RFC: a more detailed design for ThinLTO + vcall CFI

Hi all,

As promised, here is a brain dump on how I see CFI for vcalls working under ThinLTO. Most of this has been prototyped, so the design does appear to be sound. For context on how CFI currently works under regular LTO, please read:

http://llvm.org/docs/TypeMetadata.html

http://clang.llvm.org/docs/ControlFlowIntegrityDesign.html
http://clang.llvm.org/docs/LTOVisibility.html

==== Summary extensions ====

The combined summary index would be extended to include a mapping from type identifiers to “resolutions”. The resolution would control what type of code we generate to create a CFI check for that type identifier. Here are the resolutions that we would support:

Inline32, Inline64, SingleBit: these would cause us to generate code as described in “Short Inline Bit Vectors” in the design document: http://clang.llvm.org/docs/ControlFlowIntegrityDesign.html#short-inline-bit-vectors

AllOnes: this would cause us to generate code as described in “Eliminating Bit Vector Checks for All-Ones Bit Vectors” in the design document: http://clang.llvm.org/docs/ControlFlowIntegrityDesign.html#eliminating-bit-vector-checks-for-all-ones-bit-vectors

Unsat: no vtable is a member of that type identifier, so we can simply replace type checks for that type identifier with “false”
ByteArray: we emit the general form of the type check, similar to the one shown at the end of http://clang.llvm.org/docs/ControlFlowIntegrityDesign.html#forward-edge-cfi-for-virtual-calls just before “Optimizations”

Armed with that information, we have a general idea of what the code to implement the type check would look like. In fact, given one of these values, the code will be identical to any other check that uses that resolution, modulo the constant values embedded in the code.

To expand on what I mean by “constant values”, let’s look at a typical CFI check in the ByteArray case. Consider this module (based on test/Transforms/LowerTypeTests/simple.ll):

@a = constant i32 1, !type !0, !type !2
@b = constant [63 x i32] zeroinitializer, !type !0, !type !1
@c = constant i32 3, !type !1, !type !2
@d = constant [2 x i32] [i32 4, i32 5], !type !3

!0 = !{i32 0, !“typeid1”}
!3 = !{i32 4, !“typeid1”}
!1 = !{i32 0, !“typeid2”}

!2 = !{i32 0, !“typeid3”}

define i1 @baz(i32* %p) {

%x = call i1 @llvm.type.test(i8* %pi8, metadata !“typeid3”)

ret i1 %x
}

Here is the IR after the LowerTypeTests pass has run. I’ve marked the places where we use a constant value:

define i1 @baz(i32* %p) {
%pi8 = bitcast i32* %p to i8*
%1 = ptrtoint i8* %pi8 to i32
%2 = sub i32 %1, ptrtoint ({ i32, [0 x i8], [63 x i32], [4 x i8], i32, [0 x i8], [2 x i32] }* @0 to i32)
%3 = lshr i32 %2, 2 ; CONSTANT: rotate count
%4 = shl i32 %2, 30 ; CONSTANT: 32-rotate count
%5 = or i32 %3, %4
%6 = icmp ult i32 %5, 66 ; CONSTANT: size of byte array
br i1 %6, label %7, label %12

; :7: ; preds = %0
%8 = getelementptr i8, i8* @bits_use, i32 %5
%9 = load i8, i8* %8
%10 = and i8 %9, 2 ; CONSTANT: bit mask
%11 = icmp ne i8 %10, 0
br label %12

; :12: ; preds = %0, %7
%13 = phi i1 [ false, %0 ], [ %11, %7 ]
ret i1 %13
}

Here is what the asm for the above looks like:

baz:
leaq .L__unnamed_1(%rip), %rax

subl %eax, %edi
roll $30, %edi ; CONSTANT: 32-rotate count
cmpl $65, %edi ; CONSTANT: size of byte array
ja .LBB2_1

movslq %edi, %rax
leaq .Lbits_use(%rip), %rcx
movb (%rax,%rcx), %al
andb $2, %al ; CONSTANT: bit mask
shrb %al
retq
.LBB2_1:
xorl %eax, %eax
retq

