From: "Chip Davis" Subject: Re: [PATCH v2] ntdll: Force align the stack in RtlRaiseException() on i386. Message-Id: Date: Sun, 12 Jan 2020 02:38:27 +0000 In-Reply-To: <2a432543-68a5-9d73-5bf3-5df6ae0bccdf@gmail.com> References: <2a432543-68a5-9d73-5bf3-5df6ae0bccdf@gmail.com> <20200111225140.6126-1-cdavis@codeweavers.com> January 11, 2020 5:34 PM, "Zebediah Figura" wrote: > Hello Chip, > > On 1/11/20 4:51 PM, Chip Davis wrote: > >> This is an assembly function, so we have to do it manually. >> >> Signed-off-by: Chip Davis >> --- >> >> Notes: >> v2: Fix saving EAX. Unfortunately, I need a register because of the 'subl' >> instruction. >> >> dlls/ntdll/signal_i386.c | 10 +++++++++- >> 1 file changed, 9 insertions(+), 1 deletion(-) >> >> diff --git a/dlls/ntdll/signal_i386.c b/dlls/ntdll/signal_i386.c >> index e9dd0de2fc3c..099e10a3ebf2 100644 >> --- a/dlls/ntdll/signal_i386.c >> +++ b/dlls/ntdll/signal_i386.c >> @@ -2548,14 +2548,22 @@ __ASM_STDCALL_FUNC( RtlRaiseException, 4, >> "movl %esp,%ebp\n\t" >> __ASM_CFI(".cfi_def_cfa_register %ebp\n\t") >> "leal -0x2cc(%esp),%esp\n\t" /* sizeof(CONTEXT) */ >> - "pushl %esp\n\t" /* context */ >> + "andl $-16,%esp\n\t" >> + "subl $16,%esp\n\t" >> + "movl %eax,(%esp)\n\t" >> + "movl %esp,%eax\n\t" >> + "subl $12,%esp\n\t" /* stack alignment */ >> + "pushl %eax\n\t" /* context */ > > Out of curiosity, why do you need to align the stack again before > calling RtlCaptureContext()? I don't see anything in our implementation > that I'm aware relies on an aligned stack. > > And if so, perhaps it'd be better to align the stack inside > RtlCaptureContext() instead of here? It isn't needed for RtlCaptureContext(). It's hand-hacked assembly and doesn't have any instructions in it that assume a 16-byte-aligned stack. I was seeing crashes in send_debug_event(), which I tracked down to the compiler generating an aligned SSE instruction (movaps) on the assumption that the stack was 16-byte aligned. For some reason, adding this here made the crashes go away. These embedded aligned SSE instructions are part of why we compile exported functions with force_align_arg_pointer in the first place. Here we have to do it manually, because of the hand-hacked assembly. > >> + "movl 16(%esp),%eax\n\t" >> "call " __ASM_STDCALL("RtlCaptureContext",4) "\n\t" >> + "addl $12,%esp\n\t" >> "movl 4(%ebp),%eax\n\t" /* return address */ >> "movl 8(%ebp),%ecx\n\t" /* rec */ >> "movl %eax,12(%ecx)\n\t" /* rec->ExceptionAddress */ >> "leal 12(%ebp),%eax\n\t" >> "movl %eax,0xc4(%esp)\n\t" /* context->Esp */ >> "movl %esp,%eax\n\t" >> + "subl $8,%esp\n\t" /* stack alignment */ >> "pushl %eax\n\t" >> "pushl %ecx\n\t" >> "call " __ASM_NAME("raise_exception_full_context") "\n\t" Chip