Result before this patch:
Checking enum `cpu_feature_t': 113 elements; max size (CPU_FLAGS_MAX=128)... OK
Checking enum `cpu_hint_t': 1 elements; max size (CPU_HINTS_MAX=16)... OK
Checking enum `cpu_sgx_feature_t': 2 elements; max size (SGX_FLAGS_MAX=14)... OK
Finding features:
..Mismatch - cpuid_main.c:688 - `AVX512VNNI' vs `avx512_vnni'
..Mismatch - cpuid_main.c:689 - `AVX512VBMI' vs `avx512_vbmi'
..Mismatch - cpuid_main.c:690 - `AVX512VBMI2' vs `avx512_vbmi2'
cpuid_main.c: 113 features described
Found 113 total features and 113 named features
Checking whether all features have detection code... FAILED:
..No detection code for CPU_FEATURE_SSE5
Reported via http://libcpuid.sourceforge.net/bugreport.php
The test in particular has no brand string, which was causing the
misdetection (as is the case with a lot of other models, libcpuid
relies on accurate brand string being programmed by the BIOS in
order to do the detection).
The actual CPU was a Pentium-III based Celeron (SL54Q), but it
was detected as "Pentium III (Coppermine)".
A bit of historical trivia: for the related Tualatin models, if
the BIOS doesn't enter a brand string, there might be NO WAY to
tell a regular P-3 and a P-3 Celeron apart: P-3s have variants
with 256KiB and 512KiB L2 cache, while the Celerons are 256 KiB, so
a 256KiB regular P3 is no different than its corresponding Celeron.
Only the FSB is different, but there's no way to detect this via
CPUID.
For the Coppermines its an easier case: Celerons are always 128KiB,
and Pentia are 256KiB, so I've added this distinction in the tables.
Add support for detecting Coffee Lake i[357]s. Test included.
Thanks to @X0rg for reporting and to @exdeus for the raw cpu data.
The detection here relies on Kaby and Coffee lakes having different
number of cores for the equivalent brand:
i3: 2 cores in Kaby, 4 in Coffee
i5 and i7: 4 cores in Kaby, 6 in Coffee.
Add support for detecting RDSEED and ADX instructions.
Use RDSEED instead of RTM to ascertain that the CPU is
Broadwell or later in recog_intel.c. This fixes
detection discrepancies on Linux, where RTM is not
made available (I guess there's no kernel support for it).
The two new flags are also now detected in the Broadwell
and Skylake tests. Update them as well.
Improved parsing of brand strings for both Pineview and Cedarview
Atoms.
Add tests for both the Pineview Atom, and also Broadwell-E
(forgotten in a previous commit).
The test is a snapshot of a Haswell i7 (a.k.a. "Crystalwell") core.
This is the only test in the test DB right now which has lines for
L4 cache size, associativity and cacheline size different than "-1".
Also update create_test.py to accommodate for the new fields.
