gcc : ‘asm’ operand has impossible constraints, on 32bit intel

[vaintroub]

Tring to use TFM library (-DUSE_FAST_MATH) on 32bit intel results into

In file included from wolfssl/wolfssl/wolfcrypt/src/tfm.c:51:
wolfssl/wolfssl/wolfcrypt/src/tfm.c: In function ‘fp_sqr_comba’:
wolfssl/wolfcrypt/src/asm.c:709:1: error: ‘asm’ operand has impossible constraints
asm(
^~~~~~~
wolfssl/wolfssl/wolfcrypt/src/tfm.c:3148:11: note: in expansion of macro ‘SQRADD2’
SQRADD2(*tmpx++, *tmpy--);
^~~~~~~
wolfssl/wolfssl/wolfcrypt/src/asm.c:700:1: error: ‘asm’ operand has impossible constraints
asm(
^~~~~~~
/home/wlad/10.4/extra/wolfssl/wolfssl/wolfcrypt/src/tfm.c:3154:11: note: in expansion of macro ‘SQRADD’
SQRADD(A->dp[ix>>1], A->dp[ix>>1]);

Relevant compile line

cd /home/wlad/10.4/xxx/extra/wolfssl && /usr/bin/cc -DBUILDING_WOLFSSL -DDBUG_TRACE -DHAVE_CONFIG_H -DWOLFSSL_LIB -D_FILE_OFFSET_BITS=64 -I/home/wlad/10.4/extra/wolfssl/wolfssl -I/home/wlad/10.4/wsrep-lib/include -I/home/wlad/10.4/wsrep-lib/wsrep-API/v26 -I/home/wlad/10.4/xxx/include -I/home/wlad/10.4/xxx/extra/wolfssl -I/home/wlad/10.4/extra/wolfssl/wolfssl/wolfssl  -m32 -pie -fPIC -fstack-protector --param=ssp-buffer-size=4 -g1 -DENABLED_DEBUG_SYNC  -DSAFE_MUTEX -DSAFEMALLOC -DTRASH_FREED_MEMORY   -DHAVE_OPENSSL -DHAVE_WOLFSSL  -DWOLFSSL_USER_SETTINGS -fPIC -o CMakeFiles/wolfcrypt.dir/wolfssl/wolfcrypt/src/tfm.c.o   -c /home/wlad/10.4/extra/wolfssl/wolfssl/wolfcrypt/src/tfm.c

user_settings.h

#ifndef WOLFSSL_USER_SETTINGS_H
#define WOLFSSL_USER_SETTINGS_H

#define HAVE_CRL
#define WOLFSSL_MYSQL_COMPATIBLE
#define HAVE_ECC
#define ECC_TIMING_RESISTANT
#define HAVE_HASHDRBG
#define WOLFSSL_AES_DIRECT
#define WOLFSSL_SHA384
#define WOLFSSL_SHA512
#define WOLFSSL_SHA224
#define SESSION_CERT
#define KEEP_OUR_CERT
#define WOLFSSL_STATIC_RSA
#define WOLFSSL_USER_IO
#define WC_RSA_BLINDING
#define HAVE_TLS_EXTENSIONS
#define HAVE_AES_ECB
#define HAVE_AESGCM
#define WOLFSSL_AES_COUNTER
#define NO_WOLFSSL_STUB
#define OPENSSL_ALL
#define WOLFSSL_ALLOW_TLSV10
#define NO_OLD_TIMEVAL_NAME

/* TLSv1.3 definitions (all needed to build) */
#define WOLFSSL_TLS13
#define HAVE_HKDF
#define HAVE_TLS_EXTENSIONS
#define HAVE_SUPPORTED_CURVES
#define HAVE_FFDHE_2048
#define WC_RSA_PSS
/* End of TLSv1.3 defines */

/*
  FP_MAX_BITS is set high solely to satisfy ssl_8k_key.test
  WolfSSL will use more stack space with it, with fastmath
*/
#define FP_MAX_BITS 16384

/* #undef WOLFSSL_AESNI */
#define USE_FAST_MATH
#define TFM_TIMING_RESISTANT
/* #undef HAVE_INTEL_RDSEED */
/* #undef HAVE_INTEL_RDRAND */
/* #undef USE_INTEL_SPEEDUP */
#define USE_FAST_MATH
/* #undef WOLFSSL_X86_64_BUILD */

#endif  /* WOLFSSL_USER_SETTINGS_H */

[kareem-wolfssl]

Hi @vaintroub ,

Thank you for reaching out to us. I will work on reproducing your issue here.

Thanks,
Kareem

[elms]

Hi @vaintroub,

Thanks for the report. I have reproduced and believe I know a solution. I was able to compile by adding -O2. The exact command I used to test was cmake .. -DWOLFSSL_USER_SETTINGS=yes -DCMAKE_C_FLAGS="-m32 -I$PWD -O2" && make.

I'll work on putting a PR up. If you are able, can you confirm that adding -O2 solves the issue?

[vaintroub]

That does not solve it in all circumstances. We use sometimes -fno-omit-frame-pointer, also with optimized build.
I assume you need to add -fomit-frame-pointers, to free a register . Not only that, on downlevel compilers, such as GCC 4.8.5, that we are also using, you'd need to add -fno-PIC, to free yet another register - we compile everything with -fPIC. However -fno-PIC fortunately works on Intel 32bit.

Here is an example where -fomit-frame-pointer -fPIC fails :
https://buildbot.askmonty.org/buildbot/builders/kvm-bintar-centos6-x86/builds/5541/steps/compile/logs/stdio
(look for tfm.c.o on how it is compiled, and for error: on how it fails)

Here is an example where -fomit-frame-pointer -fno-PIC succeeds (same machine)
https://buildbot.askmonty.org/buildbot/builders/kvm-bintar-centos6-x86/builds/5549/steps/compile/logs/stdio

[elms]

@vaintroub Do I understand correctly that you have a workaround for now?

The only solution that I know of that would avoid this completely is using WOLFSSL_NO_ASM. I'll discuss other options with the team.

Out of curiosity have you considered WOLFSSL_SP_MATH?

[stintel]

We're also seeing this in OpenWrt after switching the default GCC version from 8 to 10.

[SparkiDev]

Hi @vaintroub,

Please try pull request #4317

I've made a small change to make the assembly code snippets work.

Thanks.

[stintel]

Please try pull request #4317

This seems to fix the problem for 32bit x86 in OpenWrt, thanks!

[vaintroub]

Unfortunately, it is a not easy to test for me at the moment, but once it is in, but I'm happy to hear about the progress,
Once it is in main tree, I will try to remove workarounds and see what it says.
But if it works for you with -fno-omit-frame-pointer -fPIC, and/or with Debug compilation, it should be fine here too.

[SparkiDev]

Hi @vaintroub

The pull request has been merged to master.

Let me know when you have a chance to test it.

Thanks,
Sean :-)

[elms]

@vaintroub Haven't heard back for a while and worked for others, so I'm marking it closed. Please open a new bug if you have any issues.