[PATCH] Fix remaining inline/tailcall unwinding breakage for x86_64

[PATCH] Fix remaining inline/tailcall unwinding breakage for x86_64

[Luis Machado]

Commit 5939967b355ba6a940887d19847b7893a4506067 fixed inline
frame unwinding breakage for some targets (aarch64, riscv, s390...)
but regressed a few amd64 testcases related to tailcalls.

Given the following example situation...

Frame #-1 - sentinel frame
Frame # 0 - inline frame
Frame # 1 - normal frame

... suppose we're at level #1 and call into dwarf2_tailcall_sniffer_first.

We'll attempt to fetch PC, which used to be done via the gdbarch_unwind_pc call
(before 5939967b355ba6a940887d19847b7893a4506067), but now it is being handled
by the get_frame_register function.

gdbarch_unwind_pc will attempt to use frame #1's cache to retrieve information
about the PC. Here's where different architectures behave differently.

x86_64 will find a dwarf rule to retrieve PC from memory, at a CFA + offset
location. So the PC value is readily available and there is no need to
create a lazy value.

For aarch64 (and others), GCC doesn't emit an explicit location for PC, so we
eventually will find that PC is DWARF2_FRAME_REG_UNSPECIFIED. This is known
and is handled by GDB by assuming GCC really meant DWARF2_FRAME_REG_SAME_VALUE.

This means we'll attempt to fetch the register value from frame #0, via a call
to frame_unwind_got_register, which will trigger the creation of a lazy value
that requires a valid frame id for frame #0.

We don't have a valid id for frame #0 yet, so we assert.

Given the above, the following patch attempts to handle the situation without
being too hacky. We verify if the next frame is an inline frame and if its
frame id has been computed already. If it hasn't been computed yet, then we
use the safer get_frame_register function, otherwise we use the regular
gdbarch_unwind_pc hook.

I've verified this makes both aarch64-linux and x86_64 happy testsuite-wise.

gdb/ChangeLog:

2020-04-25  Luis Machado  <luis.machado@linaro.org>

	* dwarf2/frame-tailcall.c (dwarf2_tailcall_sniffer_first): Handle
	problematic inline frame unwinding situation.
	* frame.c (frame_id_computed_p): New function.
	* frame.h (frame_id_computed_p): New prototype.
---
 gdb/dwarf2/frame-tailcall.c | 39 ++++++++++++++++++++++++++++++++++---
 gdb/frame.c                 |  8 ++++++++
 gdb/frame.h                 |  4 ++++
 3 files changed, 48 insertions(+), 3 deletions(-)

diff --git a/gdb/dwarf2/frame-tailcall.c b/gdb/dwarf2/frame-tailcall.c
index 01bb134a5c..16dba2b201 100644
--- a/gdb/dwarf2/frame-tailcall.c
+++ b/gdb/dwarf2/frame-tailcall.c
@@ -384,10 +384,43 @@ dwarf2_tailcall_sniffer_first (struct frame_info *this_frame,
 
       prev_gdbarch = frame_unwind_arch (this_frame);
 
+      /* The dwarf2 tailcall sniffer runs early, at the end of populating the
+	 dwarf2 frame cache for the current frame.  If there exists inline
+	 frames inner (next) to the current frame, there is a good possibility
+	 of that inline frame not having a computed frame id yet.
+
+	 This is because computing such a frame id requires us to walk through
+	 the frame chain until we find the first normal frame after the inline
+	 frame and then compute the normal frame's id first.
+
+	 Some architectures' compilers generate enough register location
+	 information for a dwarf unwinder to fetch PC without relying on inner
+	 frames (x86_64 for example).  In this case the PC is retrieved
+	 according to dwarf rules.
+
+	 But others generate less strict dwarf data for which assumptions are
+	 made (like interpreting DWARF2_FRAME_REG_UNSPECIFIED as
+	 DWARF2_FRAME_REG_SAME_VALUE).  For such cases, GDB may attempt to
+	 create lazy values for registers, and those lazy values must be
+	 created with a valid frame id, but we potentially have no valid id.
+
+	 So, to avoid breakage, if we see a dangerous situation with inline
+	 frames without a computed id, use safer functions to retrieve the
+	 current frame's PC.  Otherwise use the provided dwarf rules.  */
+      frame_info *next_frame = get_next_frame (this_frame);
+
       /* Simulate frame_unwind_pc without setting this_frame->prev_pc.p.  */
-      get_frame_register (this_frame, gdbarch_pc_regnum (prev_gdbarch),
-			  (gdb_byte *) &prev_pc);
-      prev_pc = gdbarch_addr_bits_remove (prev_gdbarch, prev_pc);
+      if (next_frame != nullptr && get_frame_type (next_frame) == INLINE_FRAME
+	  && !frame_id_computed_p (next_frame))
+	{
+	  /* The next frame is an inline frame and its frame id has not been
+	     computed yet.  */
+	  get_frame_register (this_frame, gdbarch_pc_regnum (prev_gdbarch),
+			      (gdb_byte *) &prev_pc);
+	  prev_pc = gdbarch_addr_bits_remove (prev_gdbarch, prev_pc);
+	}
+      else
+	prev_pc = gdbarch_unwind_pc (prev_gdbarch, this_frame);
 
