> On Dec 8, 2020, at 1:40 AM, Richard Biener <richard.guent...@gmail.com> wrote:
>
> On Mon, Dec 7, 2020 at 5:20 PM Qing Zhao <qing.z...@oracle.com
> <mailto:qing.z...@oracle.com>> wrote:
>>
>>
>>
>> On Dec 7, 2020, at 1:12 AM, Richard Biener <richard.guent...@gmail.com>
>> wrote:
>>
>> On Fri, Dec 4, 2020 at 5:19 PM Qing Zhao <qing.z...@oracle.com> wrote:
>>
>>
>>
>>
>> On Dec 4, 2020, at 2:50 AM, Richard Biener <richard.guent...@gmail.com>
>> wrote:
>>
>> On Thu, Dec 3, 2020 at 6:33 PM Richard Sandiford
>> <richard.sandif...@arm.com> wrote:
>>
>>
>> Richard Biener via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
>>
>> On Tue, Nov 24, 2020 at 4:47 PM Qing Zhao <qing.z...@oracle.com> wrote:
>>
>> Another issue is, in order to check whether an auto-variable has
>> initializer, I plan to add a new bit in “decl_common” as:
>> /* In a VAR_DECL, this is DECL_IS_INITIALIZED. */
>> unsigned decl_is_initialized :1;
>>
>> /* IN VAR_DECL, set when the decl is initialized at the declaration. */
>> #define DECL_IS_INITIALIZED(NODE) \
>> (DECL_COMMON_CHECK (NODE)->decl_common.decl_is_initialized)
>>
>> set this bit when setting DECL_INITIAL for the variables in FE. then keep it
>> even though DECL_INITIAL might be NULLed.
>>
>>
>> For locals it would be more reliable to set this flag during gimplification.
>>
>> Do you have any comment and suggestions?
>>
>>
>> As said above - do you want to cover registers as well as locals? I'd do
>> the actual zeroing during RTL expansion instead since otherwise you
>> have to figure youself whether a local is actually used (see
>> expand_stack_vars)
>>
>> Note that optimization will already made have use of "uninitialized" state
>> of locals so depending on what the actual goal is here "late" may be too
>> late.
>>
>>
>> Haven't thought about this much, so it might be a daft idea, but would a
>> compromise be to use a const internal function:
>>
>> X1 = .DEFERRED_INIT (X0, INIT)
>>
>> where the X0 argument is an uninitialised value and the INIT argument
>> describes the initialisation pattern? So for a decl we'd have:
>>
>> X = .DEFERRED_INIT (X, INIT)
>>
>> and for an SSA name we'd have:
>>
>> X_2 = .DEFERRED_INIT (X_1(D), INIT)
>>
>> with all other uses of X_1(D) being replaced by X_2. The idea is that:
>>
>> * Having the X0 argument would keep the uninitialised use of the
>> variable around for the later warning passes.
>>
>> * Using a const function should still allow the UB to be deleted as dead
>> if X1 isn't needed.
>>
>> * Having a function in the way should stop passes from taking advantage
>> of direct uninitialised uses for optimisation.
>>
>> This means we won't be able to optimise based on the actual init
>> value at the gimple level, but that seems like a fair trade-off.
>> AIUI this is really a security feature or anti-UB hardening feature
>> (in the sense that users are more likely to see predictable behaviour
>> “in the field” even if the program has UB).
>>
>>
>> The question is whether it's in line of peoples expectation that
>> explicitely zero-initialized code behaves differently from
>> implicitely zero-initialized code with respect to optimization
>> and secondary side-effects (late diagnostics, latent bugs, etc.).
>>
>> Introducing a new concept like .DEFERRED_INIT is much more
>> heavy-weight than an explicit zero initializer.
>>
>>
>> What exactly you mean by “heavy-weight”? More difficult to implement or much
>> more run-time overhead or both? Or something else?
>>
>> The major benefit of the approach of “.DEFERRED_INIT” is to enable us keep
>> the current -Wuninitialized analysis untouched and also pass
>> the “uninitialized” info from source code level to “pass_expand”.
>>
>>
>> Well, "untouched" is a bit oversimplified. You do need to handle
>> .DEFERRED_INIT as not
>> being an initialization which will definitely get interesting.
>>
>>
>> Yes, during uninitialized variable analysis pass, we should specially handle
>> the defs with “.DEFERRED_INIT”, to treat them as uninitializations.
>>
>> If we want to keep the current -Wuninitialized analysis untouched, this is a
>> quite reasonable approach.
>>
>> However, if it’s not required to keep the current -Wuninitialized analysis
>> untouched, adding zero-initializer directly during gimplification should
>> be much easier and simpler, and also smaller run-time overhead.
>>
>>
>> As for optimization I fear you'll get a load of redundant zero-init
>> actually emitted if you can just rely on RTL DSE/DCE to remove it.
>>
>>
>> Runtime overhead for -fauto-init=zero is one important consideration for the
>> whole feature, we should minimize the runtime overhead for zero
>> Initialization since it will be used in production build.
>> We can do some run-time performance evaluation when we have an
>> implementation ready.
