On 07/16/16 19:43, Kinney, Michael D wrote: > Laszlo, > > Symbolic debugging should be fully supported at all optimization levels. The > compiler/linker generates .pdb files for MSFT and .debug sections for GCC. > > The purpose of NOOPT is not to support source level debug. It is to make > debug easier. When optimizations are turned up, many of the call parameters > and > local variables can may be optimized into registers and calls can be inlined. > Also, the same register may be used for multiple parameters or locals > depending > on how they are used in the function. Not all debuggers are aware of these > register optimizations and may show incorrect values for parameters and > locals.
I may not have used the correct terminology, but the case when -b DEBUG enables the intrusive optimizations that you describe (i.e., the source code cannot be uniquely matched to the firmware state) is practically undistinguishable from the case when no debug symbols exist at all. It is unusable for analysis with sufficient detail. In that sense, to me at least, the difference between DEBUG and NOOPT is not "symbolic debugging is hard or easy"; it is "symbolic debugging is unusable (as if the symbols don't exist) vs. fully supported". Like everyone else, I've debugged normal Linux userspace processes that were built with "-g -O2", and their core dumps are useless for any efficient purposes. One is able to narrow it down to a more or less tight function context, but for understanding local variables, one has to look at registers and disassembly (and even that isn't guaranteed to produce results). Single stepping simply doesn't work; gcc can routinely reorder the assembly so that it loses any resemblance with the original code (the Linux kernel disables some of these gcc optimization features individually). So, as far as I'm concerned, symbolic debugging is entirely defeated by -O2 or -Os, regardless of -g. This is why I equate - NOOPT to "symbolic debug works" - DEBUG to "symbolic debug doesn't work, but ASSERT() and friends do" - RELEASE to "neither of those work". > When a difficult bug is being evaluated, it is sometimes easier to make sure > these register optimization are disabled and function inlining id disables > so the debugger can show correct values for parameters and locals on every > call > in the call stack. This is something that I expect by default from a binary whose build options are supposed to support symbolic debugging. > In this case, a single module under debug may disable > optimization in DSC <BuildOptions> or INF [BuildOptions], or if all modules > need optimization disabled to debug across the entire call stack, NOOPT > can be used. I accept that your description defines the official meaning for NOOPT / DEBUG / RELEASE, and I thank you for educating me on them. For practical purposes though, I'll have to stick with my (non-official) definitions -- if we add -Os to DEBUG (and I don't mind if we do), then to me personally, it won't be suitable for source level debugging. Only NOOPT will be (to be added later). Thanks! Laszlo > Best regards, > > Mike > >> -----Original Message----- >> From: Laszlo Ersek [mailto:ler...@redhat.com] >> Sent: Saturday, July 16, 2016 7:45 AM >> To: Ard Biesheuvel <ard.biesheu...@linaro.org> >> Cc: edk2-devel@lists.01.org <edk2-de...@ml01.01.org>; af...@apple.com; Gao, >> Liming >> <liming....@intel.com>; Shi, Steven <steven....@intel.com>; Zhu, Yonghong >> <yonghong....@intel.com>; Kinney, Michael D <michael.d.kin...@intel.com>; >> Justen, >> Jordan L <jordan.l.jus...@intel.com>; Bruce Cran <br...@cran.org.uk>; Paolo >> Bonzini >> <pbonz...@redhat.com>; Scott Duplichan <sc...@notabs.org> >> Subject: Re: [PATCH v2 3/5] BaseTools/tools_def: enable Os optimization for >> GCC X64 >> builds >> >> On 07/16/16 14:58, Ard Biesheuvel wrote: >> >>> Bottom line is that I don't really care :-) -Os for RELEASE is a clear >>> improvement. If nobody is doing source code level debugging using GCC >>> builds, it appears to be an improvement for DEBUG as well. In any >>> case, it would be good to have the numbers so we can make an informed >>> decision. >> >> At this point you've sort of convinced me that we should add -Os to >> DEBUG as well. It *doubly* aligns DEBUG_GCCxx_X64_CC_FLAGS with the >> status quo: first with GCC+IA32, second with non-GCC+X64. >> >> The gdb setup for GCC+X64 is so contrived at the moment *anyway* that >> removing -Os from the build flags as a further step is practically no >> additional burden. If we become serious about it, we can always >> introduce NOOPT later, further aligning GCC with other toolchains on >> IA32 and X64. >> >>> Another thing I noticed: OpensslLib uses -UNO_BUILTIN_VA_ARGS to >>> switch to the default va_list implementation, which is necessary since >>> its variadic functions lack an EFIAPI annotation. This means I should >>> probably revise the patch to allow the standard __builtins to be used, >>> e.g., add -DNO_MS_ABI_VARARGS to OpensslLib instead, and make the use >>> of __builtin_ms_va_list conditional on !defined(NO_MS_ABI_VARARGS) >> >> Aaargh. I've run into (independent) varargs problems with OpenSSL in >> edk2 before, so I'm not sure how my testing missed this! >> >> Ah wait, I may know how -- I think I wanted to use EnrollDefaultKeys.efi >> as a starting point for SB testing too, but I didn't get as far with it, >> because -O2 in your v1 triggered a latent bug in the app. >> >> ... So, with your next update, we won't just distinguish "builtin" from >> "no-builtin" for VA_LIST, we'll also distinguish "MS" from "SYSV" within >> "builtin:. :( >> >> <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=50818> just got twice as >> annoying. :( :( >> >> I guess I'll delay my testing until your v3. Is that okay with you? >> >> Thanks >> Laszlo _______________________________________________ edk2-devel mailing list edk2-devel@lists.01.org https://lists.01.org/mailman/listinfo/edk2-devel