       /* call_site_find_chain can throw an exception.  */
       chain = call_site_find_chain (prev_gdbarch, prev_pc, this_pc);
diff --git a/gdb/frame.c b/gdb/frame.c
index ac1016b083..ff27b9f00e 100644
--- a/gdb/frame.c
+++ b/gdb/frame.c
@@ -687,6 +687,14 @@ frame_id_build_wild (CORE_ADDR stack_addr)
   return id;
 }
 
+bool
+frame_id_computed_p (struct frame_info *frame)
+{
+  gdb_assert (frame != nullptr);
+
+  return frame->this_id.p != 0;
+}
+
 int
 frame_id_p (struct frame_id l)
 {
diff --git a/gdb/frame.h b/gdb/frame.h
index cfc15022ed..e835d49f9c 100644
--- a/gdb/frame.h
+++ b/gdb/frame.h
@@ -236,6 +236,10 @@ extern struct frame_id
    as the special identifier address are set to indicate wild cards.  */
 extern struct frame_id frame_id_build_wild (CORE_ADDR stack_addr);
 
+/* Returns true if FRAME's id has been computed.
+   Returns false otherwise.  */
+extern bool frame_id_computed_p (struct frame_info *frame);
+
 /* Returns non-zero when L is a valid frame (a valid frame has a
    non-zero .base).  The outermost frame is valid even without an
    ID.  */
-- 
2.17.1

Re: [PATCH] Fix remaining inline/tailcall unwinding breakage for x86_64

[Tom de Vries]

On 25-04-2020 06:09, Luis Machado wrote:
> Commit 5939967b355ba6a940887d19847b7893a4506067 fixed inline
> frame unwinding breakage for some targets (aarch64, riscv, s390...)
> but regressed a few amd64 testcases related to tailcalls.
> 
> Given the following example situation...
> 
> Frame #-1 - sentinel frame
> Frame # 0 - inline frame
> Frame # 1 - normal frame
> 
> ... suppose we're at level #1 and call into dwarf2_tailcall_sniffer_first.
> 
> We'll attempt to fetch PC, which used to be done via the gdbarch_unwind_pc call
> (before 5939967b355ba6a940887d19847b7893a4506067), but now it is being handled
> by the get_frame_register function.
> 
> gdbarch_unwind_pc will attempt to use frame #1's cache to retrieve information
> about the PC. Here's where different architectures behave differently.
> 
> x86_64 will find a dwarf rule to retrieve PC from memory, at a CFA + offset
> location. So the PC value is readily available and there is no need to
> create a lazy value.
> 
> For aarch64 (and others), GCC doesn't emit an explicit location for PC, so we
> eventually will find that PC is DWARF2_FRAME_REG_UNSPECIFIED. This is known
> and is handled by GDB by assuming GCC really meant DWARF2_FRAME_REG_SAME_VALUE.
> 
> This means we'll attempt to fetch the register value from frame #0, via a call
> to frame_unwind_got_register, which will trigger the creation of a lazy value
> that requires a valid frame id for frame #0.
> 
> We don't have a valid id for frame #0 yet, so we assert.
> 
> Given the above, the following patch attempts to handle the situation without
> being too hacky. We verify if the next frame is an inline frame and if its
> frame id has been computed already. If it hasn't been computed yet, then we
> use the safer get_frame_register function, otherwise we use the regular
> gdbarch_unwind_pc hook.
> 
> I've verified this makes both aarch64-linux and x86_64 happy testsuite-wise.
> 

Hi Luis,

thanks for working on this.

