Sibcall optimization rejects sret-using functions

[llvmbot]

Bugzilla Link	8233
Version	trunk
OS	All
Reporter	LLVM Bugzilla Contributor
CC	@topperc,@RKSimon,@Kojoley,@rotateright

Extended Description

Sibcall optimization does not take place when 'sret' is used on X86-64. For example:

$ cat test.ll
; ModuleID = '/home/urxae/.ccache/test.tmp.s010625.15525.ii'
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
target triple = "x86_64-unknown-linux-gnu"

%struct.Foo = type { [3 x i64] }

define void @_Z10tailcallerv(%struct.Foo* sret %agg.result) nounwind {
entry:
  tail call void @_Z10tailcalleev(%struct.Foo* sret %agg.result) nounwind
  ret void
}

declare void @_Z10tailcalleev(%struct.Foo* sret) nounwind
$ llc test.ll -o -
	.file	"test.ll"
	.text
	.globl	_Z10tailcallerv
	.align	16, 0x90
	.type	_Z10tailcallerv,@function
_Z10tailcallerv:                        # @_Z10tailcallerv
# BB#0:                                 # %entry
	pushq	%rbx
	movq	%rdi, %rbx
	callq	_Z10tailcalleev
	movq	%rbx, %rax
	popq	%rbx
	ret
.Ltmp0:
	.size	_Z10tailcallerv, .Ltmp0-_Z10tailcallerv


	.section	.note.GNU-stack,"",@progbits

This despite the fact that the code for tailcaller() here only needs to be a jump to tailcallee()...

Removing sret from tailcaller() doesn't help, but removing it from both works.
(Note: removing it only from tailcallee() shouldn't work, because then tailcaller() can't copy %rdi to %rax on exit)

32-bit x86 code seems to have this problem as well, by the way.

[llvmbot]

rdar://13017006

[Kojoley]

Strangely, ARM64 is not affected, but ARM (32bit) does.

Some abandoned tries to fix the issue:
https://reviews.llvm.org/D45653
https://reviews.llvm.org/D46262

A simple C++ reproducer that is optimized by ICC (an not by Clang or GCC):

struct X { ~X() {} };

X bar();
X foo() {
return bar();
}

https://godbolt.org/z/_zW_bG

[Kojoley]

*** Bug llvm/llvm-bugzilla-archive#28417 has been marked as a duplicate of this bug. ***