Missed optimisation from flags on SUB

[llvmbot]

Bugzilla Link	5774
Version	trunk
OS	MacOS X
Reporter	LLVM Bugzilla Contributor
CC	@asl

Extended Description

Consider the following test case:

define i32 @foo1 ( i32 %a, i32 %b ) nounwind
{
entry:
	%sub = sub i32 %a, %b
	%diff = icmp slt i32 %a, %b
	%select = select i1 %diff, i32 %b, i32 %a
	%call = call i32 @bar ( i32 %sub, i32 %select )
	ret i32 %call
}

define i32 @foo2 ( i32 %a, i32 %b ) nounwind
{
entry:
	%sub = sub i32 %a, %b
	%diff = icmp slt i32 %sub, 0
	%select = select i1 %diff, i32 %b, i32 %a
	%call = call i32 @bar ( i32 %sub, i32 %select )
	ret i32 %call
}

declare i32 @bar ( i32 %a, i32 %b ) nounwind

Whilst the effect of both foo1 and foo2 are equivalent, foo2 uses the negative flag from the result of the sub for the condition; foo1 instead generates an extra compare instruction (on X86).

[llvmbot]

Consider also xor:

declare i32 @bar ( i32 %a, i32 %b ) nounwind

define i32 @foo1 ( i32 %a, i32 %b, i32 %c ) nounwind
{
entry:
        %xor = xor i32 %a, %b
        %cmp = icmp eq i32 %a, %b
        %sel = select i1 %cmp, i32 %a, i32 %c
        %call = call i32 @bar ( i32 %xor, i32 %sel ) nounwind
        ret i32 %call
}

define i32 @foo2 ( i32 %a, i32 %b, i32 %c ) nounwind
{
entry:
        %xor = xor i32 %a, %b
        %cmp = icmp eq i32 %xor, 0
        %sel = select i1 %cmp, i32 %a, i32 %c
        %call = call i32 @bar ( i32 %xor, i32 %sel ) nounwind
        ret i32 %call
}

The former generates code with one instruction fewer than the latter (on x86-64). The situation is actually worse here because -instcombine turns the latter into the former.

[llvmbot]

Whilst the effect of both foo1 and foo2 are equivalent, foo2 uses the negative
flag from the result of the sub for the condition; foo1 instead generates an
extra compare instruction (on X86).

Upon closer consideration, I was partially mistaken: this transformation is only valid if the sub is nsw.

[RKSimon]

The xor case is still failing:

define i32 @src(i32 %a, i32 %b, i32 %c)
{
  %xor = xor i32 %a, %b
  %cmp = icmp eq i32 %a, %b
  %sel = select i1 %cmp, i32 %c, i32 %xor
  ret i32 %sel
}
define i32 @tgt(i32 %a, i32 %b, i32 %c)
{
  %xor = xor i32 %a, %b
  %cmp = icmp eq i32 %xor, 0
  %sel = select i1 %cmp, i32 %c, i32 %xor
  ret i32 %sel
}

src:                                    # @src
        movl    %edi, %eax
        xorl    %esi, %eax
        cmpl    %esi, %edi
        cmovel  %edx, %eax
        retq
tgt:                                    # @tgt
        movl    %edi, %eax
        xorl    %esi, %eax
        cmovel  %edx, %eax
        retq