File 0002-Fix-detection-of-SSE-AVX-AVX512-when-they-are-explic.patch of Package xsimd
From c2974c874e14557490eab76d2eebf9f8b9eb88f1 Mon Sep 17 00:00:00 2001
From: Dmitry Kazakov <dimula73@gmail.com>
Date: Tue, 28 May 2024 22:21:08 +0200
Subject: [PATCH 2/2] Fix detection of SSE/AVX/AVX512 when they are explicitly
disabled by OS
Some CPU vulnerability mitigations may disable AVX functionality
on the hardware level via the XCR0 register. We should check that
manually to verify that OS actually allows us to use this feature.
See https://bugs.kde.org/show_bug.cgi?id=484622
Fix #1025
---
include/xsimd/config/xsimd_cpuid.hpp | 91 ++++++++++++++++++++++------
1 file changed, 72 insertions(+), 19 deletions(-)
diff --git a/include/xsimd/config/xsimd_cpuid.hpp b/include/xsimd/config/xsimd_cpuid.hpp
index 30a9da2..8021fce 100644
--- a/include/xsimd/config/xsimd_cpuid.hpp
+++ b/include/xsimd/config/xsimd_cpuid.hpp
@@ -122,6 +122,35 @@ namespace xsimd
#endif
#elif defined(__x86_64__) || defined(__i386__) || defined(_M_AMD64) || defined(_M_IX86)
+
+ auto get_xcr0_low = []() noexcept
+ {
+ uint32_t xcr0;
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+
+ xcr0 = (uint32_t)_xgetbv(0);
+
+#elif defined(__GNUC__)
+
+ __asm__(
+ "xorl %%ecx, %%ecx\n"
+ "xgetbv\n"
+ : "=a"(xcr0)
+ :
+#if defined(__i386__)
+ : "ecx", "edx"
+#else
+ : "rcx", "rdx"
+#endif
+ );
+
+#else /* _MSC_VER < 1400 */
+#error "_MSC_VER < 1400 is not supported"
+#endif /* _MSC_VER && _MSC_VER >= 1400 */
+ return xcr0;
+ };
+
auto get_cpuid = [](int reg[4], int level, int count = 0) noexcept
{
@@ -156,19 +185,43 @@ namespace xsimd
get_cpuid(regs1, 0x1);
- sse2 = regs1[3] >> 26 & 1;
- sse3 = regs1[2] >> 0 & 1;
- ssse3 = regs1[2] >> 9 & 1;
- sse4_1 = regs1[2] >> 19 & 1;
- sse4_2 = regs1[2] >> 20 & 1;
- fma3_sse42 = regs1[2] >> 12 & 1;
+ // OS can explicitly disable the usage of SSE/AVX extensions
+ // by setting an appropriate flag in CR0 register
+ //
+ // https://docs.kernel.org/admin-guide/hw-vuln/gather_data_sampling.html
+
+ unsigned sse_state_os_enabled = 1;
+ unsigned avx_state_os_enabled = 1;
+ unsigned avx512_state_os_enabled = 1;
+
+ // OSXSAVE: A value of 1 indicates that the OS has set CR4.OSXSAVE[bit
+ // 18] to enable XSETBV/XGETBV instructions to access XCR0 and
+ // to support processor extended state management using
+ // XSAVE/XRSTOR.
+ bool osxsave = regs1[2] >> 27 & 1;
+ if (osxsave)
+ {
+
+ uint32_t xcr0 = get_xcr0_low();
+
+ sse_state_os_enabled = xcr0 >> 1 & 1;
+ avx_state_os_enabled = xcr0 >> 2 & sse_state_os_enabled;
+ avx512_state_os_enabled = xcr0 >> 6 & avx_state_os_enabled;
+ }
+
+ sse2 = regs1[3] >> 26 & sse_state_os_enabled;
+ sse3 = regs1[2] >> 0 & sse_state_os_enabled;
+ ssse3 = regs1[2] >> 9 & sse_state_os_enabled;
+ sse4_1 = regs1[2] >> 19 & sse_state_os_enabled;
+ sse4_2 = regs1[2] >> 20 & sse_state_os_enabled;
+ fma3_sse42 = regs1[2] >> 12 & sse_state_os_enabled;
- avx = regs1[2] >> 28 & 1;
+ avx = regs1[2] >> 28 & avx_state_os_enabled;
fma3_avx = avx && fma3_sse42;
int regs8[4];
get_cpuid(regs8, 0x80000001);
- fma4 = regs8[2] >> 16 & 1;
+ fma4 = regs8[2] >> 16 & avx_state_os_enabled;
// sse4a = regs[2] >> 6 & 1;
@@ -176,23 +229,23 @@ namespace xsimd
int regs7[4];
get_cpuid(regs7, 0x7);
- avx2 = regs7[1] >> 5 & 1;
+ avx2 = regs7[1] >> 5 & avx_state_os_enabled;
int regs7a[4];
get_cpuid(regs7a, 0x7, 0x1);
- avxvnni = regs7a[0] >> 4 & 1;
+ avxvnni = regs7a[0] >> 4 & avx_state_os_enabled;
fma3_avx2 = avx2 && fma3_sse42;
- avx512f = regs7[1] >> 16 & 1;
- avx512cd = regs7[1] >> 28 & 1;
- avx512dq = regs7[1] >> 17 & 1;
- avx512bw = regs7[1] >> 30 & 1;
- avx512er = regs7[1] >> 27 & 1;
- avx512pf = regs7[1] >> 26 & 1;
- avx512ifma = regs7[1] >> 21 & 1;
- avx512vbmi = regs7[2] >> 1 & 1;
- avx512vnni_bw = regs7[2] >> 11 & 1;
+ avx512f = regs7[1] >> 16 & avx512_state_os_enabled;
+ avx512cd = regs7[1] >> 28 & avx512_state_os_enabled;
+ avx512dq = regs7[1] >> 17 & avx512_state_os_enabled;
+ avx512bw = regs7[1] >> 30 & avx512_state_os_enabled;
+ avx512er = regs7[1] >> 27 & avx512_state_os_enabled;
+ avx512pf = regs7[1] >> 26 & avx512_state_os_enabled;
+ avx512ifma = regs7[1] >> 21 & avx512_state_os_enabled;
+ avx512vbmi = regs7[2] >> 1 & avx512_state_os_enabled;
+ avx512vnni_bw = regs7[2] >> 11 & avx512_state_os_enabled;
avx512vnni_vbmi = avx512vbmi && avx512vnni_bw;
#endif
}
--
2.45.2
