<div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Mon, Apr 30, 2018 at 1:05 PM, <span dir="ltr"><<a href="mailto:dmitry.mikulin@sony.com" target="_blank">dmitry.mikulin@sony.com</a>></span> wrote:<br></div><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div>
<div class="m_-2534272377643883510m_6116811485634542887BodyFragment"><font size="2"><span style="font-size:10pt">
<div class="m_-2534272377643883510m_6116811485634542887PlainText">Replacing direct calls to jump table entries with calls to real targets recovers most of the performance loss in my benchmark. Dealing with link order files is a bit cumbersome though: I can’t keep both, say “foo” and “foo.cfi” next to
each other in link order, because if both exist, the linker reorders the jump table next to the real function. This is not what we want when the goal is to get rid of calls through jump tables.</div></span></font></div></div></blockquote><div><br></div><div>I would have thought that it could be beneficial to place the jump table next to one of the target functions, as it may reduce the chance of an additional page fault when performing an indirect call via a jump table. Do your benchmarks show otherwise?</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div><div class="m_-2534272377643883510m_6116811485634542887BodyFragment"><font size="2"><span style="font-size:10pt"><div class="m_-2534272377643883510m_6116811485634542887PlainText">I had to manually pick the functions which were renamed to “foo.cfi”
and replace them in the link order file. The process can be automated, but it’s a bit flaky if we have to do it to make CFI work.<br>
<br>
I attached 2 patches here: one is the direct call replacement and the other is moving type test lowering pass to run later in the pipeline. Interestingly, running the pass later seems to help with performance as well. Though the comment in PassManagerBuilder
implies that this pass needs to run early. Why is it different from full LTO?<br></div></span></font></div></div></blockquote><div><br></div><div>The issue with running the passes early is that the llvm.assume(llvm.type.test) pattern that WholeProgramDevirt looks for is somewhat fragile, so it needs to see the IR first before one of the other passes can break it. (This is probably something that we want to fix by using a different pattern.) I don't think that the same issue exists with any of the patterns that LowerTypeTests looks for, so it would probably be fine to move it later.</div><div><br></div><div>Peter</div><div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div><div class="m_-2534272377643883510m_6116811485634542887BodyFragment"><font size="2"><span style="font-size:10pt"><div class="m_-2534272377643883510m_6116811485634542887PlainText">
<br>
<br>
</div>
</span></font></div><div><div class="m_-2534272377643883510h5">
<div class="m_-2534272377643883510m_6116811485634542887BodyFragment"><font size="2"><span style="font-size:10pt">
<div class="m_-2534272377643883510m_6116811485634542887PlainText"><br>
<br>
<br>
> On Apr 27, 2018, at 11:04 AM, via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>> wrote:<br>
> <br>
> For the test case below, I get the following IR for main() on entry to ThinLTO backend invocation of LowerTypeTestsModule::lower(). This is after I moved this pass down in the pipeline so it’s invoked after inlining.<br>
> <br>
> The declarations for foo() and bar() are read in at the time of module import, Importer.importFunctions() in lto::thinBackend(). They do not have type metadata attached to them.<br>
> In lowerTypeTestCall() we check if the pointer in the type test is of a known type, so we look at bitcast and then select operands. foo and bar in select are global objects with no type metadata, so the type check cannot be guaranteed to be true and it can’t
be eliminated. In full LTO case this works as expected: both foo and bar are the same known type and type check is gone.<br>
> <br>
> Maybe the problem is not with renaming but with the missing type metadata in this particular case.<br>
> Though having so many direct calls routed through the jump table still seems problematic. Is there a feasible solution?<br>
> <br>
> ------------------------------<wbr>--<br>
> define hidden i32 @main(i32, i8** nocapture readnone) local_unnamed_addr #0 !type !3 !type !4 {<br>
