The following test case, compiled with -OO -target=x86_64-windows-win32, returns an incorrect value. It returns 5, when it should return 9. The problem is that the frame pointer that the first finally block receives is not passed onto the second finally block, but is regenerated using the localaddr intrinsic.
The test case is:
int
main() {
int Check = 0;
__try {
Check = 3;
} __finally {
__try {
Check += 2;
} __finally {
Check += 4;
}
}
return Check;
}The code generated with "-O0 --target=x86_64-windows-win32" is:
main: # @main
.seh_proc main
# %bb.0: # %entry
subq $40, %rsp
.seh_stackalloc 40
.seh_endprologue
xorl %ecx, %ecx
.set .Lmain$frame_escape_0, 32 <==== Check is at %rsp of main + 32
movl $0, 36(%rsp)
movl $0, 32(%rsp)
movl $3, 32(%rsp)
movq %rsp, %rax
movq %rax, %rdx <==== main's %rsp is passed to fin$0
callq "?fin$0@0@main@@"
movl 32(%rsp), %eax
addq $40, %rsp
retq
?fin$0@0@main@@:
subq $56, %rsp
leaq .Lmain$frame_escape_0(%rdx), %r8
movq %rdx, 48(%rsp)
movb %cl, 47(%rsp)
movl (%r8), %r9d <==== Check is at %rsp of main + 32
addl $2, %r9d
movq %rsp, %rdx <==== %rsp of fin$0 is passed to fin$1, should be %rsp of main
movl %eax, %ecx
callq "?fin$1@0@main@@"
"?fin$1@0@main@@": # @"?fin$1@0@main@@"
# %bb.0: # %entry
subq $16, %rsp
.seh_stackalloc 16
.seh_endprologue
leaq .Lmain$frame_escape_0(%rdx), %rax
movq %rdx, 8(%rsp)
movb %cl, 7(%rsp)
movl (%rax), %r8d
addl $4, %r8d
movl %r8d, (%rax)
addq $16, %rsp
retq
.seh_handlerdata
.textWith this change, we get the following:
"?fin$0@0@main@@": # @"?fin$0@0@main@@" .seh_proc "?fin$0@0@main@@" # %bb.0: # %entry subq $56, %rsp .seh_stackalloc 56 .seh_endprologue xorl %eax, %eax leaq .Lmain$frame_escape_0(%rdx), %r8 movq %rdx, 48(%rsp) movb %cl, 47(%rsp) movl (%r8), %r9d addl $2, %r9d movl %r9d, (%r8) movl %eax, %ecx callq "?fin$1@0@main@@" <==== %rsp of main is passed in %rdx onto fin$1 nop addq $56, %rsp retq
The change assumes that no compiler generated filter function can directly invoke a compiler generated finally function. I haven't been able to come up with a test case where this occurs. Otherwise, the check for IsOutlinedSEHHelper is insufficient. All SEH finally functions have two parameters, the second being the frame pointer. This is not the case for filter functions, as they differ on x86 vs x86_64.
Please use braces consistently.