ppc64 / sha2-le

SHA2 algorithm implementation and optimization for ppc64le

This project is about SHA256 and SHA512 implementations and optimizations

The focus of this project is creating an optimized code for IBM POWER architecture, especifically using ABI v2 (Little Endian), aka ppc64le.

This is a generic project and self-contained now. We are going to start backporting it to different Languages and JIT projects. Our initial target at this moment is the following list:

• OpenJDK
• nettle
• Cryptopp
• ...

sha2-le's Issues

Undefined behavior in tests.c

When compiling with sanitizer flags -fsanitize=address,undefined the following errors are displayed:

tests.c:28:5: runtime error: left shift of 134217728 by 30 places cannot be represented in type 'int'
tests.c:28:5: runtime error: left shift of 134217728 by 19 places cannot be represented in type 'int'
tests.c:28:5: runtime error: left shift of 134217728 by 10 places cannot be represented in type 'int'
tests.c:29:8: runtime error: left shift of 134217728 by 26 places cannot be represented in type 'int'
tests.c:29:8: runtime error: left shift of 134217728 by 21 places cannot be represented in type 'int'
tests.c:29:8: runtime error: left shift of 134217728 by 7 places cannot be represented in type 'int'
./bin/test256_ll_intrinsics
./bin/test256_ll_asm
./bin/test512
tests.c:63:5: runtime error: shift exponent 63 is too large for 32-bit type 'int'
tests.c:63:5: runtime error: shift exponent 56 is too large for 32-bit type 'int'
tests.c:64:8: runtime error: left shift of 4611686018427387904 by 45 places cannot be represented in type 'long int'
tests.c:64:8: runtime error: left shift of 4611686018427387904 by 3 places cannot be represented in type 'long int'
tests.c:65:5: runtime error: left shift of 4611686018427387904 by 36 places cannot be represented in type 'long int'
tests.c:65:5: runtime error: left shift of 4611686018427387904 by 30 places cannot be represented in type 'long int'
tests.c:65:5: runtime error: left shift of 4611686018427387904 by 25 places cannot be represented in type 'long int'
tests.c:66:5: runtime error: left shift of 4611686018427387904 by 50 places cannot be represented in type 'long int'
tests.c:66:5: runtime error: left shift of 4611686018427387904 by 46 places cannot be represented in type 'long int'
tests.c:66:5: runtime error: left shift of 4611686018427387904 by 23 places cannot be represented in type 'long int'

Wrong sha256 for ubuntu image

The file http://cdimage.ubuntu.com/releases/16.10/release/ubuntu-16.10-server-ppc64el.iso should have SHA256 sum value of d14bdb413ea6cdc8d9354fcbc37a834b7de0c23f992deb0c6764d0fd5d65408e according to http://cdimage.ubuntu.com/releases/16.10/release/SHA256SUMS and the core-utils sha256sum command.

Instead, sha256 returns aef33b9074cce6a7ffa4f9450c84b21a469bc987178edebc983c3a875eb89134

Investigate efficiency of gcc builtin sigma functions and analyse generated code

Functions like __builtin_crypto_vshasigmaw can be used directly by gdb. By looking at the output with -O3 it may give us better insight of how to best develop this functionality for ppc64le on other projects.

Solve -O3 issues on vectorized variant

Currently (3f25bf2) when compiling with ´-O3thesha256_hwoutputs wrong results for inputs bigger than a block andsha512_hw` is raising segmentation faults.

The issue is due to the inline assembly found inside of calc_compression function.

Allow big inputs

Big inputs (bigger than 3mb) are not allowed by the OS (probably because we are using static vector allocation).

sha256_compress_ppc.o does not export any symbols

I think there's something missing from the assembly source file generated using the toolchain. It appears the symbol is present but it is not exported.

$ m4 common.m4 sha256_compress_ppc.m4 > sha256_compress.s
$ as -mpower8 sha256_compress.s -o sha256_compress.o
$ g++ test.cxx -o test.exe sha256_compress.o
/home/noloader/tmp/cc6pWEuv.o: In function `main':
test.cxx:(.text+0x88): undefined reference to `sha256_compress_ppc(unsigned int*, unsigned char const*, unsigned int const*)'
collect2: error: ld returned 1 exit status

And then:

$ nm -D sha256_compress.o
nm: sha256_compress.o: no symbols
$ objdump --disassemble sha256_compress.o

sha256_compress.o:     file format elf64-powerpcle

Disassembly of section .text:

0000000000000000 <sha256_compress_ppc>:
   0:   50 ff 00 38     li      r0,-176
   4:   ce 01 a1 7f     stvx    v29,r1,r0
   8:   60 ff 00 38     li      r0,-160
   c:   ce 01 81 7f     stvx    v28,r1,r0
   ...
 744:   d0 ff 00 38     li      r0,-48
 748:   ce 00 a1 7e     lvx     v21,r1,r0
 74c:   e0 ff 00 38     li      r0,-32
 750:   ce 00 81 7e     lvx     v20,r1,r0
 754:   20 00 80 4e     blr