Inlined functions are not optimized as well as the same code copied in-line

[Gramner]

Bugzilla Link	43033
Version	trunk
OS	All
CC	@dwblaikie,@fhahn,@rotateright

Extended Description

Minimal sample code:

static inline int imin(int a, int b) {
return a < b ? a : b;
}

int fn1(int a, int b) {
if (a < b) __builtin_unreachable();
return imin(a, b);
}

int fn2(int a, int b) {
if (a < b) __builtin_unreachable();
return a < b ? a : b;
}

compiled with -O3 produces (Linux x86-64):

fn1:
mov eax, esi
cmp edi, esi
cmovle eax, edi
ret
fn2:
mov eax, esi
ret

Only in fn2 is the ternary operation correctly optimized away.

[fhahn]

This is an interesting example, thanks for sharing! I had a look, and here is what I think is going on:

After initial cfg simplifications and memory promotion, we get the following for fn1

; Function Attrs: nounwind ssp uwtable
define i32 @​fn1(i32 %a, i32 %b) #​0 {
entry:
%cmp = icmp slt i32 %a, %b
%0 = xor i1 %cmp, true
call void @​llvm.assume(i1 %0)
%call = call i32 @​imin(i32 %a, i32 %b)
ret i32 %call
}

and for fn2

define i32 @​fn2(i32 %a, i32 %b) #​0 {
entry:
%cmp = icmp slt i32 %a, %b
%0 = xor i1 %cmp, true
call void @​llvm.assume(i1 %0)
%cmp1 = icmp slt i32 %a, %b
%cond = select i1 %cmp1, i32 %a, i32 %b
ret i32 %cond
}

For fn2, early CSE will remove %cmp1, but for fn1, we do not, as @​imin is not inlined yet. Before it gets inlined, instcombine folds %cmp and %0 to %cmp = icmp sge %a, %b and after inlining @​imin, CSE won't do anything. We end up with the code below, and fail to use the fact that the assumed comparison is the negated comparison of the select.

define i32 @​fn1(i32 %a, i32 %b) local_unnamed_addr #​0 {
entry:
%cmp = icmp sge i32 %a, %b
call void @​llvm.assume(i1 %cmp)
%cmp.1 = icmp slt i32 %a, %b
%cond = select i1 %cmp.1, i32 %a, i32 %b
ret i32 %cond
}