[PATCH for-8.0 16/29] accel/tcg: Use have_atomic16 in ldst_atomicity.c.inc

[Richard Henderson <richard.henderson@linaro.org>]

Hosts using Intel and AMD AVX cpus are quite common.
Add fast paths through ldst_atomicity using this.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 accel/tcg/ldst_atomicity.c.inc | 76 +++++++++++++++++++++++++++-------
 1 file changed, 60 insertions(+), 16 deletions(-)

diff --git a/accel/tcg/ldst_atomicity.c.inc b/accel/tcg/ldst_atomicity.c.inc
index 68edab4398..d2a3783193 100644
--- a/accel/tcg/ldst_atomicity.c.inc
+++ b/accel/tcg/ldst_atomicity.c.inc
@@ -35,6 +35,14 @@
 
 #if defined(CONFIG_ATOMIC128)
 # define HAVE_al16_fast    true
+#elif defined(CONFIG_TCG_INTERPRETER)
+/*
+ * FIXME: host specific detection this is in tcg/$host/,
+ * but we're using tcg/tci/ instead.
+ */
+# define HAVE_al16_fast    false
+#elif defined(__x86_64__)
+# define HAVE_al16_fast    likely(have_atomic16)
 #else
 # define HAVE_al16_fast    false
 #endif
@@ -160,6 +168,12 @@ load_atomic16(void *pv)
 
     r.u = qatomic_read__nocheck(p);
     return r.s;
+#elif defined(__x86_64__)
+    Int128Alias r;
+
+    /* Via HAVE_al16_fast, have_atomic16 is true. */
+    asm("vmovdqa %1, %0" : "=x" (r.u) : "m" (*(Int128 *)pv));
+    return r.s;
 #else
     qemu_build_not_reached();
 #endif
@@ -379,6 +393,24 @@ load_atom_extract_al16_or_al8(void *pv, int s)
         r = qatomic_read__nocheck(p16);
     }
     return r >> shr;
+#elif defined(__x86_64__)
+    uintptr_t pi = (uintptr_t)pv;
+    int shr = (pi & 7) * 8;
+    uint64_t a, b;
+
+    /* Via HAVE_al16_fast, have_atomic16 is true. */
+    pv = (void *)(pi & ~7);
+    if (pi & 8) {
+        uint64_t *p8 = __builtin_assume_aligned(pv, 16, 8);
+        a = qatomic_read__nocheck(p8);
+        b = qatomic_read__nocheck(p8 + 1);
+    } else {
+        asm("vmovdqa %2, %0\n\tvpextrq $1, %0, %1"
+            : "=x"(a), "=r"(b) : "m" (*(__uint128_t *)pv));
+    }
+    asm("shrd %b2, %1, %0" : "+r"(a) : "r"(b), "c"(shr));
+
+    return a;
 #else
     qemu_build_not_reached();
 #endif
@@ -695,23 +727,35 @@ static inline void ATTRIBUTE_ATOMIC128_OPT
 store_atomic16(void *pv, Int128Alias val)
 {
 #if defined(CONFIG_ATOMIC128)
-    __uint128_t *pu = __builtin_assume_aligned(pv, 16);
-    qatomic_set__nocheck(pu, val.u);
-#elif defined(CONFIG_CMPXCHG128)
-    __uint128_t *pu = __builtin_assume_aligned(pv, 16);
-    __uint128_t o;
-
-    /*
-     * Without CONFIG_ATOMIC128, __atomic_compare_exchange_n will always
-     * defer to libatomic, so we must use __sync_val_compare_and_swap_16
-     * and accept the sequential consistency that comes with it.
-     */
-    do {
-        o = *pu;
-    } while (!__sync_bool_compare_and_swap_16(pu, o, val.u));
-#else
-    qemu_build_not_reached();
+    {
+        __uint128_t *pu = __builtin_assume_aligned(pv, 16);
+        qatomic_set__nocheck(pu, val.u);
+        return;
+    }
 #endif
+#if defined(__x86_64__)
+    if (HAVE_al16_fast) {
+        asm("vmovdqa %1, %0" : "=m"(*(__uint128_t *)pv) : "x" (val.u));
+        return;
+    }
+#endif
+#if defined(CONFIG_CMPXCHG128)
+    {
+        __uint128_t *pu = __builtin_assume_aligned(pv, 16);
+        __uint128_t o;
+
+        /*
+         * Without CONFIG_ATOMIC128, __atomic_compare_exchange_n will always
+         * defer to libatomic, so we must use __sync_val_compare_and_swap_16
+         * and accept the sequential consistency that comes with it.
+         */
+        do {
+            o = *pu;
+        } while (!__sync_bool_compare_and_swap_16(pu, o, val.u));
+        return;
+    }
+#endif
+    qemu_build_not_reached();
 }
 
 /**
-- 
2.34.1