meson: Detect atomic128 support with optimization

There is an edge condition prior to gcc13 for which optimization
is required to generate 16-byte atomic sequences.  Detect this.

Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

accel/tcg/ldst_atomicity.c.inc

@@ -16,6 +16,23 @@
 #endif
 #define HAVE_al8_fast      (ATOMIC_REG_SIZE >= 8)
 
+/*
+ * If __alignof(unsigned __int128) < 16, GCC may refuse to inline atomics
+ * that are supported by the host, e.g. s390x.  We can force the pointer to
+ * have our known alignment with __builtin_assume_aligned, however prior to
+ * GCC 13 that was only reliable with optimization enabled.  See
+ *   https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107389
+ */
+#if defined(CONFIG_ATOMIC128_OPT)
+# if !defined(__OPTIMIZE__)
+#  define ATTRIBUTE_ATOMIC128_OPT  __attribute__((optimize("O1")))
+# endif
+# define CONFIG_ATOMIC128
+#endif
+#ifndef ATTRIBUTE_ATOMIC128_OPT
+# define ATTRIBUTE_ATOMIC128_OPT
+#endif
+
 #if defined(CONFIG_ATOMIC128)
 # define HAVE_al16_fast    true
 #else
@@ -152,7 +169,8 @@ static inline uint64_t load_atomic8(void *pv)
  *
  * Atomically load 16 aligned bytes from @pv.
  */
-static inline Int128 load_atomic16(void *pv)
+static inline Int128 ATTRIBUTE_ATOMIC128_OPT
+load_atomic16(void *pv)
 {
 #ifdef CONFIG_ATOMIC128
     __uint128_t *p = __builtin_assume_aligned(pv, 16);
@@ -356,7 +374,8 @@ static uint64_t load_atom_extract_al16_or_exit(CPUArchState *env, uintptr_t ra,
  * cross an 16-byte boundary then the access must be 16-byte atomic,
  * otherwise the access must be 8-byte atomic.
  */
-static inline uint64_t load_atom_extract_al16_or_al8(void *pv, int s)
+static inline uint64_t ATTRIBUTE_ATOMIC128_OPT
+load_atom_extract_al16_or_al8(void *pv, int s)
 {
 #if defined(CONFIG_ATOMIC128)
     uintptr_t pi = (uintptr_t)pv;
@@ -692,7 +711,8 @@ static inline void store_atomic8(void *pv, uint64_t val)
  *
  * Atomically store 16 aligned bytes to @pv.
  */
-static inline void store_atomic16(void *pv, Int128Alias val)
+static inline void ATTRIBUTE_ATOMIC128_OPT
+store_atomic16(void *pv, Int128Alias val)
 {
 #if defined(CONFIG_ATOMIC128)
     __uint128_t *pu = __builtin_assume_aligned(pv, 16);
@@ -790,7 +810,8 @@ static void store_atom_insert_al8(uint64_t *p, uint64_t val, uint64_t msk)
  *
  * Atomically store @val to @p masked by @msk.
  */
-static void store_atom_insert_al16(Int128 *ps, Int128Alias val, Int128Alias msk)
+static void ATTRIBUTE_ATOMIC128_OPT
+store_atom_insert_al16(Int128 *ps, Int128Alias val, Int128Alias msk)
 {
 #if defined(CONFIG_ATOMIC128)
     __uint128_t *pu, old, new;

meson.build

@@ -2249,23 +2249,21 @@ config_host_data.set('HAVE_BROKEN_SIZE_MAX', not cc.compiles('''
         return printf("%zu", SIZE_MAX);
     }''', args: ['-Werror']))
 
-atomic_test = '''
+# See if 64-bit atomic operations are supported.
+# Note that without __atomic builtins, we can only
+# assume atomic loads/stores max at pointer size.
+config_host_data.set('CONFIG_ATOMIC64', cc.links('''
   #include <stdint.h>
   int main(void)
   {
-    @0@ x = 0, y = 0;
+    uint64_t x = 0, y = 0;
     y = __atomic_load_n(&x, __ATOMIC_RELAXED);
     __atomic_store_n(&x, y, __ATOMIC_RELAXED);
     __atomic_compare_exchange_n(&x, &y, x, 0, __ATOMIC_RELAXED, __ATOMIC_RELAXED);
     __atomic_exchange_n(&x, y, __ATOMIC_RELAXED);
     __atomic_fetch_add(&x, y, __ATOMIC_RELAXED);
     return 0;
-  }'''
-
-# See if 64-bit atomic operations are supported.
-# Note that without __atomic builtins, we can only
-# assume atomic loads/stores max at pointer size.
-config_host_data.set('CONFIG_ATOMIC64', cc.links(atomic_test.format('uint64_t')))
+  }'''))
 
 has_int128 = cc.links('''
   __int128_t a;
@@ -2283,21 +2281,39 @@ if has_int128
   # "do we have 128-bit atomics which are handled inline and specifically not
   # via libatomic". The reason we can't use libatomic is documented in the
   # comment starting "GCC is a house divided" in include/qemu/atomic128.h.
-  has_atomic128 = cc.links(atomic_test.format('unsigned __int128'))
+  # We only care about these operations on 16-byte aligned pointers, so
+  # force 16-byte alignment of the pointer, which may be greater than
+  # __alignof(unsigned __int128) for the host.
+  atomic_test_128 = '''
+    int main(int ac, char **av) {
+      unsigned __int128 *p = __builtin_assume_aligned(av[ac - 1], sizeof(16));
+      p[1] = __atomic_load_n(&p[0], __ATOMIC_RELAXED);
+      __atomic_store_n(&p[2], p[3], __ATOMIC_RELAXED);
+      __atomic_compare_exchange_n(&p[4], &p[5], p[6], 0, __ATOMIC_RELAXED, __ATOMIC_RELAXED);
+      return 0;
+    }'''
+  has_atomic128 = cc.links(atomic_test_128)
 
   config_host_data.set('CONFIG_ATOMIC128', has_atomic128)
 
   if not has_atomic128
-    has_cmpxchg128 = cc.links('''
-      int main(void)
-      {
-        unsigned __int128 x = 0, y = 0;
-        __sync_val_compare_and_swap_16(&x, y, x);
-        return 0;
-      }
-    ''')
-
-    config_host_data.set('CONFIG_CMPXCHG128', has_cmpxchg128)
+    # Even with __builtin_assume_aligned, the above test may have failed
+    # without optimization enabled.  Try again with optimizations locally
+    # enabled for the function.  See
+    #   https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107389
+    has_atomic128_opt = cc.links('__attribute__((optimize("O1")))' + atomic_test_128)
+    config_host_data.set('CONFIG_ATOMIC128_OPT', has_atomic128_opt)
+
+    if not has_atomic128_opt
+      config_host_data.set('CONFIG_CMPXCHG128', cc.links('''
+        int main(void)
+        {
+          unsigned __int128 x = 0, y = 0;
+          __sync_val_compare_and_swap_16(&x, y, x);
+          return 0;
+        }
+      '''))
+    endif
   endif
 endif