build: Configuring with `-mno-sse4.1` does not fail the sse4.1 instrinsics check

dergoegge commented at 4:59 pm on November 13, 2023: member

I would have expected that CXXFLAGS="-mno-sse4.1" ./configure would make the sse4.1 intrinsics check fail but it doesn’t:

0configure:23377: checking for SSE4.1 intrinsics
1configure:23396: clang++ -std=c++17 -c -msse4.1 -mno-sse4.1  conftest.cpp >&5
2configure:23396: $? = 0
3configure:23398: result: yes

Not sure what is going on here, seems like a bug.

dergoegge renamed this:
~~build: Compiling with `-mno-sse4.1` does not fail the sse4.1 instrinsics check~~
build: Configuring with `-mno-sse4.1` does not fail the sse4.1 instrinsics check
on Nov 13, 2023

fanquake commented at 3:52 pm on November 14, 2023: member

Note that this ultimately results in compile failure, but because we try and compile the sha256_x86_shani.cpp code, but fail because it "needs isa option -msse4.1":

 0In file included from /usr/lib/gcc/x86_64-linux-gnu/13/include/immintrin.h:39,
 1                 from crypto/sha256_x86_shani.cpp:12:
 2In function 'void {anonymous}::Shuffle(__m128i&, __m128i&)',
 3    inlined from 'void sha256_x86_shani::Transform(uint32_t*, const unsigned char*, size_t)' at crypto/sha256_x86_shani.cpp:87:12:
 4crypto/sha256_x86_shani.cpp:57:10: error: '__builtin_ia32_pblendw128' needs isa option -msse4.1
 5   57 |     s1 = _mm_blend_epi16(t2, t1, 0xF0);
 6      |          ^~~~~~~~~~~~~~~
 7In function 'void {anonymous}::Unshuffle(__m128i&, __m128i&)',
 8    inlined from 'void sha256_x86_shani::Transform(uint32_t*, const unsigned char*, size_t)' at crypto/sha256_x86_shani.cpp:138:14:
 9crypto/sha256_x86_shani.cpp:64:10: error: '__builtin_ia32_pblendw128' needs isa option -msse4.1
10   64 |     s0 = _mm_blend_epi16(t1, t2, 0xF0);
11      |          ^~~~~~~~~~~~~~~
12In function 'void {anonymous}::Unshuffle(__m128i&, __m128i&)',
13    inlined from 'void sha256d64_x86_shani::Transform_2way(unsigned char*, const unsigned char*)' at crypto/sha256_x86_shani.cpp:270:14:
14crypto/sha256_x86_shani.cpp:64:10: error: '__builtin_ia32_pblendw128' needs isa option -msse4.1
15   64 |     s0 = _mm_blend_epi16(t1, t2, 0xF0);
16      |          ^~~~~~~~~~~~~~~
17In function 'void {anonymous}::Unshuffle(__m128i&, __m128i&)',
18    inlined from 'void sha256d64_x86_shani::Transform_2way(unsigned char*, const unsigned char*)' at crypto/sha256_x86_shani.cpp:271:14:
19crypto/sha256_x86_shani.cpp:64:10: error: '__builtin_ia32_pblendw128' needs isa option -msse4.1
20   64 |     s0 = _mm_blend_epi16(t1, t2, 0xF0);
21      |          ^~~~~~~~~~~~~~~
22In function 'void {anonymous}::Unshuffle(__m128i&, __m128i&)',
23    inlined from 'void sha256d64_x86_shani::Transform_2way(unsigned char*, const unsigned char*)' at crypto/sha256_x86_shani.cpp:348:14:
24crypto/sha256_x86_shani.cpp:64:10: error: '__builtin_ia32_pblendw128' needs isa option -msse4.1
25   64 |     s0 = _mm_blend_epi16(t1, t2, 0xF0);
26      |          ^~~~~~~~~~~~~~~
27In function 'void {anonymous}::Unshuffle(__m128i&, __m128i&)',
28    inlined from 'void sha256d64_x86_shani::Transform_2way(unsigned char*, const unsigned char*)' at crypto/sha256_x86_shani.cpp:349:14:
29crypto/sha256_x86_shani.cpp:64:10: error: '__builtin_ia32_pblendw128' needs isa option -msse4.1
30   64 |     s0 = _mm_blend_epi16(t1, t2, 0xF0);
31      |          ^~~~~~~~~~~~~~~
32make[2]: *** [Makefile:14527: crypto/libbitcoin_crypto_x86_shani_la-sha256_x86_shani.lo] Error 1

and also because we then try and compile the sse4.1 code anyways:

 0In file included from /usr/lib/gcc/x86_64-linux-gnu/13/include/immintrin.h:37,
 1                 from crypto/sha256_sse41.cpp:8:
 2/usr/lib/gcc/x86_64-linux-gnu/13/include/tmmintrin.h: In function '__m128i sha256d64_sse41::{anonymous}::Read4(const unsigned char*, int)':
 3/usr/lib/gcc/x86_64-linux-gnu/13/include/tmmintrin.h:136:1: error: inlining failed in call to 'always_inline' '__m128i _mm_shuffle_epi8(__m128i, __m128i)': target specific option mismatch
 4  136 | _mm_shuffle_epi8 (__m128i __X, __m128i __Y)
 5      | ^~~~~~~~~~~~~~~~
 6crypto/sha256_sse41.cpp:55:28: note: called from here
 7   55 |     return _mm_shuffle_epi8(ret, _mm_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL));
 8      |            ~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 9/usr/lib/gcc/x86_64-linux-gnu/13/include/tmmintrin.h:136:1: error: inlining failed in call to 'always_inline' '__m128i _mm_shuffle_epi8(__m128i, __m128i)': target specific option mismatch
10  136 | _mm_shuffle_epi8 (__m128i __X, __m128i __Y)
11      | ^~~~~~~~~~~~~~~~
12crypto/sha256_sse41.cpp:55:28: note: called from here
13   55 |     return _mm_shuffle_epi8(ret, _mm_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL));
14      |            ~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
15make[2]: *** [Makefile:14520: crypto/libbitcoin_crypto_sse41_la-sha256_sse41.lo] Error 1

Both of these can / should be fixed up.

hebasto commented at 12:06 pm on November 16, 2023: member

FWIW, adding -O1 to the test makes it fail with CXXFLAGS="-mno-sse4.1".

luke-jr commented at 1:13 am on December 5, 2023: member

This is fixed by #13789 which was inexplicably closed for no reason

maflcko commented at 7:49 am on December 5, 2023: member

This is fixed by #13789 which was inexplicably closed for no reason

It was closed due to #13758 (comment)

willcl-ark added the label Build system on Jan 24, 2024

willcl-ark added the label Bug on Jan 24, 2024

fanquake closed this on Jul 17, 2024

fanquake referenced this in commit 3679fa167f on Jul 17, 2024

build: Configuring with -mno-sse4.1 does not fail the sse4.1 instrinsics check #28864

build: Configuring with `-mno-sse4.1` does not fail the sse4.1 instrinsics check #28864