X-Git-Url: https://git.distorted.org.uk/~mdw/catacomb/blobdiff_plain/a0e9bb8a70bf2ac9ddb7d1f638ce6af2c80d13e7..4bf3072e34661918a8e6c5e95c3e2ab05c76089b:/base/dispatch.c

diff --git a/base/dispatch.c b/base/dispatch.c
index 63649748..9ba6a7cd 100644
--- a/base/dispatch.c
+++ b/base/dispatch.c
@@ -47,6 +47,7 @@
 #  define CPUID1D_SSE2 (1u << 26)
 #  define CPUID1D_FXSR (1u << 24)
 #  define CPUID1C_AESNI (1u << 25)
+#  define CPUID1C_AVX (1u << 28)
 #  define CPUID1C_RDRAND (1u << 30)
 
 struct cpuid { unsigned a, b, c, d; };
@@ -72,8 +73,8 @@ static __inline__ unsigned setflags(unsigned f)
 {
   unsigned ff;
   __asm__ ("pushf; pushl %1; popf; pushf; popl %0; popf"
-	   : "=g" (ff)
-	   : "g" (f));
+	   : "=r" (ff)
+	   : "r" (f));
   return (ff);
 }
 #  else
@@ -83,8 +84,8 @@ static __inline__ unsigned long long setflags(unsigned long f)
 {
   unsigned long ff;
   __asm__ ("pushf; pushq %1; popf; pushf; popq %0; popf"
-	   : "=g" (ff)
-	   : "g" (f));
+	   : "=r" (ff)
+	   : "r" (f));
   return (ff);
 }
 #  endif
@@ -229,6 +230,16 @@ struct auxentry { unsigned long type; union auxval value; };
 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 #endif
 
+#if defined(AT_HWCAP) && CPUFAM_ARM64
+#  define WANT_ANY 1
+#  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
+#endif
+
+#if defined(AT_HWCAP2) && CPUFAM_ARMEL
+#  define WANT_ANY 1
+#  define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
+#endif
+
 /* If we couldn't find any interesting entries then we can switch all of this
  * machinery off.  Also do that if we have no means for atomic updates.
  */
@@ -246,6 +257,9 @@ static unsigned hwcaps = 0;
 #ifndef WANT_AT_HWCAP
 #  define WANT_AT_HWCAP(_)
 #endif
+#ifndef WANT_AT_HWCAP2
+#  define WANT_AT_HWCAP2(_)
+#endif
 
 /* For each CPU family, define two lists.
  *
@@ -261,12 +275,20 @@ static unsigned hwcaps = 0;
  */
 #if CPUFAM_ARMEL
 #  define WANTAUX(_)							\
-	WANT_AT_HWCAP(_)
+	WANT_AT_HWCAP(_)						\
+	WANT_AT_HWCAP2(_)
 #  define CAPMAP(_)							\
 	_(ARM_VFP, "arm:vfp")						\
 	_(ARM_NEON, "arm:neon")						\
 	_(ARM_V4, "arm:v4")						\
-	_(ARM_D32, "arm:d32")
+	_(ARM_D32, "arm:d32")						\
+	_(ARM_AES, "arm:aes")
+#endif
+#if CPUFAM_ARM64
+#  define WANTAUX(_)							\
+	WANT_AT_HWCAP(_)
+#  define CAPMAP(_)							\
+	_(ARM_AES, "arm:aes")
 #endif
 
 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
@@ -319,8 +341,8 @@ static void probe_hwcaps(void)
   /* Shiny new libc lets us request individual entry types.  This is almost
    * too easy.
    */
-#  define CAP__GET(type, slot, ubranch)					\
-	probed.slot.ubranch = (AUXUTYPE_##ubranch)getauxval(type);
+#  define CAP__GET(type, ubranch, slot)					\
+	probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
   WANTAUX(CAP__GET)
 #else
   /* Otherwise we're a bit stuck, really.  Modern Linux kernels make a copy
@@ -377,6 +399,12 @@ static void probe_hwcaps(void)
   if (probed.hwcap & HWCAP_NEON) hw |= HF_ARM_NEON;
   if (probed.hwcap & HWCAP_VFPD32) hw |= HF_ARM_D32;
   if (probed.hwcap & HWCAP_VFPv4) hw |= HF_ARM_V4;
+#  ifdef HWCAP2_AES
+  if (probed.hwcap2 & HWCAP2_AES) hw |= HF_ARM_AES;
+#  endif
+#endif
+#if CPUFAM_ARM64
+  if (probed.hwcap & HWCAP_AES) hw |= HF_ARM_AES;
 #endif
 
   /* Store the bitmask of features we probed for everyone to see. */
@@ -518,6 +546,9 @@ int cpu_feature_p(int feat)
 		 cpuid_features_p(CPUID1D_SSE2, CPUID1C_AESNI));
     CASE_CPUFEAT(X86_RDRAND, "x86:rdrand",
 		 cpuid_features_p(0, CPUID1C_RDRAND));
+    CASE_CPUFEAT(X86_AVX, "x86:avx",
+		 xmm_registers_available_p() &&
+		 cpuid_features_p(0, CPUID1C_AVX));
 #endif
 #ifdef CAPMAP
 #  define FEATP__CASE(feat, tok)					\