> %3 = add nsw i32 %0, -2<br>
> store i32 %3, i32* @i, align 4, !tbaa !5<br>
> %4 = icmp sgt i32 %0, 1<br>
> %5 = select i1 %4, i32 ()* @foo, i32 ()* @bar<br>
> %6 = bitcast i32 ()* %5 to i8*, !nosanitize !9<br>
> %7 = tail call i1 @llvm.type.test(i8* nonnull %6, metadata !"_ZTSFivE"), !nosanitize !9<br>
> br i1 %7, label %10, label %8, !prof !10, !nosanitize !9<br>
> <br>
> ; <label>:8: <wbr> ; preds = %2<br>
> %9 = ptrtoint i32 ()* %5 to i64<br>
> tail call void @__ubsan_handle_cfi_check_fail<wbr>_abort(i8* getelementptr inbounds ({ i8, { [4 x i8]*, i32, i32 }, { i16, i16, [13 x i8] }* }, { i8, { [4 x i8]*, i32, i32 }, { i16, i16, [13 x i8] }* }* @anon.fad58de7366495db4650cfef<wbr>ac2fcd61.1, i64 0, i32 0),
i64 %9, i64 undef) #4, !nosanitize !9<br>
> unreachable, !nosanitize !9<br>
> <br>
> ; <label>:10: <wbr> ; preds = %2<br>
> %11 = tail call i32 %5() #5<br>
> ret i32 %11<br>
> }<br>
> <br>
> . . .<br>
> declare hidden i32 @foo() #3<br>
> declare hidden i32 @bar() #3<br>
> <br>
> <br>
> —————————test case---------------<br>
> b.c<br>
> =================<br>
> typedef int (*fptr_t) (void);<br>
> fptr_t get_fptr();<br>
> extern int i;<br>
> <br>
> int main(int argc, char *argv[])<br>
> {<br>
> i = argc - 2;<br>
> fptr_t fp = get_fptr();<br>
> return fp();<br>
> }<br>
> <br>
> v.c<br>
> ================<br>
> int i;<br>
> typedef int (*fptr_t) (void);<br>
> int foo(void) { return 11; }<br>
> int bar(void) { return 22; }<br>
> fptr_t get_fptr(void) { return (i >= 0) ? foo : bar; }<br>
> <br>
> <br>
>> On Apr 26, 2018, at 5:29 PM, Peter Collingbourne <<a href="mailto:peter@pcc.me.uk" target="_blank">peter@pcc.me.uk</a>> wrote:<br>
>> <br>
>>> We could probably tolerate a certain amount of unused jump table entries. However, I just realized that all non-inline imported calls end up going through a jump table entry. Is that correct?<br>
>> <br>
>> In fact it is all calls that go through a function pointer type that is used anywhere in the program for an indirect call, but depending on your program that could be very close to "yes".<br>
>> <br>
>>> Moving the type check lowering pass further down the pipeline (after inlining) still does not solve the problem because CFI renaming happens early and symbols attached to the jump table do not have a matching type.<br>
>> <br>
>> As far as I know, renaming happens during the LowerTypeTests pass, after the type checks are lowered.<br>
>> Lowering: <a href="http://llvm-cs.pcc.me.uk/lib/Transforms/IPO/LowerTypeTests.cpp#1620" target="_blank">
http://llvm-cs.pcc.me.uk/lib/T<wbr>ransforms/IPO/LowerTypeTests.c<wbr>pp#1620</a><br>
>> Renaming: <a href="http://llvm-cs.pcc.me.uk/lib/Transforms/IPO/LowerTypeTests.cpp#1642" target="_blank">
http://llvm-cs.pcc.me.uk/lib/T<wbr>ransforms/IPO/LowerTypeTests.c<wbr>pp#1642</a><br>
>> Do you have an example of what you are seeing?<br>
>> <br>
>> Peter<br>
>> <br>
>> On Thu, Apr 26, 2018 at 4:54 PM, <<a href="mailto:dmitry.mikulin@sony.com" target="_blank">dmitry.mikulin@sony.com</a>> wrote:<br>
>> Hi Peter,<br>
>> <br>
>> We could probably tolerate a certain amount of unused jump table entries. However, I just realized that all non-inline imported calls end up going through a jump table entry. Is that correct? Initially I thought you meant calls promoted from indirect. While
this can be fixed by replacing direct calls to jump tables with direct calls to real targets, I found other cases where ThinLTO+CFI has issues.<br>
>> <br>
>> In ThinLTO backend, type test lowering happens very early in the pipeline, before inlining. When the type check after the call to get_fptr() is lowered (in my original example, below), the compiler cannot see that both targets belong to the same type and
that the type check will always return ‘true’ and can be eliminated. Moving the type check lowering pass further down the pipeline (after inlining) still does not solve the problem because CFI renaming happens early and symbols attached to the jump table do
not have a matching type.<br>
>> <br>
>> I’m trying to think if there’s a way to delay renaming until ThinLTO backend type check lowering pass. It would help with solving both problems.<br>
>> <br>
>> Thanks.<br>
>> Dmitry.<br>
>> <br>
>> <br>
>> <br>
>>>>>> <br>
>>>>>> a.c<br>
>>>>>> =============================<br>
>>>>>> typedef int (*fptr_t) (void);<br>
