MC generated wrong code for label arithmetic

[llvmbot]

Bugzilla Link	22060
Resolution	FIXED
Resolved on	Feb 09, 2015 00:31
Version	trunk
OS	Windows NT
Blocks	#22748
Attachments	the sample
Reporter	LLVM Bugzilla Contributor
CC	@majnemer,@zmodem

Extended Description

see aa.s in attachment, then

llvm-mc -arch=x86 -triple=pc-windows-msvc -filetype=obj aa.s -o aa.obj

we disassemble the aa.obj and can found following

calll	L18$pb

L18$pb:
popl %eax
leal "?WndProc@@YGJPAUHWND__@@iij@Z"-L18$pb(%eax), %eax
movl 24(%ebp), %ecx
movl 20(%ebp), %edx
movl 16(%ebp), %esi
movl 12(%ebp), %edi
movl 8(%ebp), %ebx
movl %eax, 76(%esp) # 4-byte Spill
retl

.def	 "?WndProc@@YGJPAUHWND__@@IIJ@Z";
.scl	2;
.type	32;
.endef
.globl	"?WndProc@@YGJPAUHWND__@@IIJ@Z"

"?WndProc@@YGJPAUHWND__@@iij@Z": # @"\01?WndProc@@YGJPAUHWND__@@iij@Z"

BB#0:

pushl	%ebp
movl	%esp, %ebp

retl

is assembled to

.text:00000000 public ?SetUpWindow@@YAPAUHWND__@@PBD0HHPAUHINSTANCE__@@@z
.text:00000000 ?SetUpWindow@@YAPAUHWND__@@PBD0HHPAUHINSTANCE__@@@z proc near
.text:00000000
.text:00000000 arg_0 = dword ptr 8
.text:00000000 arg_4 = dword ptr 0Ch
.text:00000000 arg_8 = dword ptr 10h
.text:00000000 arg_C = dword ptr 14h
.text:00000000 arg_10 = dword ptr 18h
.text:00000000
.text:00000000 push ebp
.text:00000001 mov ebp, esp
.text:00000003
.text:00000003 loc_3:
.text:00000003 push ebx
.text:00000004 push edi
.text:00000005 push esi
.text:00000006 and esp, 0FFFFFFF8h
.text:00000009 sub esp, 80h
.text:0000000F call $+5
.text:00000014 pop eax
.text:00000015
.text:00000015 loc_15:
.text:00000015 lea eax, [eax+3]
.text:0000001B mov ecx, [ebp+arg_10]
.text:0000001E mov edx, [ebp+arg_C]
.text:00000021 mov esi, [ebp+arg_8]
.text:00000024 mov edi, [ebp+arg_4]
.text:00000027 mov ebx, [ebp+arg_0]
.text:0000002A mov [esp+4Ch], eax
.text:0000002E retn
.text:0000002E ?SetUpWindow@@YAPAUHWND__@@PBD0HHPAUHINSTANCE__@@@z endp ; sp-analysis failed

.text:00000015 lea eax, [eax+3]

obviuosly it is not the right code for:

leal "?WndProc@@YGJPAUHWND__@@iij@Z"-L18$pb(%eax), %eax

[llvmbot]

assigned to @majnemer

[zmodem]

Confirmed that this is a regression from 3.5 to 3.6 (and trunk).

objdump diff from 3.5 to 3.6:

@@ -14,7 +14,7 @@
9: 81 ec 80 00 00 00 sub $0x80,%esp
f: e8 00 00 00 00 call 14 <?SetUpWindow@@YAPAUHWND__@@PBD0HHPAUHINSTANCE__@@@z+0x14>
14: 58 pop %eax

• 15: 8d 80 1b 00 00 00 lea 0x1b(%eax),%eax

• 15: 8d 80 03 00 00 00 lea 0x3(%eax),%eax
  1b: 8b 4d 18 mov 0x18(%ebp),%ecx
  1e: 8b 55 14 mov 0x14(%ebp),%edx
  21: 8b 75 10 mov 0x10(%ebp),%esi

That certainly looks like a real problem.

[zmodem]

Bisection points to r221678.

David, can you take a look?

[majnemer]

Hi,

Can I please get some IR? I'm a little confused as to how we could generate this assembly.

[majnemer]

Fixed in r228565.

[zmodem]

mentioned in issue #22748