[regression] missed LTO oportunity

[llvmbot]

Bugzilla Link	11367
Resolution	WONTFIX
Resolved on	Nov 14, 2011 14:54
Version	trunk
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@asl,@lattner,@echristo,@tlattner

Extended Description

given 3 IL files:

test1.ll:

define i32 @​f() {
entry:
ret i32 42
}

test2.ll:

define available_externally i32 @​g() {
entry:
ret i32 42
}

test3.ll:

define i32 @​h() {
entry:
%call = tail call i32 @​g()
ret i32 %call
}
declare i32 @​g()

if we link and optimized them with

$ llvm-link test1.ll test2.ll test3.ll -o all.bc
$ opt -std-link-opts all.bc -o opt.ll -S

we get:

define i32 @​f() nounwind readnone {
entry:
ret i32 42
}
define i32 @​h() {
entry:
%call = tail call i32 @​g()
ret i32 %call
}
declare i32 @​g()

with 9af37a3 (143524) reverted, we get:

define i32 @​f() nounwind readnone {
entry:
ret i32 42
}
define i32 @​h() nounwind readnone {
entry:
ret i32 42
}

[lattner]

This seems completely theoretical to me. No frontend (that I'm aware of) would ever generate a dead available_externally function.

[llvmbot]

This seems completely theoretical to me. No frontend (that I'm aware of) would
ever generate a dead available_externally function.

I got to mostly the same conclusion and sent an email to the list. Sorry for splitting the discussion in two places.

[lattner]

Should this bug be closed?

[llvmbot]

I think so. It would be nice to still have the old API available, but we can discuss that on the list.