Re: [Lldb-commits] [lldb] r257117 - Performance improvement: Change lldb so that it puts a breakpoint

[Jason Molenda via lldb-commits]

Thanks Siva, I'll revert the change.  I have access to a linux machine but I 
haven't used it in a long time.  Would it be easy to send me the binary (with 
debug info) for this test case so I can see what gcc emitted here and figure 
out how it went wrong?  This test passes with clang on x86_64 darwin but there 
must be something in gcc's debug info that is confusing it.

Jason


> On Jan 7, 2016, at 6:16 PM, Siva Chandra  wrote:
> 
> This broke TestStepNoDebug, atleast on Linux when inferior is compiled
> with GCC: 
> http://lab.llvm.org:8011/builders/lldb-x86_64-ubuntu-14.04-cmake/builds/10091
> 
> I am able to reproduce this on my machine. Let me know if you need any
> help with anything.
> 
> On Thu, Jan 7, 2016 at 4:06 PM, Jason Molenda via lldb-commits
>  wrote:
>> Author: jmolenda
>> Date: Thu Jan  7 18:06:03 2016
>> New Revision: 257117
>> 
>> URL: http://llvm.org/viewvc/llvm-project?rev=257117=rev
>> Log:
>> Performance improvement: Change lldb so that it puts a breakpoint
>> on the first branch instruction after a function return (or the end
>> of a source line), instead of a breakpoint on the return address,
>> to skip an extra stop & start of the inferior process.
>> 
>> I changed Process::AdvanceAddressToNextBranchInstruction to not
>> take an optional InstructionList argument - no callers are providing
>> a cached InstructionList today, and if this function was going to
>> do that, the right thing to do would be to fill out / use a
>> DisassemblerSP which is a disassembler with the InstructionList for
>> this address range.
>> 
>> 
>> http://reviews.llvm.org/D15708
>> 
>> 
>> 
>> Modified:
>>lldb/trunk/include/lldb/Target/Process.h
>>lldb/trunk/include/lldb/Target/Thread.h
>>lldb/trunk/include/lldb/Target/ThreadPlanStepOut.h
>>lldb/trunk/source/Target/Process.cpp
>>lldb/trunk/source/Target/Thread.cpp
>>lldb/trunk/source/Target/ThreadPlanShouldStopHere.cpp
>>lldb/trunk/source/Target/ThreadPlanStepOut.cpp
>>lldb/trunk/source/Target/ThreadPlanStepOverRange.cpp
>> 
>> Modified: lldb/trunk/include/lldb/Target/Process.h
>> URL: 
>> http://llvm.org/viewvc/llvm-project/lldb/trunk/include/lldb/Target/Process.h?rev=257117=257116=257117=diff
>> ==
>> --- lldb/trunk/include/lldb/Target/Process.h (original)
>> +++ lldb/trunk/include/lldb/Target/Process.h Thu Jan  7 18:06:03 2016
>> @@ -3149,6 +3149,34 @@ public:
>> void
>> ResetImageToken(size_t token);
>> 
>> +//--
>> +/// Find the next branch instruction to set a breakpoint on
>> +///
>> +/// When instruction stepping through a source line, instead of
>> +/// stepping through each instruction, we can put a breakpoint on
>> +/// the next branch instruction (within the range of instructions
>> +/// we are stepping through) and continue the process to there,
>> +/// yielding significant performance benefits over instruction
>> +/// stepping.
>> +///
>> +/// @param[in] default_stop_addr
>> +/// The address of the instruction where lldb would put a
>> +/// breakpoint normally.
>> +///
>> +/// @param[in] range_bounds
>> +/// The range which the breakpoint must be contained within.
>> +/// Typically a source line.
>> +///
>> +/// @return
>> +/// The address of the next branch instruction, or the end of
>> +/// the range provided in range_bounds.  If there are any
>> +/// problems with the disassembly or getting the instructions,
>> +/// the original default_stop_addr will be returned.
>> +//--
>> +Address
>> +AdvanceAddressToNextBranchInstruction (Address default_stop_addr,
>> +   AddressRange range_bounds);
>> +
>> protected:
>> void
>> SetState (lldb::EventSP _sp);
>> 
>> Modified: lldb/trunk/include/lldb/Target/Thread.h
>> URL: 
>> http://llvm.org/viewvc/llvm-project/lldb/trunk/include/lldb/Target/Thread.h?rev=257117=257116=257117=diff
>> ==
>> --- lldb/trunk/include/lldb/Target/Thread.h (original)
>> +++ lldb/trunk/include/lldb/Target/Thread.h Thu Jan  7 18:06:03 2016
>> @@ -888,6 +888,16 @@ public:
>> /// @param[in] run_vote
>> ///See standard meanings for the stop & run votes in ThreadPlan.h.
>> ///
>> +/// @param[in] continue_to_next_branch
>> +///Normally this will enqueue a plan that will put a breakpoint on 
>> the return address and continue
>> +///to there.  If continue_to_next_branch is true, this is an 
>> operation not involving the user --
>> +///e.g. stepping "next" in a source line and we instruction stepped 
>> into another function --
>> +///

