Compiler Bugs - simd-everywhere/simde GitHub Wiki
SIMDe has proven to be rather good at finding compiler bugs.
It is generally possible for us to work around these issues, usually by using the preprocessor to choose a different code path. When we find bugs in the current versions of compilers we try to report them so they can be fixed; the GCC and Emscripten developers in particular have been very responsive and helpful.
Bugs We Can't Work Around
Unfortunately, sometimes working around a bug simply isn't feasible. What follows is a list of bugs we can't currently work around.
Clang
- Clang 3.8 hits an internal compiler error when trying to auto-vectorize some functions using AVX-512VL. This isn't usually a problem outside of our test suite since if AVX-512VL is enabled SIMDe will typically use the AVX-512VL intrinsics instead of auto-vectorizing. Either upgrade to clang 3.9, don't use
-DSIMDE_NO_NATIVE(which is really only for testing), or don't auto-vectorize to AVX-512VL (don't use-mavx512vlor add-mno-avx512vl).
GCC
- All GCC 7.x releases hit a floating point exception internal compiler error when compiling many AVX-512 functions if optimization is enabled (
-O1or higher), debugging is enabled (-g), and AVX is not enabled. For details, see issue #278. This only effects applications which use the SIMDe versions of certain AVX-512 instructions. - GCC 4.7 - 5 don't work reliably with AVX-512 support as many functions will generate internal compiler errors. You can still use there compilers with SIMDe, just don't pass -mavx512f (or otherwise enable it, such as with -march=native on an AVX-512 machine) or pass -mno-avx512f.
PGI
- PGI doesn't work in C++ mode. PGI is tracking the issue internally as TPR #28482. For details, see their community forum
Bugs Discovered By SIMDe
As mentioned in the introduction, we seem to find a fair number of compiler bugs. What follows is an incomplete list of some of the issues we've found in compilers which have been fixed, and/or for which there is a work-around in SIMDe.
| Compiler | Bug | Present In | Fixed In | Type | Description |
|---|---|---|---|---|---|
| GCC | 95227 | Spurious warning | vec_extract doesn't mark input as used in C++ mode | ||
| GCC | 10 11 12 | 10.5 12.3 13 | Spurious warning | Unable to ignore -Wattribute-warning in macro | |
| GCC | 95421 | API missing / incorrect | [AArch64] Missing NEON functions documented on ARM's web site | ||
| GCC | 95471 | API missing / incorrect | [ARMv8] vrndvq_f32 defined even when not supported by ARMv8 | ||
| GCC | 95782 | ICE | [ppc64le] ICE in _cpp_pop_context | ||
| GCC | 97016 | API missing / incorrect | _MM_CMPINT_ENUM type is missing | ||
| GCC | 11 | API missing / incorrect | [i386] Missing SIMD functions. | ||
| GCC | 95399 | API missing / incorrect | [ARM] 32/64-bit vcvtnq_* functions are missing | ||
| GCC | 8.5 | API missing / incorrect | [AArch64] vqmovun* return types should be unsigned. | ||
| GCC | 8 / 9 | 8.5/9.4/10 | Incorrect optimization | [8/9 Regression] Inserting into vector with optimization enabled on x86 generates incorrect result. | |
| GCC | 9.3/10.1 | Spurious warning | __builtin_convertvector doesn't mark input as used. | ||
| GCC | 9.4/10.1 | ICE | [10 Regression] Internal compiler error for __builtin_convertvector + statement expr. Latent bug in GCC 9. | ||
| GCC | 8.5/9.4/10.1 | ICE | [AArch64] ICE on right shift of V2DImode by DImode shift. | ||
| GCC | 9.4/10.2 | API missing / incorrect | AVX-512 functions missing when compiled without optimization. | ||
| GCC | 11 | 11.1 | ICE | ICE with omp simd loop + optimization | |
| GCC | 10.2.1 | 10.3 | API missing / incorrect | [x86] _mm256_cmov_si256 XOP function is missing. | |
| GCC | 97248 | ICE | [mips] unrecognizable insn when left shifting uint64 vector by scalar with MSA | ||
| GCC | 11 12 | 11.4 12.1 | API missing / incorrect | [sse2] new _mm_loadu_si16 and _mm_loadu_si32 implemented incorrectly | |
| GCC | 10.2 | 12.1 | ICE | [mips + msa] ICE: maximum number of generated reload insns per insn achieved | |
| GCC | 10.2 | 12.1 | ICE | [mips+msa] ICE when using __builtin_convertvector to convert from u8x8 to u8x16 | |
| GCC | 10.2 | 12.1 | ICE | [mips+msa] ICE when comparing 64 bit vectors | |
| GCC | 100927 | 10.2 | Incorrect optimization | [sse2] floating point to integer conversion functions incorrect results w/ NaN constants + optimization | |