A naive summary encoding would map a type identifier to a tuple of (resolution, rotate count, size of byte array, bit mask), and pull the latter three out of the summary as constants. However, the disadvantage of hard coding the constants in the IR like this is that any change to one of the constants will invalidate any cache entry that depends on a constant value. For example, if I add a class to a hierarchy, that would most likely increase the size of the byte array, which may invalidate every cache entry containing a check for a class in that hierarchy. To avoid this, we only include resolutions in the summary and obtain the constants using absolute symbol references. Here is what the asm code may look like:

baz:
leaq __typeid_typeid3_global_addr(%rip), %rax
subl %eax, %edi
rorl $__typeid_typeid3_rotate_count, %edi
cmpl $__typeid_typeid3_size, %edi
ja .LBB2_1

movslq %edi, %rax
leaq __typeid_typeid3_byte_array(%rip), %rcx
movb (%rax,%rcx), %al
andb $__typeid_typeid3_bitmask, %al
shrb %al
retq
.LBB2_1:
xorl %eax, %eax
retq

The appropriate representation for this at the IR level is the subject of the RFC entitled ‘Absolute or “fixed address” symbols as immediate operands’. We can imagine, though, that it could look something like this:

@__typeid_typeid3_rotate_count = external globalconst i8
@__typeid_typeid3_size = external globalconst i32
@__typeid_typeid3_bit_mask = external globalconst i8
@__typeid_typeid3_byte_array = external global i8

@__typeid_typeid3_global_addr = external global i8

define i1 @baz(i32* %p) {
%pi8 = bitcast i32* %p to i8*
%1 = ptrtoint i8* %pi8 to i32
%2 = sub i32 %1, ptrtoint (i8* @__typeid_typeid3_global_addr to i32)
%3 = lshr i32 %2, i8 @__typeid_typeid3_rotate_count
%4 = shl i32 %2, sub i8 (i8 32, i8 @__typeid_typeid3_rotate_count)
%5 = or i32 %3, %4
%6 = icmp ult i32 %5, i32 @__typeid_typeid3_size
br i1 %6, label %7, label %12

; :7: ; preds = %0
%8 = getelementptr i8, i8* @__typeid_typeid3_byte_array, i32 %5
%9 = load i8, i8* %8
%10 = and i8 %9, @__typeid_typeid3_bitmask
%11 = icmp ne i8 %10, 0
br label %12

; :12: ; preds = %0, %7
%13 = phi i1 [ false, %0 ], [ %11, %7 ]
ret i1 %13
}

==== Getting resolutions into the summary ====

Now that we’ve looked at how code is generated for the individual modules, let’s look at how the resolutions are computed and put into the summary.

The first step happens when we compile the translation unit with clang. I propose to change the bitcode format for any module that defines virtual tables with hidden LTO visibility. The bitcode would contain two modules: one to be compiled with regular LTO and the other to be compiled with ThinLTO. The regular LTO module would contain only the following:

  • the definitions of those virtual tables that have hidden LTO visibility

  • the !type metadata for those virtual tables

  • a list of type tests required by the corresponding ThinLTO module, in a GlobalMDNode named “llvm.export.type.tests”

The ThinLTO module would contain the rest of the original module.

For example, here is what the regular LTO module for our example above would look like:

@a = constant i32 1, !type !0, !type !2
@b = constant [63 x i32] zeroinitializer, !type !0, !type !1
@c = constant i32 3, !type !1, !type !2
@d = constant [2 x i32] [i32 4, i32 5], !type !3

!0 = !{i32 0, !“typeid1”}
!3 = !{i32 4, !“typeid1”}
!1 = !{i32 0, !“typeid2”}
!2 = !{i32 0, !“typeid3”}

!8 = !{!“typeid3”}

!llvm.export.type.tests = !{!8}

and here is the ThinLTO module:

define i1 @baz(i32* %p) {
%x = call i1 @llvm.type.test(i8* %pi8, metadata !“typeid3”)
ret i1 %x
}

The regular LTO modules are merged and compiled in the usual way, so we would end up merging the type definitions together with the list of required data.