Re: [Lldb-commits] [lldb] r257117 - Performance improvement: Change lldb so that it puts a breakpoint

[Jason Molenda via lldb-commits]

Thanks for the help with this.  

The problem is that lldb would try to advance the return breakpoint in 
no_debug_caller:

(lldb) disass -b
a.out`no_debug_caller:
0x4005dd <+0>:  55pushq  %rbp
0x4005de <+1>:  48 89 e5  movq   %rsp, %rbp
0x4005e1 <+4>:  48 83 ec 20   subq   $0x20, %rsp
0x4005e5 <+8>:  89 7d ec  movl   %edi, -0x14(%rbp)
0x4005e8 <+11>: 48 89 75 e0   movq   %rsi, -0x20(%rbp)
0x4005ec <+15>: c7 45 fc 00 00 00 00  movl   $0x0, -0x4(%rbp)
0x4005f3 <+22>: 48 8b 55 e0   movq   -0x20(%rbp), %rdx
0x4005f7 <+26>: 8b 45 ec  movl   -0x14(%rbp), %eax
0x4005fa <+29>: 48 89 d6  movq   %rdx, %rsi
0x4005fd <+32>: 89 c7 movl   %eax, %edi
0x4005ff <+34>: e8 b0 ff ff ffcallq  0x4005b4  ; 
no_debug_caller_intermediate
0x400604 <+39>: 89 45 fc  movl   %eax, -0x4(%rbp)
0x400607 <+42>: 8b 45 fc  movl   -0x4(%rbp), %eax
0x40060a <+45>: c9leave  
0x40060b <+46>: c3retq   

which has no source line information so ThreadPlanStepOut::ThreadPlanStepOut() 
would advance that to the last instruction of the function (retq) and stop 
there.  This is arguably not necessary or even helpful, so I could simply 
change ThreadPlanStepOut::ThreadPlanStepOut to not advance the pc unless we 
have source line information (the only time we're really doing anything to help 
the user / performance).

When we stop on the retq instruction, we expose the fact that lldb doesn't know 
what the 'leave' instruction does so the unwind information is wrong.  Instead 
of finding the caller is main(), we find the stack frame above that, 
__libc_start_main() which has some random debug info for line 321 (which is how 
that line # came up in the testsuite failure).  You can see this with the 
unwind info difference between the assembly profiling and the eh_frame 
information:

Assembly language inspection UnwindPlan:
This UnwindPlan originally sourced from assembly insn profiling
This UnwindPlan is sourced from the compiler: no.
This UnwindPlan is valid at all instruction locations: yes.
Address range of this UnwindPlan: [a.out..text + 413-0x01cc)
row[0]:0: CFA=rsp +8 => rsp=CFA+0 rip=[CFA-8] 
row[1]:1: CFA=rsp+16 => rbp=[CFA-16] rsp=CFA+0 rip=[CFA-8] 
row[2]:4: CFA=rbp+16 => rbp=[CFA-16] rsp=CFA+0 rip=[CFA-8] 
row[3]:   46: CFA=rbp+16 => rbp=[CFA-16] rsp=CFA+0 rip=[CFA-8] 

eh_frame UnwindPlan:
This UnwindPlan originally sourced from eh_frame CFI
This UnwindPlan is sourced from the compiler: yes.
This UnwindPlan is valid at all instruction locations: no.
Address range of this UnwindPlan: [a.out..text + 413-0x01cc)
row[0]:0: CFA=rsp +8 => rip=[CFA-8] 
row[1]:1: CFA=rsp+16 => rbp=[CFA-16] rip=[CFA-8] 
row[2]:4: CFA=rbp+16 => rbp=[CFA-16] rip=[CFA-8] 
row[3]:   46: CFA=rsp +8 => rbp=[CFA-16] rip=[CFA-8] 

clang doesn't emit the LEAVE instruction which is why I didn't see this on the 
mac when I was testing there.



I have a patch to handle the LEAVE instruction but I need to head home in a bit 
- I'll check that in tomorrow after I've had a chance to look at it more 
closely.  And I'll talk to Jim about advancing the pc when we don't have line 
level information; I'm inclined to not do that. 

Thanks again - having the same binary that was failing was a big help.



> On Jan 7, 2016, at 6:16 PM, Siva Chandra  wrote:
> 
> This broke TestStepNoDebug, atleast on Linux when inferior is compiled
> with GCC: 
> http://lab.llvm.org:8011/builders/lldb-x86_64-ubuntu-14.04-cmake/builds/10091
> 
> I am able to reproduce this on my machine. Let me know if you need any
> help with anything.
> 
> On Thu, Jan 7, 2016 at 4:06 PM, Jason Molenda via lldb-commits
>  wrote:
>> Author: jmolenda
>> Date: Thu Jan  7 18:06:03 2016
>> New Revision: 257117
>> 
>> URL: http://llvm.org/viewvc/llvm-project?rev=257117=rev
>> Log:
>> Performance improvement: Change lldb so that it puts a breakpoint
>> on the first branch instruction after a function return (or the end
>> of a source line), instead of a breakpoint on the return address,
>> to skip an extra stop & start of the inferior process.
>> 
>> I changed Process::AdvanceAddressToNextBranchInstruction to not
>> take an optional InstructionList argument - no callers are providing
>> a cached InstructionList today, and if this function was going to
>> do that, the right thing to do would be to fill out / use a
>> DisassemblerSP which is a disassembler with the InstructionList for
>> this address range.
>> 
>> 
>> http://reviews.llvm.org/D15708
>> 
>> 
>> 
>> Modified:
>>lldb/trunk/include/lldb/Target/Process.h
>>lldb/trunk/include/lldb/Target/Thread.h
>>lldb/trunk/include/lldb/Target/ThreadPlanStepOut.h
>>lldb/trunk/source/Target/Process.cpp
>>

Re: [Lldb-commits] [lldb] r257117 - Performance improvement: Change lldb so that it puts a breakpoint

[Siva Chandra via lldb-commits]

This broke TestStepNoDebug, atleast on Linux when inferior is compiled
with GCC: 
http://lab.llvm.org:8011/builders/lldb-x86_64-ubuntu-14.04-cmake/builds/10091

I am able to reproduce this on my machine. Let me know if you need any
help with anything.

