[PATCH][AArch64] PR71951: Fix unwinding with -fomit-frame-pointer

[PATCH][AArch64] PR71951: Fix unwinding with -fomit-frame-pointer

[Wilco Dijkstra]

As described in PR71951, if libgcc is built with -fomit-frame-pointer,
unwinding crashes, for example while doing a backtrace.  The underlying
reason is the Dwarf unwinder does not setup the frame pointer register
in the initialization code.  When later unwinding a function that uses
the frame pointer, it tries to read FP using _Unwind_GetGR, and this
crashes if has never restored FP.  To unwind correctly the first frame
must save and restore FP (it is unwound in a special way so that it
uses SP instead of FP).  This is done by adding -fno-omit-frame-pointer.

OK for commit and backport to GCC6/7?

ChangeLog:
2017-07-31  Wilco Dijkstra  <wdijkstr@arm.com>

	PR target/71951
	* config/aarch64/aarch64.h (LIBGCC2_UNWIND_ATTRIBUTE): Define.

--
diff --git a/gcc/config/aarch64/aarch64.h b/gcc/config/aarch64/aarch64.h
index 7f91edb5713d7e8eda2f0a024a0f97b4e111c4b0..03fd93046bdbdb03bd7d0c4573928f504640f7e1 100644
--- a/gcc/config/aarch64/aarch64.h
+++ b/gcc/config/aarch64/aarch64.h
@@ -971,4 +971,12 @@ extern const char *host_detect_local_cpu (int argc, const char **argv);
 extern tree aarch64_fp16_type_node;
 extern tree aarch64_fp16_ptr_type_node;
 
+/* The generic unwind code in libgcc does not initialize the frame pointer.
+   So in order to unwind a function using a frame pointer, the very first
+   function that is unwound must save the frame pointer.  That way the frame
+   pointer is restored and its value is now valid - otherwise _Unwind_GetGR
+   crashes.  Libgcc can now be safely built with -fomit-frame-pointer.  */
+#define LIBGCC2_UNWIND_ATTRIBUTE \
+  __attribute__((optimize ("no-omit-frame-pointer")))
+
 #endif /* GCC_AARCH64_H */

Re: [PATCH][AArch64] PR71951: Fix unwinding with -fomit-frame-pointer

[Wilco Dijkstra]

ping


From: Wilco Dijkstra
Sent: 31 July 2017 16:57
To: GCC Patches; James Greenhalgh
Cc: nd
Subject: [PATCH][AArch64] PR71951: Fix unwinding with -fomit-frame-pointer
    
As described in PR71951, if libgcc is built with -fomit-frame-pointer,
unwinding crashes, for example while doing a backtrace.  The underlying
reason is the Dwarf unwinder does not setup the frame pointer register
in the initialization code.  When later unwinding a function that uses
the frame pointer, it tries to read FP using _Unwind_GetGR, and this
crashes if has never restored FP.  To unwind correctly the first frame
must save and restore FP (it is unwound in a special way so that it
uses SP instead of FP).  This is done by adding -fno-omit-frame-pointer.

OK for commit and backport to GCC6/7?

ChangeLog:
2017-07-31  Wilco Dijkstra  <wdijkstr@arm.com>

        PR target/71951
        * config/aarch64/aarch64.h (LIBGCC2_UNWIND_ATTRIBUTE): Define.

--
diff --git a/gcc/config/aarch64/aarch64.h b/gcc/config/aarch64/aarch64.h
index 7f91edb5713d7e8eda2f0a024a0f97b4e111c4b0..03fd93046bdbdb03bd7d0c4573928f504640f7e1 100644
--- a/gcc/config/aarch64/aarch64.h
+++ b/gcc/config/aarch64/aarch64.h
@@ -971,4 +971,12 @@ extern const char *host_detect_local_cpu (int argc, const char **argv);
 extern tree aarch64_fp16_type_node;
 extern tree aarch64_fp16_ptr_type_node;
 
