Support ILP32 in AArch64 assembly routines (GCC)

(back-port of commit 3f41ecb from 'release-7_6')

Issue #38 (libatomic_ops).

* src/atomic_ops/sysdeps/gcc/aarch64.h [(!__clang__
|| AO_AARCH64_ASM_LOAD_STORE_CAS) && !AO_PREFER_GENERALIZED
&& __ILP32__] (AO_double_load, AO_double_load_acquire, AO_double_store,
AO_double_store_release): Specify the size ("w") of arguments of the
aarch64 instructions.
* src/atomic_ops/sysdeps/gcc/aarch64.h [(!__clang__
|| AO_AARCH64_ASM_LOAD_STORE_CAS) && __ILP32__]
(AO_double_compare_and_swap, AO_double_compare_and_swap_acquire,
AO_double_compare_and_swap_release, AO_double_compare_and_swap_full):
Likewise.
* src/atomic_ops/sysdeps/standard_ao_double_t.h [__aarch64__
&& __ILP32__] (double_ptr_storage): Define as unsigned long long
(instead of unsigned __int128).

diff --git a/src/atomic_ops/sysdeps/gcc/aarch64.h b/src/atomic_ops/sysdeps/gcc/aarch64.h
index b9624ef..6db47fb 100644 (file)
--- a/src/atomic_ops/sysdeps/gcc/aarch64.h
+++ b/src/atomic_ops/sysdeps/gcc/aarch64.h
@@ -41,8 +41,13 @@
     /* single-copy atomic (unlike LDREXD for 32-bit ARM).               */
     do {
       __asm__ __volatile__("//AO_double_load\n"
-      "       ldxp  %0, %1, %3\n"
-      "       stxp %w2, %0, %1, %3"
+#       ifdef __ILP32__
+          "       ldxp  %w0, %w1, %3\n"
+          "       stxp %w2, %w0, %w1, %3"
+#       else
+          "       ldxp  %0, %1, %3\n"
+          "       stxp %w2, %0, %1, %3"
+#       endif
       : "=&r" (result.AO_val1), "=&r" (result.AO_val2), "=&r" (status)
       : "Q" (*addr));
     } while (AO_EXPECT_FALSE(status));
@@ -58,8 +63,13 @@
 
     do {
       __asm__ __volatile__("//AO_double_load_acquire\n"
-      "       ldaxp  %0, %1, %3\n"
-      "       stxp %w2, %0, %1, %3"
+#       ifdef __ILP32__
+          "       ldaxp  %w0, %w1, %3\n"
+          "       stxp %w2, %w0, %w1, %3"
+#       else
+          "       ldaxp  %0, %1, %3\n"
+          "       stxp %w2, %0, %1, %3"
+#       endif
       : "=&r" (result.AO_val1), "=&r" (result.AO_val2), "=&r" (status)
       : "Q" (*addr));
     } while (AO_EXPECT_FALSE(status));
@@ -75,8 +85,13 @@
 
     do {
       __asm__ __volatile__("//AO_double_store\n"
-      "       ldxp  %0, %1, %3\n"
-      "       stxp %w2, %4, %5, %3"
+#       ifdef __ILP32__
+          "       ldxp  %w0, %w1, %3\n"
+          "       stxp %w2, %w4, %w5, %3"
+#       else
+          "       ldxp  %0, %1, %3\n"
+          "       stxp %w2, %4, %5, %3"
+#       endif
       : "=&r" (old_val.AO_val1), "=&r" (old_val.AO_val2), "=&r" (status),
         "=Q" (*addr)
       : "r" (value.AO_val1), "r" (value.AO_val2));
@@ -94,8 +109,13 @@
 
     do {
       __asm__ __volatile__("//AO_double_store_release\n"
-      "       ldxp  %0, %1, %3\n"
-      "       stlxp %w2, %4, %5, %3"
+#       ifdef __ILP32__
+          "       ldxp  %w0, %w1, %3\n"
+          "       stlxp %w2, %w4, %w5, %3"
+#       else
+          "       ldxp  %0, %1, %3\n"
+          "       stlxp %w2, %4, %5, %3"
+#       endif
       : "=&r" (old_val.AO_val1), "=&r" (old_val.AO_val2), "=&r" (status),
         "=Q" (*addr)
       : "r" (value.AO_val1), "r" (value.AO_val2));
@@ -112,13 +132,21 @@
 
