On 16 July 2016 at 14:29, Laszlo Ersek <[email protected]> wrote: > On 07/16/16 00:16, Ard Biesheuvel wrote: >> Now that we switched to the __builtin_ms_va_list VA_LIST type for >> GCC/X64, we can trust the compiler to do the right thing even under >> optimization, and so we can enable -Os optimization all the way back >> to GCC44, and drop the -D define that prevents the use of the __builtin >> VA_LIST types. Note that this requires the -maccumulate-outgoing-args >> switch as well. >> >> Contributed-under: TianoCore Contribution Agreement 1.0 >> Signed-off-by: Ard Biesheuvel <[email protected]> >> --- >> BaseTools/Conf/tools_def.template | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/BaseTools/Conf/tools_def.template >> b/BaseTools/Conf/tools_def.template >> index 2065fa34998f..a7da6741611d 100644 >> --- a/BaseTools/Conf/tools_def.template >> +++ b/BaseTools/Conf/tools_def.template >> @@ -4353,7 +4353,7 @@ DEFINE GCC_AARCH64_RC_FLAGS = -I binary -O >> elf64-littleaarch64 -B aarch64 >> >> DEFINE GCC44_ALL_CC_FLAGS = -g -fshort-wchar >> -fno-strict-aliasing -Wall -Werror -Wno-array-bounds -ffunction-sections >> -fdata-sections -c -include AutoGen.h -fno-common >> -DSTRING_ARRAY_NAME=$(BASE_NAME)Strings >> DEFINE GCC44_IA32_CC_FLAGS = DEF(GCC44_ALL_CC_FLAGS) -m32 >> -march=i586 -malign-double -fno-stack-protector -D EFI32 >> -fno-asynchronous-unwind-tables >> -DEFINE GCC44_X64_CC_FLAGS = DEF(GCC44_ALL_CC_FLAGS) -m64 >> -fno-stack-protector "-DEFIAPI=__attribute__((ms_abi))" >> -DNO_BUILTIN_VA_FUNCS -mno-red-zone -Wno-address -mcmodel=large >> -fno-asynchronous-unwind-tables >> +DEFINE GCC44_X64_CC_FLAGS = DEF(GCC44_ALL_CC_FLAGS) -m64 >> -fno-stack-protector "-DEFIAPI=__attribute__((ms_abi))" -Os >> -maccumulate-outgoing-args -mno-red-zone -Wno-address -mcmodel=large >> -fno-asynchronous-unwind-tables >> DEFINE GCC44_IA32_X64_DLINK_COMMON = -nostdlib -n -q --gc-sections -z >> common-page-size=0x20 >> DEFINE GCC44_IA32_X64_ASLDLINK_FLAGS = DEF(GCC44_IA32_X64_DLINK_COMMON) >> --entry ReferenceAcpiTable -u ReferenceAcpiTable >> DEFINE GCC44_IA32_X64_DLINK_FLAGS = DEF(GCC44_IA32_X64_DLINK_COMMON) >> --entry $(IMAGE_ENTRY_POINT) -u $(IMAGE_ENTRY_POINT) -Map >> $(DEST_DIR_DEBUG)/$(BASE_NAME).map >> > > Before I embark on build-testing this series too with my "build farm", > I'd like to point out this thread: > > http://thread.gmane.org/gmane.comp.bios.tianocore.devel/10741/focus=10961 > > Now, the assumption that -Os itself was causing the corruption has been > laid to rest; we now know that the corruption was the product of the > VA_LIST implementation, which is exactly what this series is replacing. > So that's not why I'm pointing at the thread. > > However, I recall from the thread that -Os enables -fomit-frame-pointer, > which might make source level debugging impossible (according to the GCC > manual). > > Now, we're not big on source level debugging in GCC builds, at least > right now, plus I also cannot claim that that -fomit-frame-pointer is > never enabled *otherwise*. Much as I know -fomit-frame-pointer could be > enabled with -O1, -O2, even with -O0?... >
I simply extrapolated from IA32, which uses -Os for DEBUG as well. Since IA32 has even fewer general purpose registers, I would assume that by the same reasoning, this is OK for X64 as well. I wouldn't have given it any thought if you hadn't mentioned it :-) > I'd just like to avoid a setting that *guarantees* that source level > debugging would be impossible or garbled. Ard, can you comment on that? > That seems like a genuine concern, and -Os optimization but with a frame pointer should be perfectly feasible, so perhaps it is better to simply add -fno-omit-frame-pointer in the DEBUG case, especially since it does not hurt X64 as much as it hurts IA32 code. > ... Actually, just now I'm remembering something Scott explained to me: > the difference between DEBUG, RELEASE, and NOOPT. Both DEBUG and RELEASE > are supposed to be optimized (they differ in the compilation of DEBUG, > DEBUG_CODE, ASSERT etc; not in optimization). NOOPT on the other hand is > supposed to keep DEBUGs, but also disable optimization (for source level > debugging). > > At the moment, we have no NOOPT settings for GCC. We only have RELEASE > (in the "supposed" meaning of RELEASE), and DEBUG (which has, > traditionally, stood for the NOOPT behavior actually). > > Version 1 of this patch set uses -O2 instead of -Os, but another > difference is that v1 only added optimization to RELEASE. This version > adds optimization (-Os) to DEBUG too (*) -- I guess in no small part > because I expressed a wish for that? --, but it doesn't introduce a > NOOPT target. I'm concerned that this might cause us to lose any usable > source level debugging, even though our current "source level debugging" > facility means a super contrived, out-of-tree gdb setup. > > (*) This is not my "discovery" of course, it's announced in the v2 blurb. > > I don't really know what to ask for / wish for :) I think introducing > NOOPT might be a sizeable task, and it would even require changes to > platforms (OvmfPkg and ArmVirtPkg minimally). So I don't feel good about > asking Ard to add NOOPT as well. > On ARM/AARCH64, we (I) heavily rely on the DEBUG builds to produce code that is suitable for single step debugging in GDB. I understand that this change will make X64 deviate from ARM/AARCH64 in that respect, but it does align it with IA32, so this is really a judgement call. I am not crazy about enlarging the validation space by adding a third flavor. > Instead, I admit that my suggestion (implied request?) in the > v1 thread -- i.e., to add optimization to DEBUG -- broke the GCC > toolchain tradition of DEBUG standing for NOOPT actually. > > I'm very sorry about that. :( > As you point out in your next reply, this is not true for IA32. > In order to uphold the GCC toolchain tradition for DEBUG, should we add > -Os (and whatever else -Os requires) to RELEASE only? > Optimized code is a pain to debug [even with frame pointers] so in the ARM/AARCH64 case, I would strongly object to enabling optimization for the DEBUG target. I suppose for X86, it makes sense to align with other toolchains in this case. 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. 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) _______________________________________________ edk2-devel mailing list [email protected] https://lists.01.org/mailman/listinfo/edk2-devel