When the LowerTypeTests pass is given a module with the “llvm.export.type.tests” global MDNode, it “exports” each of the type identifiers mentioned in the MDNode by storing a resolution in the combined summary, and by creating definitions of each of the symbols that shall be required to satisfy the link time dependencies in the individual ThinLTO modules. Here is an example of what the combined module would look like:

@0 = […] ; combined global for a, b, c and d
@bits = […] ; combined global byte array

@__typeid_typeid3_global_addr = hidden alias i8, bitcast ({…}* @0 to i8*)
@__typeid_typeid3_rotate_count = hidden globalconst i8 2

@__typeid_typeid3_size = hidden globalconst i32 65

@__typeid_typeid3_byte_array = hidden alias i8, i8* @bits
@__typeid_typeid3_bit_mask = hidden globalconst i8 2

The ThinLTO backend processes would run after regular LTO, so they would have access to the resolutions in the summary. If the module summary is present, LowerTypeTests will “import” the appropriate type identifier by generating code containing external references to these symbols, as described previously.

Thanks,

Hi Peter,

Thanks for sending this and sorry for the slow response. Some questions below.

Teresa

Hi Peter,

Thanks for sending this and sorry for the slow response. Some questions
below.

Teresa

Hi all,

As promised, here is a brain dump on how I see CFI for vcalls working
under ThinLTO. Most of this has been prototyped, so the design does appear
to be sound. For context on how CFI currently works under regular LTO,
please read:

http://llvm.org/docs/TypeMetadata.html
http://clang.llvm.org/docs/ControlFlowIntegrityDesign.html
http://clang.llvm.org/docs/LTOVisibility.html

==== Summary extensions ====

The combined summary index would be extended to include a mapping from
type identifiers to "resolutions". The resolution would control what type
of code we generate to create a CFI check for that type identifier. Here
are the resolutions that we would support:

Inline32, Inline64, SingleBit: these would cause us to generate code as
described in "Short Inline Bit Vectors" in the design document:
http://clang.llvm.org/docs/ControlFlowIntegrityDes
ign.html#short-inline-bit-vectors
AllOnes: this would cause us to generate code as described
in "Eliminating Bit Vector Checks for All-Ones Bit Vectors" in the design
document: http://clang.llvm.org/docs/ControlFlowIntegrityDes
ign.html#eliminating-bit-vector-checks-for-all-ones-bit-vectors
Unsat: no vtable is a member of that type identifier, so we can simply
replace type checks for that type identifier with "false"
ByteArray: we emit the general form of the type check, similar to the one
shown at the end of http://clang.llvm.org/docs/
ControlFlowIntegrityDesign.html#forward-edge-cfi-for-virtual-calls just
before "Optimizations"

Armed with that information, we have a general idea of what the code to
implement the type check would look like. In fact, given one of these
values, the code will be identical to any other check that uses that
resolution, modulo the constant values embedded in the code.

To expand on what I mean by "constant values", let's look at a typical
CFI check in the ByteArray case. Consider this module (based on
test/Transforms/LowerTypeTests/simple.ll):

@a = constant i32 1, !type !0, !type !2
@b = constant [63 x i32] zeroinitializer, !type !0, !type !1
@c = constant i32 3, !type !1, !type !2
@d = constant [2 x i32] [i32 4, i32 5], !type !3

!0 = !{i32 0, !"typeid1"}
!3 = !{i32 4, !"typeid1"}
!1 = !{i32 0, !"typeid2"}
!2 = !{i32 0, !"typeid3"}

define i1 @baz(i32* %p) {
  %x = call i1 @llvm.type.test(i8* %pi8, metadata !"typeid3")
  ret i1 %x
}

Here is the IR after the LowerTypeTests pass has run. I've marked the
places where we use a constant value:

define i1 @baz(i32* %p) {
  %pi8 = bitcast i32* %p to i8*
  %1 = ptrtoint i8* %pi8 to i32
  %2 = sub i32 %1, ptrtoint ({ i32, [0 x i8], [63 x i32], [4 x i8], i32,
[0 x i8], [2 x i32] }* @0 to i32)
  %3 = lshr i32 %2, 2 ; CONSTANT: rotate count
  %4 = shl i32 %2, 30 ; CONSTANT: 32-rotate count
  %5 = or i32 %3, %4
  %6 = icmp ult i32 %5, 66 ; CONSTANT: size of byte array
  br i1 %6, label %7, label %12