On Thu, Jan 7, 2016 at 4:06 PM, Jason Molenda via lldb-commits
 wrote:
> Author: jmolenda
> Date: Thu Jan  7 18:06:03 2016
> New Revision: 257117
>
> URL: http://llvm.org/viewvc/llvm-project?rev=257117=rev
> Log:
> Performance improvement: Change lldb so that it puts a breakpoint
> on the first branch instruction after a function return (or the end
> of a source line), instead of a breakpoint on the return address,
> to skip an extra stop & start of the inferior process.
>
> I changed Process::AdvanceAddressToNextBranchInstruction to not
> take an optional InstructionList argument - no callers are providing
> a cached InstructionList today, and if this function was going to
> do that, the right thing to do would be to fill out / use a
> DisassemblerSP which is a disassembler with the InstructionList for
> this address range.
>
>
> http://reviews.llvm.org/D15708
> 
>
>
> Modified:
> lldb/trunk/include/lldb/Target/Process.h
> lldb/trunk/include/lldb/Target/Thread.h
> lldb/trunk/include/lldb/Target/ThreadPlanStepOut.h
> lldb/trunk/source/Target/Process.cpp
> lldb/trunk/source/Target/Thread.cpp
> lldb/trunk/source/Target/ThreadPlanShouldStopHere.cpp
> lldb/trunk/source/Target/ThreadPlanStepOut.cpp
> lldb/trunk/source/Target/ThreadPlanStepOverRange.cpp
>
> Modified: lldb/trunk/include/lldb/Target/Process.h
> URL: 
> http://llvm.org/viewvc/llvm-project/lldb/trunk/include/lldb/Target/Process.h?rev=257117=257116=257117=diff
> ==
> --- lldb/trunk/include/lldb/Target/Process.h (original)
> +++ lldb/trunk/include/lldb/Target/Process.h Thu Jan  7 18:06:03 2016
> @@ -3149,6 +3149,34 @@ public:
>  void
>  ResetImageToken(size_t token);
>
> +//--
> +/// Find the next branch instruction to set a breakpoint on
> +///
> +/// When instruction stepping through a source line, instead of
> +/// stepping through each instruction, we can put a breakpoint on
> +/// the next branch instruction (within the range of instructions
> +/// we are stepping through) and continue the process to there,
> +/// yielding significant performance benefits over instruction
> +/// stepping.
> +///
> +/// @param[in] default_stop_addr
> +/// The address of the instruction where lldb would put a
> +/// breakpoint normally.
> +///
> +/// @param[in] range_bounds
> +/// The range which the breakpoint must be contained within.
> +/// Typically a source line.
> +///
> +/// @return
> +/// The address of the next branch instruction, or the end of
> +/// the range provided in range_bounds.  If there are any
> +/// problems with the disassembly or getting the instructions,
> +/// the original default_stop_addr will be returned.
> +//--
> +Address
> +AdvanceAddressToNextBranchInstruction (Address default_stop_addr,
> +   AddressRange range_bounds);
> +
>  protected:
>  void
>  SetState (lldb::EventSP _sp);
>
> Modified: lldb/trunk/include/lldb/Target/Thread.h
> URL: 
> http://llvm.org/viewvc/llvm-project/lldb/trunk/include/lldb/Target/Thread.h?rev=257117=257116=257117=diff
> ==
> --- lldb/trunk/include/lldb/Target/Thread.h (original)
> +++ lldb/trunk/include/lldb/Target/Thread.h Thu Jan  7 18:06:03 2016
> @@ -888,6 +888,16 @@ public:
>  /// @param[in] run_vote
>  ///See standard meanings for the stop & run votes in ThreadPlan.h.
>  ///
> +/// @param[in] continue_to_next_branch
> +///Normally this will enqueue a plan that will put a breakpoint on 
> the return address and continue
> +///to there.  If continue_to_next_branch is true, this is an 
> operation not involving the user --
> +///e.g. stepping "next" in a source line and we instruction stepped 
> into another function --
> +///so instead of putting a breakpoint on the return address, advance 
> the breakpoint to the
> +///end of the source line that is doing the call, or until the next 
> flow control instruction.
> +///If the return value from the function call is to be retrieved / 
> displayed to the user, you must stop
> +///on the return address.  The return value may be stored in 
> volatile registers which are overwritten
> +///before the next branch instruction.
> +///
>  /// @return
>  /// A shared pointer to the newly queued