SimplifyCFG may de-optimize code with highly predictable branches

[rotateright]

Bugzilla Link	24743
Version	trunk
OS	All
CC	@majnemer,@hfinkel

Extended Description

As mentioned in:
https://llvm.org/bugs/show_bug.cgi?id=23827#c7
SimplifyCFG (eg, FoldBranchToCommonDest()) transforms code based on a static "BonusInstThreshold". This can waste time because this change in the CFG may later be undone in CodeGenPrepare::splitBranchCondition() or SelectionDAGBuilder::visitBr().

Without external information (target hook, profile data, or programmer hint) to justify the transform, SimplifyCFG shouldn't be doing this:

$ cat predictable_branches.ll
define void @​x_or_y(i32 %x, i32 %y) #​0 {
entry:
%cmp = icmp slt i32 %x, 3
br i1 %cmp, label %if.then, label %lhs.false

lhs.false:
%cmp1 = icmp sgt i32 %y, 10
br i1 %cmp1, label %if.then, label %end

if.then:
%call = call i32 (...) @​foo() #​2
br label %end

end:
ret void
}

declare i32 @​foo(...) #​1

$ ./opt -simplifycfg predictable_branches.ll -S
define void @​x_or_y(i32 %x, i32 %y) {
entry:
%cmp = icmp slt i32 %x, 3
%cmp1 = icmp sgt i32 %y, 10
%or.cond = or i1 %cmp, %cmp1
br i1 %or.cond, label %if.then, label %end

if.then:
%call = call i32 (...) @​foo()
br label %end

end:
ret void
}

[majnemer]

I'm less sure that this is a pessimization. Turning that branch tree into an or of icmp permits follow-up transformation of the or in InstCombine.

[rotateright]

I'm less sure that this is a pessimization. Turning that branch tree into an
or of icmp permits follow-up transformation of the or in InstCombine.

Hi David -
Can you provide an example of this? I may be missing it, but I don't see any further optimization for this particular test. Maybe there's a way to limit SimplifyCFG to only try this transform if we think there's room for more optimization by other passes?