; <label>:7: ; preds = %0
  %8 = getelementptr i8, i8* @bits_use, i32 %5
  %9 = load i8, i8* %8
  %10 = and i8 %9, 2 ; CONSTANT: bit mask
  %11 = icmp ne i8 %10, 0
  br label %12

; <label>:12: ; preds = %0, %7
  %13 = phi i1 [ false, %0 ], [ %11, %7 ]
  ret i1 %13
}

Here is what the asm for the above looks like:

baz:
leaq .L__unnamed_1(%rip), %rax
subl %eax, %edi
roll $30, %edi ; CONSTANT: 32-rotate count
cmpl $65, %edi ; CONSTANT: size of byte array
ja .LBB2_1

movslq %edi, %rax
leaq .Lbits_use(%rip), %rcx
movb (%rax,%rcx), %al
andb $2, %al ; CONSTANT: bit mask
shrb %al
retq
.LBB2_1:
xorl %eax, %eax
retq

A naive summary encoding would map a type identifier to a tuple of
(resolution, rotate count, size of byte array, bit mask), and pull the
latter three out of the summary as constants. However, the disadvantage of
hard coding the constants in the IR like this is that any change to one of
the constants will invalidate any cache entry that depends on a constant
value. For example, if I add a class to a hierarchy, that would most likely
increase the size of the byte array, which may invalidate every cache entry
containing a check for a class in that hierarchy. To avoid this, we only
include resolutions in the summary and obtain the constants using absolute
symbol references. Here is what the asm code may look like:

baz:
leaq __typeid_typeid3_global_addr(%rip), %rax
subl %eax, %edi
rorl $__typeid_typeid3_rotate_count, %edi
cmpl $__typeid_typeid3_size, %edi
ja .LBB2_1

movslq %edi, %rax
leaq __typeid_typeid3_byte_array(%rip), %rcx
movb (%rax,%rcx), %al
andb $__typeid_typeid3_bitmask, %al
shrb %al
retq
.LBB2_1:
xorl %eax, %eax
retq

The appropriate representation for this at the IR level is the subject of
the RFC entitled 'Absolute or "fixed address" symbols as immediate
operands'. We can imagine, though, that it could look something like this:

@__typeid_typeid3_rotate_count = external globalconst i8
@__typeid_typeid3_size = external globalconst i32
@__typeid_typeid3_bit_mask = external globalconst i8
@__typeid_typeid3_byte_array = external global i8
@__typeid_typeid3_global_addr = external global i8

Naive question: These will be defined in the object file containing the
definition of the key method and therefore the vtable definition, right? So
only that object would need to be recompiled on a change like the one you
mentioned earlier when talking about caching (adding a class to a
hierarchy), right?

These will all be defined in the object file produced during regular LTO.
In principle we could support caching that object, but I was thinking more
about the objects produced by the thin backends.

define i1 @baz(i32* %p) {
  %pi8 = bitcast i32* %p to i8*
  %1 = ptrtoint i8* %pi8 to i32
  %2 = sub i32 %1, ptrtoint (i8* @__typeid_typeid3_global_addr to i32)
  %3 = lshr i32 %2, i8 @__typeid_typeid3_rotate_count
  %4 = shl i32 %2, sub i8 (i8 32, i8 @__typeid_typeid3_rotate_count)
  %5 = or i32 %3, %4
  %6 = icmp ult i32 %5, i32 @__typeid_typeid3_size
  br i1 %6, label %7, label %12

; <label>:7: ; preds = %0
  %8 = getelementptr i8, i8* @__typeid_typeid3_byte_array, i32 %5
  %9 = load i8, i8* %8
  %10 = and i8 %9, @__typeid_typeid3_bitmask
  %11 = icmp ne i8 %10, 0
  br label %12

; <label>:12: ; preds = %0, %7
  %13 = phi i1 [ false, %0 ], [ %11, %7 ]
  ret i1 %13
}

==== Getting resolutions into the summary ====

Now that we've looked at how code is generated for the individual
modules, let's look at how the resolutions are computed and put into the
summary.

The first step happens when we compile the translation unit with clang. I
propose to change the bitcode format for any module that defines virtual
tables with hidden LTO visibility. The bitcode would contain two modules:
one to be compiled with regular LTO and the other to be compiled with
ThinLTO. The regular LTO module would contain only the following:

   - the definitions of those virtual tables that have hidden LTO
   visibility

Another naive question: What happens in the non-hidden case and why is it
different here?