| GCC | 9, 10, 11, 12 | 9.5, 10.4, 11.4, 12.1 | API missing / incorrect | [x86] missing AVX-512F scalef functions when optimization is disabled | |
| GCC | 10.2.1 | 12.1 | Implementation incorrect | [POWER] vec_min / vec_max handles NaN incorrectly when evaluated at compile time | |
| GCC | 9, 10, 11 | 12.0 | Implementation incorrect | [POWER] vec_cpsgn parameter order was reversed | |
| GCC | 14.0 | 14.1 | miscompilation | [s390x] Do not emulate vectors containing floats | |
| GCC | 113065 | 10.2 | ICE | [ARM32] shift of 128-bit NEON types | |
| GCC | 13.1.0, 13.3.0, 14.1.0, 14.2.0, 9.1.0 | 12.5.0, 13.4.0, 14.3.0, 15.0, 8.5.0 | ICE | [i586] invalid 'PHI' argument | |
| Clang | 15.0 | Incorrect optimization | [AArch64] Incorrect result for vector conversion with -O2 | ||
| Clang | 11 | API missing / incorrect | Many AVX-512 functions take an int instead of unsigned int. | ||
| Clang | BZ45959 GH45304 | Spurious warning | SIMD & reduction on signed types emits sign-conversion diagnostic | ||
| Clang | 12 | API missing / incorrect | [ppc] vec_sel variants missing | ||
| Clang | 46844 GH46188 | 12 (AArch64-only) | Incorrect optimization | [AArch64] incorrect results from vcvt* functions with negative inputs when optimization is enabled. | |
| Clang | 12 | API missing / incorrect | [AArch64] vqmovun* return types should be unsigned. | ||
| Clang | 5 | API missing / incorrect | _mm_set_pd1 missing from emmintrin.h | ||
| Clang | Spurious warning | _mm_extract_pi16 and _mm_insert_pi16 warn with -Wvector-conversion. Fixed in clang 11.0 | |||
| Clang | BZ48257 GH47601 | Spurious warning | vget_lane_p64 triggers -Wvector-conversion diagnostic | ||
| Clang | BZ48673 GH48017 | API missing / incorrect | [x86] _mm_frcz_ss and _mm_frcz_sd should take two parameters | ||
| Clang | 12 | API missing / incorrect | NEON scalar comparison functions should return unsigned values | ||
| Clang | 7+ | 13 | ICE | clang segfault at -O2 in C mode | |
| Clang | 11.0 | 13.0 | ICE | Compiler error when converting from vector of 2x32 to 2x64-bit int on POWER7 | |
| Clang | 11.0 | 13.0 | ICE | [power7] error in backend: Cannot select v16i8 = PPCISD::SCALAR_TO_VECTOR_PERMUTED | |
| Clang | 11.0 | trunk | Perf | slow code for absolute value of int8 x 16 vector on POWER9 at -O3 | |
| Clang | 14.0 | API missing / incorrect | [POWER] vec_bperm missing documented signatures | ||
| Clang | 14-dev | 14-dev | ICE | Regression [VectorCombine] ScalarizationResult destructor assertion due to SafeWithFreeze and scalarizeLoadExtract | |
| Clang | 71362 | Incorrect Implementation | NEON intrinsic compilation error occurs when using -fno-lax-vector-conversions | ||
| Clang | 71365 | Incorrect Implementation | Wrong return type in some ARM NEON intrinsics | ||
| Clang | 15 | trunk | API missing / incorrect | Wrong return type of NEON intrinsic vqrshrunh_n_s16 in arm_neon.h | |
| Clang | trunk | Incorrect Implementation | Wrong result of NEON intrinsic vld2q_dup_p16 with -march=armv7-a | ||
| Emscripten | 1.39.7 | 1.39.11 | ICE | LLVM assertion failure when casting a SIMD type with -O3 | |
| Emscripten | ICE | Crash in WebAssembly Register Stackify with tot | |||
| Emscripten | 1.39.16-dev | 1.39.17 | ICE | "Not a vector MVT!" clang error with -s SIMD=1 on tot | |
| Emscripten | 1.39.17 | ICE | emcc crashes with SIMD without optimization | ||
| Emscripten | 1.39.9 | 1.39.10 | API missing / incorrect | emscripten errors in wasm_simd128.h | |
| Emscripten | 2.0.27 | ICE | UNREACHABLE executed at …/binaryen/src/wasm-interpreter.h:503! | ||
| Emscripten | 3.1.34 | 3.1.43 | API incorrect | Compiler advertises __builtin_roundeven{f,} but it is not implemented |
|
| ICC | ??? | Incorrect optimization | ICC generates incorrect code for signed absolute difference | ||
| ICC | ??? | False advertisement of feature | __builtin_expect_with_probability unsupported but __has_builtin claims otherwise | ||
| ICC | ??? | False advertisement of feature | fallthrough attribute unsupported, contrary to icc's claims | ||
| ICC | ??? | ICE | Internal error: 04010002_1671 | ||
| NVC | 30104 | 21.3 | ICE | ICE from % operator on vector extensions | |
| NVC | 30107 | 21.3 | ICE | Using _mm_cvtpd_epi32 results in compiler error | |
| NVC | 30106 | 21.3 | API missing/incorrect | Undefined reference to `__builtin_ia32_palignr256’ when calling _mm256_alignr_epi8 |
Note: some of the version number information may be incorrect or missing. Sorry, it's a recent addition to the table.