>>
>>
>> Note there will be other passes "confused" by .DEFERRED_INIT. Note
>> that there's going to be other
>> considerations - namely where to emit the .DEFERRED_INIT - when
>> emitting it during gimplification
>> you can emit it at the start of the block of block-scope variables.
>> When emitting after gimplification
>> you have to emit at function start which will probably make stack slot
>> sharing inefficient because
>> the deferred init will cause overlapping lifetimes. With emitting at
>> block boundary the .DEFERRED_INIT
>> will act as code-motion barrier (and it itself likely cannot be moved)
>> so for example invariant motion
>> will no longer happen. Likewise optimizations like SRA will be
>> confused by .DEFERRED_INIT which
>> again will lead to bigger stack usage (and less optimization).
>>
>>
>> Yes, looks like that the inserted “.DEFERRED_INIT” function calls will
>> negatively impact tree optimizations.
>>
>>
>> But sure, you can try implement a few variants but definitely
>> .DEFERRED_INIT will be the most
>> work.
>>
>>
>> How about implement the following two approaches and compare the run-time
>> cost:
>>
>> A. Insert the real initialization during gimplification phase.
>> B. Insert the .DEFERRED_INIT during gimplification phase, and then expand
>> this call to real initialization during expand phase.
>>
>> The Approach A will have less run-time overhead, but will mess up the
>> current uninitialized variable analysis in GCC.
>> The Approach B will have more run-time overhead, but will keep the current
>> uninitialized variable analysis in GCC.
>>
>> And then decide which approach we will go with?
>>
>> What’s your opinion on this?
>
> Well, in the end you have to try. Note for the purpose of stack slot
> sharing you do want the
> instrumentation to happen during gimplification.
>
> Another possibility is to materialize .DEFERRED_INIT earlier than
> expand, for example shortly
> after IPA optimizations to avoid pessimizing loop transforms and allow
> SRA. At the point you
> materialize the inits you could run the late uninit warning pass
> (which would then be earlier
> than regular but would still see the .DEFERRED_INIT).
If we put the “materializing .DEFERRED_INIT” phase earlier as you suggested
above,
the late uninitialized warning pass has to be moved earlier in order to utilize
the “.DEFERRED_INIT”.
Then we might miss some opportunities for the late uninitialized warning. I
think that this is not we really
want.
>
> While users may be happy to pay some performance stack usage is
> probably more critical
So, which pass is for computing the stack usage?
> (just thinking of the kernel) so not regressing there should be as
> important as preserving
> uninit warnings (which I for practical purposes see not important at
> all - people can do
> "debug" builds without -fzero-init).
Looks like that the major issue with the “.DERERRED_INIT” approach is: the new
inserted calls to internal const function
might inhibit some important tree optimizations.
So, I am thinking again the following another approach I raised in the very
beginning:
During gimplification phase, mark the DECL for an auto variable without
initialization as “no_explicit_init”, then maintain this
“no_explicit_init” bit till after pass_late_warn_uninitialized, or till
pass_expand, add zero-iniitiazation for all DECLs that are
marked with “no_explicit_init”.
This approach will not have the issue to interrupt tree optimizations, however,
I guess that “maintaining this “no_explicit_init” bit
might be very difficult?
Do you have any comments on this approach?
thanks.
Qing
>
> Richard.
>
>>
>> Btw, I don't think theres any reason to cling onto clangs semantics
>> for a particular switch. We'll never be able to emulate 1:1 behavior
>> and our -Wuninit behavior is probably wastly different already.
>>
>>
>> From my study so far, yes, the currently behavior of -Wunit for Clang and
>> GCC is not exactly the same.
>>
>> For example, for the following small testing case:
>> void blah(int);
>>
>> int foo_2 (int n, int l, int m, int r)
>> {
>> int v;
>>
>> if ( (n > 10) && (m != 100) && (r < 20) )
>> v = r;
>>
>> if (l > 100)
>> if ( (n <= 8) && (m < 102) && (r < 19) )
>> blah(v); /* { dg-warning "uninitialized" "real warning" } */
>>
>> return 0;
>> }
>>
>> GCC is able to report maybe uninitialized warning, but Clang cannot.
>> Looks like that GCC’s uninitialized analysis relies on more analysis and
>> optimization information than CLANG.
>>
>> Really curious on how clang implement its uninitialized analysis?
>>
>>
>>
>> Actually, I studied a little bit on how clang implement its uninitialized
>> analysis last Friday.
>> And noticed that CLANG has a data flow analysis phase based on CLANG's AST.
>> http://clang-developers.42468.n3.nabble.com/A-survey-of-dataflow-analyses-in-Clang-td4069644.html
>>
>> And clang’s uninitialized analysis is based on this data flow analysis.
>>
>> Therefore, adding initialization AFTER clang’s uninitialization analysis
>> phase is straightforward.
>>
>> However, for GCC, we don’t have data flow analysis in FE. The uninitialized
>> variable analysis is put in TREE optimization phase,
>> Therefore, it’s much more difficult to implement this feature in GCC than
>> that in CLANG.
>>
>> Qing
>>
>>
>>
>> Qing
>>
>>
>>
>>
>> Richard.
>>
>> Thanks,
>> Richard
