crypto/x86_64cpuid.pl: detect if kernel preserves %zmm registers.
authorAndy Polyakov <appro@openssl.org>
Fri, 27 Jan 2017 18:03:37 +0000 (19:03 +0100)
committerAndy Polyakov <appro@openssl.org>
Fri, 3 Feb 2017 11:21:50 +0000 (12:21 +0100)
Reviewed-by: Rich Salz <rsalz@openssl.org>
crypto/x86_64cpuid.pl

index 6cb152148b5b6eae2d61da23e18dd442e3edfff5..3082253aea947082d12b456e9399d20ea3c253aa 100644 (file)
@@ -175,13 +175,21 @@ OPENSSL_ia32_cpuid:
        jnc     .Lclear_avx
        xor     %ecx,%ecx               # XCR0
        .byte   0x0f,0x01,0xd0          # xgetbv
+       and     \$0xe6,%eax             # isolate XMM, YMM and ZMM state support
+       cmp     \$0xe6,%eax
+       je      .Ldone
+       andl    \$0xfffeffff,8(%rdi)    # clear AVX512F, ~(1<<16)
+                                       # note that we don't touch other AVX512
+                                       # extensions, because they can be used
+                                       # with YMM (without opmasking though)
        and     \$6,%eax                # isolate XMM and YMM state support
        cmp     \$6,%eax
        je      .Ldone
 .Lclear_avx:
        mov     \$0xefffe7ff,%eax       # ~(1<<28|1<<12|1<<11)
        and     %eax,%r9d               # clear AVX, FMA and AMD XOP bits
-       andl    \$0xffffffdf,8(%rdi)    # cleax AVX2, ~(1<<5)
+       mov     \$0x3fdeffdf,%eax       # ~(1<<31|1<<30|1<<21|1<<16|1<<5)
+       and     %eax,8(%rdi)            # cleax AVX2 and AVX512* bits
 .Ldone:
        shl     \$32,%r9
        mov     %r10d,%eax