i = bitcast i8* %retval.0.in.i to %0*
ret %0* %retval.0.i
}