Mike,

Just my two cents:

I think adding intrinsic libraries is a mixed bag, for the following reasons:

1) The intrinsic libraries are completely internal to the compilers. Breaking down that toolchain/code barrier is not a good idea if we want the project to compile using a good variety of compilers. The API is unstable (as it's internal to the compiler + version) and sometimes undocumented; while in reality the ABI doesn't really change (at least in the LLVM/GCC world, not sure about the other compilers), it's something to consider.

2) Linking the compiler's builtin libraries would fix our issues, except that it doesn't work in cases where object files are tagged with ABI (such as hard FP vs soft FP).

On the other hand, the latest libgcc_s.so symbol I can find was introduced in GCC 7.0 (2017) and is completely unrelated to any simple 64-bit integer math we may want to do. So, in our very simple integer-mathy case, it may end up being a solid option.
LLVM also has a testsuite for its intrinsics, which is handy :)

It's worth noting that (at least) LLVM/GCC expect standard freestanding <string.h> functions like memcpy, strlen, ... to be available; I think it would be a good idea to supply these and drop "-fno-builtin" so that the compiler can do its job and optimize code further.

Also, if we're going to add runtime library code, UBSan and Asan might be worth a look ;)

Best regards,
Pedro

On Mon, Jan 24, 2022 at 5:29 PM Michael D Kinney <michael.d.kinney@intel.com> wrote:

Hi Kilian,

 

I am in favor of an intrinsic lib to improve the EDK II development environment.

 

This has already been done for ARM compilers.  The solution should mirror that approach.

 

It would be best if we had source code (either in the edk2 repo or through a submodule) for

the required intrinsic APIs.  If source code is not possible and we have to use a binary, then

that must be accessed through a submodule.  The edk2 repo does not host binaries.  We use

repos such as edk2-non-osi for binaries.

 

We also have to provide a solution that works with supported compilers (VS, GCC, CLANG, XCODE).

 

One of the challenges is that compilers are allowed to add/remove/modify intrinsic APIs

across compiler releases.  We would need to define a solution that will work if there are

these types of changes, which would potentially mean a different instance of the intrinsic

library for each tool chain tag.  I think in practice, the intrinsic APIs we are seeing from use

of C code from submodules is a very limited set that do not change across compiler releases,

so the maintenance of these intrinsic libs would be manageable.

 

If we go down the source code path, we can break it up into the following tasks:

  1. Identify the specific subset of intrinsic APIs from each compiler that is required for the edk2 use cases. 
  2. Obtain the function prototype and full documentation for each intrinsic API to support implementation and unit tests.
  3. Implement the APIs for all compilers.
  4. Implement unit tests for all APIs for all compilers using UnitTestFrameworkPkg unit tests.
  5. Update MdeLibs.dsc.inc with the NULL instances for the intrinsic libs
  6. Remove intrinsic APIs from EDK II modules that currently maintain their own implementations of intrinsic APIs.

 

Best regards,

 

Mike

 

From: Kilian Kegel <kilian_kegel@outlook.com>
Sent: Monday, January 24, 2022 8:25 AM
To: devel@edk2.groups.io; Kinney, Michael D <michael.d.kinney@intel.com>; kraxel@redhat.com; Yao, Jiewen <jiewen.yao@intel.com>; Sean Brogan <sean.brogan@microsoft.com>; Bret Barkelew <Bret.Barkelew@microsoft.com>
Cc: devel@edk2.groups.io; Wang, Jian J <jian.j.wang@intel.com>; Jiang, Guomin <guomin.jiang@intel.com>; Pawel Polawski <ppolawsk@redhat.com>; Lu, XiaoyuX <xiaoyux.lu@intel.com>
Subject: RE: [edk2-devel] [PATCH 00/24] CryptoPkg/openssl: update openssl submodule to v3.0

 

The 64-bit integer math intrinsics and other intrinsic

problems could be solved easily for ever:

 

  1. Putting all .OBJ files together from LIBCMT.H or INT64.LIB (for ll*.obj and ull*.obj only)

ltod3.obj

ftol2.obj

lldiv.obj

lldvrm.obj

llmul.obj

llrem.obj

llshl.obj

llshr.obj

ulldiv.obj

ulldvrm.obj

ullrem.obj

ullshr.obj

memcmp.obj

memcpycpy.obj

                and adjust for usability in EDK2 (remove / solve further internal dependencies or rewrite “*cpy” and “*cmp” functions)

This is already done in IntrinsicLib.lib for some of the above functions, just complete this task!

  1. Put all the .OBJ into a e.g. edk2\Conf \“MSFTINTRINx86-32<compilerversion>.lib”
  2. Update the MSFT_DEF.txt tool chain definition path

DEBUG_*_IA32_DLINK_FLAGS     = %CONF_PATH%\ MSFTINTRINx86-32<compilerversion>.lib

RELEASE_*_IA32_DLINK_FLAGS   = %CONF_PATH%\ MSFTINTRINx86-32<compilerversion>.lib

  1. Resolve build conflicts with other existing intrinsic libraries from CryptoPkg, RedfishPkg… – remove these libraries

 

From now on all existing 32Bit components have access to their own compiler intrinsics without

touching any .INF file and the problem is instantly gone.

 

Please do the same for

  • GCCINTRINx86-32<compilerversion>.lib

 

Leave the intrinsic restrictions behind and just provide all required intrinsics the compiler needs

to fulfil the C-Standard!