>>>>>> fptr_t get_fptr();<br>
>>>>>> int main(int argc, char *argv[])<br>
>>>>>> {<br>
>>>>>> fptr_t fp = get_fptr();<br>
>>>>>> return fp();<br>
>>>>>> }<br>
>>>>>> <br>
>>>>>> <br>
>>>>>> b.c<br>
>>>>>> =============================<br>
>>>>>> typedef int (*fptr_t) (void);<br>
>>>>>> int foo(void) { return 11; }<br>
>>>>>> int bar(void) { return 22; }<br>
>>>>>> <br>
>>>>>> static fptr_t fptr = bar;<br>
>>>>>> static int i = 53;<br>
>>>>>> <br>
>>>>>> fptr_t get_fptr(void)<br>
>>>>>> {<br>
>>>>>> if (i >= 0)<br>
>>>>>> fptr = foo;<br>
>>>>>> else<br>
>>>>>> fptr = bar;<br>
>>>>>> <br>
>>>>>> return fptr;<br>
>>>>>> }<br>
>>>>>> <br>
>> <br>
>> <br>
>> <br>
>> <br>
>>> On Apr 19, 2018, at 6:18 PM, Peter Collingbourne <<a href="mailto:peter@pcc.me.uk" target="_blank">peter@pcc.me.uk</a>> wrote:<br>
>>> <br>
>>> Regarding the orderfile, yes, I was thinking more about ordering the real functions.<br>
>>> <br>
>>> In that case it sounds like your best option may be to implement the optimization pass to make direct calls go directly to the real function. From a performance perspective I don't think it would make much difference if there are unused jump table entries.<br>
>>> <br>
>>> Peter<br>
>>> <br>
>>> On Thu, Apr 19, 2018 at 6:09 PM, via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>> wrote:<br>
>>> Teresa, Peter,<br>
>>> <br>
>>> Thanks for your help!<br>
>>> I need to re-run my experiments as the compiler I used did not have the latest changes like r327254.<br>
>>> The fact that the decision about routing calls through jump table entries is made early may be problematic. In my experiments with FreeBSD kernel, ThinLTO produced thousands jump table entries compared to only dozens with full LTO. As for re-ordering jump
table entries, I don’t think it’s going to work as they are placed in the same section. Including *.cfi names into a link order file will take care of re-ordering real functions routed through jump table entries, but in our case we need to force some functions
to be on the same page. So not having jump table entries for the functions that don't really need them would be ideal.<br>
>>> <br>
>>> Thanks.<br>
>>> Dmitry.<br>
>>> <br>
>>> <br>
>>>> On Apr 18, 2018, at 6:11 PM, Teresa Johnson <<a href="mailto:tejohnson@google.com" target="_blank">tejohnson@google.com</a>> wrote:<br>
>>>> <br>
>>>> <br>
>>>> <br>
>>>> On Wed, Apr 18, 2018 at 4:49 PM, <<a href="mailto:dmitry.mikulin@sony.com" target="_blank">dmitry.mikulin@sony.com</a>> wrote:<br>
>>>> Hi Teresa,<br>
>>>> <br>
>>>> Thanks for the info!<br>
>>>> This example is my attempt to reduce FreeBSD kernel to something more manageable :)<br>
>>>> <br>
>>>> I will take a look at why globals are not being imported in this case. What’s the best tool to look into ThinLTO objects and their summaries? Most dumping tools don’t seem to like ThinLTO bitcode files…<br>
>>>> <br>
>>>> Sadly there isn't a really great way to dump the summaries. =( There was a patch awhile back by a GSOC student to dump in YAML format, but there was resistance from some who preferred dumping to llvm assembly via llvm-dis and support reading in the summary
from llvm assembly. It's been on my list of things to do, hasn't yet risen high enough in priority to work on that. For now, you have to use llvm-bcanalyzer -dump and look at the raw format.<br>
>>>> <br>
>>>> Teresa<br>
>>>> <br>
>>>> <br>
>>>> Hopefully Peter can chime in regarding CFI related issues.<br>
>>>> <br>
>>>> Thanks.<br>
>>>> Dmitry. <br>
>>>> <br>
>>>> <br>
>>>>> On Apr 17, 2018, at 9:37 AM, Teresa Johnson <<a href="mailto:tejohnson@google.com" target="_blank">tejohnson@google.com</a>> wrote:<br>
>>>>> <br>
>>>>> Hi Dmitry,<br>
>>>>> <br>
>>>>> Sorry for the late reply. For CFI specific code generation, pcc is a better person to answer. But on the issue of global variables being optimized, that hasn't happened yet. That would be great if you wanted to pick that up!<br>
>>>>> <br>
>>>>> In your original email example, it seems like the file static i=53 could be constant propagated since there are no other defs, and the code in get_fptr simplified during the compile step, but I assume this is part of a more complex example where it is
not possible to do this? Also note that with r327254 we started importing global variables. Do you know why we don't import in your case? I wonder if it has to do with it being CFI inserted code?<br>