I've tested this patch on x86_64-linux and can confirm that this fixes
all the regressions I saw.

I've reviewed the patch and it looks ok to me.

Please check this in, given that if fixes regressions.  If there are any
comments from others to be addressed, that can still be done post-commit.

Thanks,
- Tom

> gdb/ChangeLog:
> 
> 2020-04-25  Luis Machado  <luis.machado@linaro.org>
> 
> 	* dwarf2/frame-tailcall.c (dwarf2_tailcall_sniffer_first): Handle
> 	problematic inline frame unwinding situation.
> 	* frame.c (frame_id_computed_p): New function.
> 	* frame.h (frame_id_computed_p): New prototype.
> ---
>  gdb/dwarf2/frame-tailcall.c | 39 ++++++++++++++++++++++++++++++++++---
>  gdb/frame.c                 |  8 ++++++++
>  gdb/frame.h                 |  4 ++++
>  3 files changed, 48 insertions(+), 3 deletions(-)
> 
> diff --git a/gdb/dwarf2/frame-tailcall.c b/gdb/dwarf2/frame-tailcall.c
> index 01bb134a5c..16dba2b201 100644
> --- a/gdb/dwarf2/frame-tailcall.c
> +++ b/gdb/dwarf2/frame-tailcall.c
> @@ -384,10 +384,43 @@ dwarf2_tailcall_sniffer_first (struct frame_info *this_frame,
>  
>        prev_gdbarch = frame_unwind_arch (this_frame);
>  
> +      /* The dwarf2 tailcall sniffer runs early, at the end of populating the
> +	 dwarf2 frame cache for the current frame.  If there exists inline
> +	 frames inner (next) to the current frame, there is a good possibility
> +	 of that inline frame not having a computed frame id yet.
> +
> +	 This is because computing such a frame id requires us to walk through
> +	 the frame chain until we find the first normal frame after the inline
> +	 frame and then compute the normal frame's id first.
> +
> +	 Some architectures' compilers generate enough register location
> +	 information for a dwarf unwinder to fetch PC without relying on inner
> +	 frames (x86_64 for example).  In this case the PC is retrieved
> +	 according to dwarf rules.
> +
> +	 But others generate less strict dwarf data for which assumptions are
> +	 made (like interpreting DWARF2_FRAME_REG_UNSPECIFIED as
> +	 DWARF2_FRAME_REG_SAME_VALUE).  For such cases, GDB may attempt to
> +	 create lazy values for registers, and those lazy values must be
> +	 created with a valid frame id, but we potentially have no valid id.
> +
> +	 So, to avoid breakage, if we see a dangerous situation with inline
> +	 frames without a computed id, use safer functions to retrieve the
> +	 current frame's PC.  Otherwise use the provided dwarf rules.  */
> +      frame_info *next_frame = get_next_frame (this_frame);
> +
>        /* Simulate frame_unwind_pc without setting this_frame->prev_pc.p.  */
> -      get_frame_register (this_frame, gdbarch_pc_regnum (prev_gdbarch),
> -			  (gdb_byte *) &prev_pc);
> -      prev_pc = gdbarch_addr_bits_remove (prev_gdbarch, prev_pc);
> +      if (next_frame != nullptr && get_frame_type (next_frame) == INLINE_FRAME
> +	  && !frame_id_computed_p (next_frame))
> +	{
> +	  /* The next frame is an inline frame and its frame id has not been
> +	     computed yet.  */
> +	  get_frame_register (this_frame, gdbarch_pc_regnum (prev_gdbarch),
> +			      (gdb_byte *) &prev_pc);
> +	  prev_pc = gdbarch_addr_bits_remove (prev_gdbarch, prev_pc);
> +	}
> +      else
> +	prev_pc = gdbarch_unwind_pc (prev_gdbarch, this_frame);
>  
>        /* call_site_find_chain can throw an exception.  */
>        chain = call_site_find_chain (prev_gdbarch, prev_pc, this_pc);
> diff --git a/gdb/frame.c b/gdb/frame.c
> index ac1016b083..ff27b9f00e 100644
> --- a/gdb/frame.c
> +++ b/gdb/frame.c
> @@ -687,6 +687,14 @@ frame_id_build_wild (CORE_ADDR stack_addr)
>    return id;
>  }
>  
> +bool
> +frame_id_computed_p (struct frame_info *frame)
> +{
> +  gdb_assert (frame != nullptr);
> +
> +  return frame->this_id.p != 0;
> +}
> +
>  int
>  frame_id_p (struct frame_id l)
>  {
> diff --git a/gdb/frame.h b/gdb/frame.h
> index cfc15022ed..e835d49f9c 100644
> --- a/gdb/frame.h
> +++ b/gdb/frame.h
> @@ -236,6 +236,10 @@ extern struct frame_id
>     as the special identifier address are set to indicate wild cards.  */
>  extern struct frame_id frame_id_build_wild (CORE_ADDR stack_addr);
>  
> +/* Returns true if FRAME's id has been computed.
> +   Returns false otherwise.  */
> +extern bool frame_id_computed_p (struct frame_info *frame);
> +
>  /* Returns non-zero when L is a valid frame (a valid frame has a
>     non-zero .base).  The outermost frame is valid even without an
>     ID.  */
> 

