[PATCH] Expand SimplifyCFG to convert certain simple switches to selects

[Marcello Maggioni]

================
Comment at: lib/Transforms/Utils/SimplifyCFG.cpp:3766
@@ +3765,3 @@
+// select.
+static Value *
+ConvertTwoCaseSwitch(const SmallVectorImpl<ConstantInt *> &FirstGroupCases,
----------------
Philip Reames wrote:
> It looks like this code is actually handling two cases + default.  Update the comment?
> 
> Also, is it always profitable to replace one switch with two selects?  I would guess not, but will defer to actual measurement.  
On X86 from the measurements I did with a very simple micro benchmark it seems profitable.

This is the code of the benchmark:

int foo1_with_def(int a) {
  switch(a) {
    case 10:
      return 10;
    case 20:
      return 2;
  }
  return 4;
}

int main() {
  int a = 0;
  int ret = 0;
  for (; a < 2000000000; ++a) {
    ret |= foo1_with_def(a);
  }

  return ret;
}

And the generated code for the functions containing the switch is (+ performance results):

LLVM without patch:
_foo1_with_def:
	movl	$10, %eax
	cmpl	$10, %edi
	je	LBB0_4
	cmpl	$20, %edi
	jne	LBB0_3
	movl	$2, %eax
	retq
LBB0_3:
	movl	$4, %eax
LBB0_4:
	retq

real	0m5.522s
user	0m5.519s
sys	0m0.002s

LLVM with patch:

_foo1_with_def:
	cmpl	$10, %edi
	movl	$4, %ecx
	cmovel	%edi, %ecx
	cmpl	$20, %edi
	movl	$2, %eax
	cmovnel	%ecx, %eax
	retq

real	0m3.233s
user	0m3.230s
sys	0m0.002s

I don't have any real world application though to back this up (it looks like a good result though).

In addition to that X86 is quite a control flow "friendly" architecture with respect to many others and if the results are positive on X86 I would expect that on some more custom architectures where control flow is more expensive this optimization might actually be even more effective.

================
Comment at: lib/Transforms/Utils/SimplifyCFG.cpp:3955
@@ +3954,3 @@
+                          (FirstPossibleHits + SecondPossibleHits)) != 0);
+    DefaultCanTrigger &= DefaultResult != nullptr;
+
----------------
Philip Reames wrote:
> It seems like proving the switch default is dead is a useful optimization even if we can make no other changes.  Should this be separated and generalized?
Do you mean making it a generic independent static function in SimplifyCFG.cpp ?

http://reviews.llvm.org/D4219