[PATCH v5 14/54] accel/tcg: Add have_lse2 support in ldst_atomicity

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

Add fast paths for FEAT_LSE2, using the detection in tcg.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 accel/tcg/ldst_atomicity.c.inc | 37 ++++++++++++++++++++++++++++++----
 1 file changed, 33 insertions(+), 4 deletions(-)

diff --git a/accel/tcg/ldst_atomicity.c.inc b/accel/tcg/ldst_atomicity.c.inc
index c3b2b35823..baa935ab8b 100644
--- a/accel/tcg/ldst_atomicity.c.inc
+++ b/accel/tcg/ldst_atomicity.c.inc
@@ -41,6 +41,8 @@
  * but we're using tcg/tci/ instead.
  */
 # define HAVE_al16_fast    false
+#elif defined(__aarch64__)
+# define HAVE_al16_fast    likely(have_lse2)
 #elif defined(__x86_64__) && defined(CONFIG_INT128)
 # define HAVE_al16_fast    likely(have_atomic16)
 #else
@@ -48,6 +50,8 @@
 #endif
 #if defined(CONFIG_ATOMIC128) || defined(CONFIG_CMPXCHG128)
 # define HAVE_al16         true
+#elif defined(__aarch64__)
+# define HAVE_al16         true
 #else
 # define HAVE_al16         false
 #endif
@@ -186,6 +190,14 @@ load_atomic16(void *pv)
 
     r.u = qatomic_read__nocheck(p);
     return r.s;
+#elif defined(__aarch64__)
+    uint64_t l, h;
+
+    /* Via HAVE_al16_fast, FEAT_LSE2 is present: LDP becomes atomic. */
+    asm("ldp %0, %1, %2" : "=r"(l), "=r"(h) : "m"(*(__uint128_t *)pv));
+
+    qemu_build_assert(!HOST_BIG_ENDIAN);
+    return int128_make128(l, h);
 #elif defined(__x86_64__) && defined(CONFIG_INT128)
     Int128Alias r;
 
@@ -428,6 +440,18 @@ load_atom_extract_al16_or_al8(void *pv, int s)
         r = qatomic_read__nocheck(p16);
     }
     return r >> shr;
+#elif defined(__aarch64__)
+    /*
+     * Via HAVE_al16_fast, FEAT_LSE2 is present.
+     * LDP becomes single-copy atomic if 16-byte aligned, and
+     * single-copy atomic on the parts if 8-byte aligned.
+     */
+    uintptr_t pi = (uintptr_t)pv;
+    int shr = (pi & 7) * 8;
+    uint64_t l, h;
+
+    asm("ldp %0, %1, %2" : "=r"(l), "=r"(h) : "m"(*(__uint128_t *)(pi & ~7)));
+    return (l >> shr) | (h << (-shr & 63));
 #elif defined(__x86_64__) && defined(CONFIG_INT128)
     uintptr_t pi = (uintptr_t)pv;
     int shr = (pi & 7) * 8;
@@ -783,10 +807,15 @@ store_atomic16(void *pv, Int128Alias val)
         l = int128_getlo(val.s);
         h = int128_gethi(val.s);
 
-        asm("0: ldxp %0, xzr, %1\n\t"
-            "stxp %w0, %2, %3, %1\n\t"
-            "cbnz %w0, 0b"
-            : "=&r"(t), "=Q"(*(__uint128_t *)pv) : "r"(l), "r"(h));
+        if (HAVE_al16_fast) {
+            /* Via HAVE_al16_fast, FEAT_LSE2 is present: STP becomes atomic. */
+            asm("stp %1, %2, %0" : "=Q"(*(__uint128_t *)pv) : "r"(l), "r"(h));
+        } else {
+            asm("0: ldxp %0, xzr, %1\n\t"
+                "stxp %w0, %2, %3, %1\n\t"
+                "cbnz %w0, 0b"
+                : "=&r"(t), "=Q"(*(__uint128_t *)pv) : "r"(l), "r"(h));
+        }
         return;
     }
 #elif defined(CONFIG_CMPXCHG128)
-- 
2.34.1