+/* The generic unwind code in libgcc does not initialize the frame pointer.
+   So in order to unwind a function using a frame pointer, the very first
+   function that is unwound must save the frame pointer.  That way the frame
+   pointer is restored and its value is now valid - otherwise _Unwind_GetGR
+   crashes.  Libgcc can now be safely built with -fomit-frame-pointer.  */
+#define LIBGCC2_UNWIND_ATTRIBUTE \
+  __attribute__((optimize ("no-omit-frame-pointer")))
+
 #endif /* GCC_AARCH64_H */
    

Re: [PATCH][AArch64] PR71951: Fix unwinding with -fomit-frame-pointer

[James Greenhalgh]

On Tue, Aug 15, 2017 at 05:27:37PM +0100, Wilco Dijkstra wrote:
> 
> ping

This seems like a bit of a theoretical issue as we would normally build
libgcc with -fno-omit-frame-pointer anyway, but it can't hurt to guarantee
this, so OK.

Thanks,
James

> From: Wilco Dijkstra
> Sent: 31 July 2017 16:57
> To: GCC Patches; James Greenhalgh
> Cc: nd
> Subject: [PATCH][AArch64] PR71951: Fix unwinding with -fomit-frame-pointer
>     
> As described in PR71951, if libgcc is built with -fomit-frame-pointer,
> unwinding crashes, for example while doing a backtrace.  The underlying
> reason is the Dwarf unwinder does not setup the frame pointer register
> in the initialization code.  When later unwinding a function that uses
> the frame pointer, it tries to read FP using _Unwind_GetGR, and this
> crashes if has never restored FP.  To unwind correctly the first frame
> must save and restore FP (it is unwound in a special way so that it
> uses SP instead of FP).  This is done by adding -fno-omit-frame-pointer.
> 
> OK for commit and backport to GCC6/7?
> 
> ChangeLog:
> 2017-07-31  Wilco Dijkstra  <wdijkstr@arm.com>
> 
>         PR target/71951
>         * config/aarch64/aarch64.h (LIBGCC2_UNWIND_ATTRIBUTE): Define.
> 
> --
> diff --git a/gcc/config/aarch64/aarch64.h b/gcc/config/aarch64/aarch64.h
> index 7f91edb5713d7e8eda2f0a024a0f97b4e111c4b0..03fd93046bdbdb03bd7d0c4573928f504640f7e1 100644
> --- a/gcc/config/aarch64/aarch64.h
> +++ b/gcc/config/aarch64/aarch64.h
> @@ -971,4 +971,12 @@ extern const char *host_detect_local_cpu (int argc, const char **argv);
>  extern tree aarch64_fp16_type_node;
>  extern tree aarch64_fp16_ptr_type_node;
>  
> +/* The generic unwind code in libgcc does not initialize the frame pointer.
> +   So in order to unwind a function using a frame pointer, the very first
> +   function that is unwound must save the frame pointer.  That way the frame
> +   pointer is restored and its value is now valid - otherwise _Unwind_GetGR
> +   crashes.  Libgcc can now be safely built with -fomit-frame-pointer.  */
> +#define LIBGCC2_UNWIND_ATTRIBUTE \
> +  __attribute__((optimize ("no-omit-frame-pointer")))
> +
>  #endif /* GCC_AARCH64_H */
>     

Re: [PATCH][AArch64] PR71951: Fix unwinding with -fomit-frame-pointer

[Wilco Dijkstra]

James Greenhalgh wrote:

> This seems like a bit of a theoretical issue as we would normally build
> libgcc with -fno-omit-frame-pointer anyway, but it can't hurt to guarantee
> this, so OK.

It's not theoretical since there were multiple users reporting unwinding issues,
so clearly doing CFLAGS="-fomit-frame-pointer" is popular. While this patch 
fixes AArch64, this bug affects all targets. Obviously -fno-omit-frame-pointer
should be an automatic default for the libgcc unwind code until the Dwarf
engine is fixed and it is made clear what the target unwind interface is.

Wilco