We do not normally enable CFI checks for virtual calls to non-hidden
classes, so there's nothing to do about that case.

   - the !type metadata for those virtual tables
   - a list of type tests required by the corresponding ThinLTO module,
   in a GlobalMDNode named "llvm.export.type.tests"

The ThinLTO module would contain the rest of the original module.

The two modules would presumably be distinguished because the regular LTO
module would not have a summary block, because the summary for it would be
generated on the fly during LTO processing of the merged regular LTO
modules, right?

Right.

For example, here is what the regular LTO module for our example above
would look like:

@a = constant i32 1, !type !0, !type !2
@b = constant [63 x i32] zeroinitializer, !type !0, !type !1
@c = constant i32 3, !type !1, !type !2
@d = constant [2 x i32] [i32 4, i32 5], !type !3

!0 = !{i32 0, !"typeid1"}
!3 = !{i32 4, !"typeid1"}
!1 = !{i32 0, !"typeid2"}
!2 = !{i32 0, !"typeid3"}

!8 = !{!"typeid3"}
!llvm.export.type.tests = !{!8}

and here is the ThinLTO module:

define i1 @baz(i32* %p) {
  %x = call i1 @llvm.type.test(i8* %pi8, metadata !"typeid3")
  ret i1 %x
}

The regular LTO modules are merged and compiled in the usual way, so we
would end up merging the type definitions together with the list of
required data.

When the LowerTypeTests pass is given a module with the
"llvm.export.type.tests" global MDNode, it "exports" each of the type
identifiers mentioned in the MDNode by storing a resolution in the combined
summary, and by creating

Assuming I understand this correctly, that means that there would be some
kind of type table section in the combined summary index bitcode file, that
would look something like the following for the above case:

<TYPEID_SUMMARY_BLOCK ...>
    <TYPE abbrevid=... resolutionid, "typeid3">
</TYPEID_SUMMARY_BLOCK>

and included in the on-disk summary as a StringMap, right?

Right.

definitions of each of the symbols that shall be required to satisfy the
link time dependencies in the individual ThinLTO modules. Here is an
example of what the combined module would look like:

@0 = [...] ; combined global for a, b, c and d
@bits = [...] ; combined global byte array

@__typeid_typeid3_global_addr = hidden alias i8, bitcast ({...}* @0 to
i8*)
@__typeid_typeid3_rotate_count = hidden globalconst i8 2
@__typeid_typeid3_size = hidden globalconst i32 65
@__typeid_typeid3_byte_array = hidden alias i8, i8* @bits
@__typeid_typeid3_bit_mask = hidden globalconst i8 2

The ThinLTO backend processes would run after regular LTO, so they would
have access to the resolutions in the summary. If the module summary is
present, LowerTypeTests will "import" the appropriate type identifier by
generating code containing external references to these symbols, as
described previously.

So in the above case where the summary says "typeid3" -> ByteArray,
ThinLTO would on the ThinLTO module would see the @llvm.type.test of
typeid3 in the IR, consult the type resolution table in the combined
summary, and know to generate decls for those @__typeid_typeid3_* variables
and the associated code, right?

Right.

One thing that needs some thought is for the distributed case, where the
thin link serializes out individual index files for each ThinLTO module.
Presumably that step would first create the merged regular LTO module and
emit a regular object file for it during the thin link - we'll just need to
anticipate that under the appropriate options. But the bigger issue that
needs thought now is how to know which subset of the type resolution
summary table to emit into each individual index file. That needs to be
known when processing the ThinLTO modules through the
WriteIndexesThinBackend - won't the ThinLTO summaries need to include info
about which typeids are referenced from that module? Otherwise you'd have
to conservatively emit the entire type resolution summary table in each
individual index file - doable but presumably a fair amount of overhead.

Yes, that's a good point. Note that this is exactly the list of type
identifiers that is emitted into "llvm.export.type.tests" in the regular
part of each module. If we store this list in the individual module
summaries instead, we can use it to produce both "llvm.export.type.tests"
for the combined regular LTO module and the type resolution summaries in
the individual index files.

Thanks,