>>>>> <br>
>>>>> Teresa<br>
>>>>> <br>
>>>>> On Tue, Apr 17, 2018 at 9:17 AM <<a href="mailto:dmitry.mikulin@sony.com" target="_blank">dmitry.mikulin@sony.com</a>> wrote:<br>
>>>>> I watched Teresa’s talk on ThinLTO from last year’s CppCon, and it sounded like adding global variable information to the summaries was in the works, or at least in planning. Can someone (Teresa?) please share the current status? If it’s part of future
plans, are there any specific proposals that can be picked up and worked on?<br>
>>>>> <br>
>>>>> Thanks!<br>
>>>>> <br>
>>>>> <br>
>>>>>> On Apr 9, 2018, at 6:51 PM, via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>> wrote:<br>
>>>>>> <br>
>>>>>> Hi,<br>
>>>>>> <br>
>>>>>> I’m working on setting up ThinLTO+CFI for a C application which uses a lot of function pointers. While functionally it appears stable, it’s performance is significantly degraded, to the tune of double digit percentage points compared to regular LTO+CFI.<br>
>>>>>> <br>
>>>>>> Looking into possible causes I see that under ThinLTO+CFI iCall type checks almost always generate jump table entries for indirect calls, which creates another level of indirection for every such call. On top of that it breaks the link order layout because
real function names point to jump table entries. It appears that I’m hitting a limitation in ThinLTO on how much information it can propagate across modules, particularly information about constants. In the example below, the fact that “i” is effectively a
constant, is lost under ThinLTO, and the inlined copy of b.c:get_fptr() in a.c does not eliminate the conditional, which, for CFI purposes requires to generate a type check/jump table.<br>
>>>>>> <br>
>>>>>> I was wondering if there was a way to mitigate this limitation.<br>
>>>>>> <br>
>>>>>> a.c<br>
>>>>>> =============================<br>
>>>>>> typedef int (*fptr_t) (void);<br>
>>>>>> fptr_t get_fptr();<br>
>>>>>> int main(int argc, char *argv[])<br>
>>>>>> {<br>
>>>>>> fptr_t fp = get_fptr();<br>
>>>>>> return fp();<br>
>>>>>> }<br>
>>>>>> <br>
>>>>>> <br>
>>>>>> b.c<br>
>>>>>> =============================<br>
>>>>>> typedef int (*fptr_t) (void);<br>
>>>>>> int foo(void) { return 11; }<br>
>>>>>> int bar(void) { return 22; }<br>
>>>>>> <br>
>>>>>> static fptr_t fptr = bar;<br>
>>>>>> static int i = 53;<br>
>>>>>> <br>
>>>>>> fptr_t get_fptr(void)<br>
>>>>>> {<br>
>>>>>> if (i >= 0)<br>
>>>>>> fptr = foo;<br>
>>>>>> else<br>
>>>>>> fptr = bar;<br>
>>>>>> <br>
>>>>>> return fptr;<br>
>>>>>> }<br>
>>>>>> <br>
>>>>>> ______________________________<wbr>_________________<br>
>>>>>> LLVM Developers mailing list<br>
>>>>>> <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
>>>>>> <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" target="_blank">http://lists.llvm.org/cgi-bin/<wbr>mailman/listinfo/llvm-dev</a><br>
>>>>> <br>
>>>>> <br>
>>>>> <br>
>>>>> -- <br>
>>>>> Teresa Johnson | Software Engineer | <a href="mailto:tejohnson@google.com" target="_blank">tejohnson@google.com</a> | 408-460-2413<br>
>>>> <br>
>>>> <br>
>>>> <br>
>>>> <br>
>>>> -- <br>
>>>> Teresa Johnson | Software Engineer | <a href="mailto:tejohnson@google.com" target="_blank">tejohnson@google.com</a> | 408-460-2413<br>
>>> <br>
>>> <br>
>>> ______________________________<wbr>_________________<br>
>>> LLVM Developers mailing list<br>
>>> <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
>>> <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" target="_blank">http://lists.llvm.org/cgi-bin/<wbr>mailman/listinfo/llvm-dev</a><br>
>>> <br>
>>> <br>
>>> <br>
>>> <br>
>>> -- <br>
>>> -- <br>
>>> Peter<br>
>> <br>
>> <br>
>> <br>
>> <br>
>> -- <br>
>> -- <br>
>> Peter<br>
> <br>
> ______________________________<wbr>_________________<br>
> LLVM Developers mailing list<br>
> <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br>
> <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" target="_blank">http://lists.llvm.org/cgi-bin/<wbr>mailman/listinfo/llvm-dev</a><br>
<br>
</div>
</span></font></div>
</div></div></div>
</blockquote></div><br><br clear="all"><div><br></div>-- <br><div class="m_-2534272377643883510gmail_signature" data-smartmail="gmail_signature"><div dir="ltr">-- <div>Peter</div></div></div>
</div></div>