Re: [PATCH] Fix remaining inline/tailcall unwinding breakage for x86_64

[Luis Machado]

On 4/25/20 12:53 PM, Tom de Vries wrote:
> On 25-04-2020 06:09, Luis Machado wrote:
>> Commit 5939967b355ba6a940887d19847b7893a4506067 fixed inline
>> frame unwinding breakage for some targets (aarch64, riscv, s390...)
>> but regressed a few amd64 testcases related to tailcalls.
>>
>> Given the following example situation...
>>
>> Frame #-1 - sentinel frame
>> Frame # 0 - inline frame
>> Frame # 1 - normal frame
>>
>> ... suppose we're at level #1 and call into dwarf2_tailcall_sniffer_first.
>>
>> We'll attempt to fetch PC, which used to be done via the gdbarch_unwind_pc call
>> (before 5939967b355ba6a940887d19847b7893a4506067), but now it is being handled
>> by the get_frame_register function.
>>
>> gdbarch_unwind_pc will attempt to use frame #1's cache to retrieve information
>> about the PC. Here's where different architectures behave differently.
>>
>> x86_64 will find a dwarf rule to retrieve PC from memory, at a CFA + offset
>> location. So the PC value is readily available and there is no need to
>> create a lazy value.
>>
>> For aarch64 (and others), GCC doesn't emit an explicit location for PC, so we
>> eventually will find that PC is DWARF2_FRAME_REG_UNSPECIFIED. This is known
>> and is handled by GDB by assuming GCC really meant DWARF2_FRAME_REG_SAME_VALUE.
>>
>> This means we'll attempt to fetch the register value from frame #0, via a call
>> to frame_unwind_got_register, which will trigger the creation of a lazy value
>> that requires a valid frame id for frame #0.
>>
>> We don't have a valid id for frame #0 yet, so we assert.
>>
>> Given the above, the following patch attempts to handle the situation without
>> being too hacky. We verify if the next frame is an inline frame and if its
>> frame id has been computed already. If it hasn't been computed yet, then we
>> use the safer get_frame_register function, otherwise we use the regular
>> gdbarch_unwind_pc hook.
>>
>> I've verified this makes both aarch64-linux and x86_64 happy testsuite-wise.
>>
> 
> Hi Luis,
> 
> thanks for working on this.
> 
> I've tested this patch on x86_64-linux and can confirm that this fixes
> all the regressions I saw.
> 
> I've reviewed the patch and it looks ok to me.
> 
> Please check this in, given that if fixes regressions.  If there are any
> comments from others to be addressed, that can still be done post-commit.
> 
> Thanks,
> - Tom
> 

Thanks for the quick review. Pushed now.