 

UEFI shall conform the execution environment described in the C Specification

http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1256.pdf#page=23

and shall not try to create a new restricted “UEFI execution environment”

that currently prohibits some “expressions” (shift << >> , divide / % ) on some “data types” (64bit “long long”)

but maybe in the future will prohibit some more “expressions” (logical AND &&, relational-expression < >) on

still speculative “data types” (e.g. a 128bit “extended long”) or just because a new compiler

(version) with some new optimization(ultra slow)/security(specdown/meltre) capabilities introduces

some new intrinsic functions.

Who knows…

 

In contrast to:

“I think we shouldn't add any intrinsics unless we are absolutely forced

to. I do agree however that, for those intrinsics that we cannot at all

avoid reimplementing, we should at least collect them in a common

library.

(In theory, I can also imagine reimplementing all possible intrinsics

*if* the edk2 coding style spec / requirements are updated in parallel,

permitting all new code to universally rely on the intrinsics, rather

than the BaseLib / BaseMemoryLib functions.)”

https://bugzilla.tianocore.org/show_bug.cgi?id=1516#c2

 

This mindset violates edk2 coding style spec too:

https://edk2-docs.gitbook.io/edk-ii-c-coding-standards-specification/2_guiding_principles

  • Maintainability
  • Extensibility
  • Intellectual manageability
  • Portability
  • Reusability
  • Standard techniques

 

Have fun,

Kilian

 

From: Michael D Kinney
Sent: Friday, January 21, 2022 05:39 PM
To: kraxel@redhat.com; Yao, Jiewen; Sean Brogan; Bret Barkelew; Kinney, Michael D
Cc: devel@edk2.groups.io; Wang, Jian J; Jiang, Guomin; Pawel Polawski; Lu, XiaoyuX
Subject: Re: [edk2-devel] [PATCH 00/24] CryptoPkg/openssl: update openssl submodule to v3.0

 

Comments below.

Mike

> -----Original Message-----
> From: kraxel@redhat.com <kraxel@redhat.com>
> Sent: Friday, January 21, 2022 12:31 AM
> To: Yao, Jiewen <jiewen.yao@intel.com>
> Cc: devel@edk2.groups.io; Kinney, Michael D <michael.d.kinney@intel.com>; Wang, Jian J <jian.j.wang@intel.com>; Jiang, Guomin
> <guomin.jiang@intel.com>; Pawel Polawski <ppolawsk@redhat.com>; Lu, XiaoyuX <xiaoyux.lu@intel.com>
> Subject: Re: [edk2-devel] [PATCH 00/24] CryptoPkg/openssl: update openssl submodule to v3.0
>
> > > No changes in SEC and PEI.
> > [Jiewen] Do you mean the Crypto consumer in PEI has no size difference? Such as
> > https://github.com/tianocore/edk2/tree/master/SecurityPkg/Tcg/Tcg2Pei ,
> > https://github.com/tianocore/edk2/tree/master/SecurityPkg/FvReportPei ,
> > https://github.com/tianocore/edk2/tree/master/SignedCapsulePkg/Universal/RecoveryModuleLoadPei linking
> https://github.com/tianocore/edk2/tree/master/SecurityPkg/Library/FmpAuthenticationLibRsa2048Sha256.
>
> PEI has this (OvmfIa32X64Pkg build):
>
>     7062 TpmMmioSevDecryptPei
>     7830 StatusCodeHandlerPei
>     7902 ReportStatusCodeRouterPei
>     8470 FaultTolerantWritePei
>     9734 SmmAccessPei
>    11206 Tcg2ConfigPei
>    11842 PeiVariable
>    14730 Tcg2PlatformPei
>    17274 TcgPei
>    18438 S3Resume2Pei
>    18682 DxeIpl
>    18938 PcdPeim
>    38014 CpuMpPei
>    39554 PlatformPei
>    45050 PeiCore
>    49274 Tcg2Pei
>
> No size change for Tcg2Pei.
>
> The other modules are not there.  Seems they are related to firmware
> updates.  We don't have that on ovmf as we can simply update the
> firmware image files on the host machine ...
>
> Is there some target I could use to test-build those modules?
>
> > > INFO - OpensslLibCrypto.lib(rsa_lib.obj) : error LNK2001: unresolved external
> > > symbol __allmul
> > > INFO - OpensslLibCrypto.lib(rsa_lib.obj) : error LNK2001: unresolved external
> > > symbol __aulldiv
> > > INFO - OpensslLibCrypto.lib(bio_print.obj) : error LNK2001: unresolved external
> > > symbol __aulldvrm
> > > INFO - OpensslLibCrypto.lib(bio_print.obj) : error LNK2001: unresolved external
> > > symbol __ftol2_sse
> > >
> > > Those symbols look like they reference helper functions to do 64bit math
> > > on 32bit architecture.  Any hints how to fix that?
> > [Jiewen] Please add them to https://github.com/tianocore/edk2/tree/master/CryptoPkg/Library/IntrinsicLib
>
> Any hints where I could get them?  Given this happens on windows builds
> it's probably somewhere in the microsoft standard C library?  Is that
> available as open source somewhere?

Sean and Bret may be able to help with these.

There is also a BZ on this topic.

https://bugzilla.tianocore.org/show_bug.cgi?id=1516

>
> > > (3) Some NOOPT builds are failing due to the size growing ...
> > [Jiewen] Size becomes big challenge...
> > Have you tried to use https://github.com/tianocore/edk2/tree/master/CryptoPkg/Driver solution?
>
> Seems the idea is to have only one openssl copy in the dxe image by
> calling a protocol instead of linking a lib.  Makes sense.
>
> Is this documented somewhere?  Is there some easy way to use that as
> drop-in replacement?  Or do we have to change all crypto users to call
> the driver instead of linking the lib?
>
> take care,
>   Gerd



 



--
Pedro Falcato