     do {
       __asm__ __volatile__("//AO_double_compare_and_swap\n"
-        "       ldxp  %0, %1, %2\n"
+#       ifdef __ILP32__
+          "       ldxp  %w0, %w1, %2\n"
+#       else
+          "       ldxp  %0, %1, %2\n"
+#       endif
         : "=&r" (tmp.AO_val1), "=&r" (tmp.AO_val2)
         : "Q" (*addr));
       if (tmp.AO_val1 != old_val.AO_val1 || tmp.AO_val2 != old_val.AO_val2)
         break;
       __asm__ __volatile__(
-        "       stxp %w0, %2, %3, %1\n"
+#       ifdef __ILP32__
+          "       stxp %w0, %w2, %w3, %1\n"
+#       else
+          "       stxp %w0, %2, %3, %1\n"
+#       endif
         : "=&r" (result), "=Q" (*addr)
         : "r" (new_val.AO_val1), "r" (new_val.AO_val2));
     } while (AO_EXPECT_FALSE(result));
@@ -135,13 +163,21 @@
 
     do {
       __asm__ __volatile__("//AO_double_compare_and_swap_acquire\n"
-        "       ldaxp  %0, %1, %2\n"
+#       ifdef __ILP32__
+          "       ldaxp  %w0, %w1, %2\n"
+#       else
+          "       ldaxp  %0, %1, %2\n"
+#       endif
         : "=&r" (tmp.AO_val1), "=&r" (tmp.AO_val2)
         : "Q" (*addr));
       if (tmp.AO_val1 != old_val.AO_val1 || tmp.AO_val2 != old_val.AO_val2)
         break;
       __asm__ __volatile__(
-        "       stxp %w0, %2, %3, %1\n"
+#       ifdef __ILP32__
+          "       stxp %w0, %w2, %w3, %1\n"
+#       else
+          "       stxp %w0, %2, %3, %1\n"
+#       endif
         : "=&r" (result), "=Q" (*addr)
         : "r" (new_val.AO_val1), "r" (new_val.AO_val2));
     } while (AO_EXPECT_FALSE(result));
@@ -158,13 +194,21 @@
 
     do {
       __asm__ __volatile__("//AO_double_compare_and_swap_release\n"
-        "       ldxp  %0, %1, %2\n"
+#       ifdef __ILP32__
+          "       ldxp  %w0, %w1, %2\n"
+#       else
+          "       ldxp  %0, %1, %2\n"
+#       endif
         : "=&r" (tmp.AO_val1), "=&r" (tmp.AO_val2)
         : "Q" (*addr));
       if (tmp.AO_val1 != old_val.AO_val1 || tmp.AO_val2 != old_val.AO_val2)
         break;
       __asm__ __volatile__(
-        "       stlxp %w0, %2, %3, %1\n"
+#       ifdef __ILP32__
+          "       stlxp %w0, %w2, %w3, %1\n"
+#       else
+          "       stlxp %w0, %2, %3, %1\n"
+#       endif
         : "=&r" (result), "=Q" (*addr)
         : "r" (new_val.AO_val1), "r" (new_val.AO_val2));
     } while (AO_EXPECT_FALSE(result));
@@ -181,13 +225,21 @@
 
     do {
       __asm__ __volatile__("//AO_double_compare_and_swap_full\n"
-        "       ldaxp  %0, %1, %2\n"
+#       ifdef __ILP32__
+          "       ldaxp  %w0, %w1, %2\n"
+#       else
+          "       ldaxp  %0, %1, %2\n"
+#       endif
         : "=&r" (tmp.AO_val1), "=&r" (tmp.AO_val2)
         : "Q" (*addr));
       if (tmp.AO_val1 != old_val.AO_val1 || tmp.AO_val2 != old_val.AO_val2)
         break;
       __asm__ __volatile__(
-        "       stlxp %w0, %2, %3, %1\n"
+#       ifdef __ILP32__
+          "       stlxp %w0, %w2, %w3, %1\n"
+#       else
+          "       stlxp %w0, %2, %3, %1\n"
+#       endif
         : "=&r" (result), "=Q" (*addr)
         : "r" (new_val.AO_val1), "r" (new_val.AO_val2));
     } while (AO_EXPECT_FALSE(result));

diff --git a/src/atomic_ops/sysdeps/standard_ao_double_t.h b/src/atomic_ops/sysdeps/standard_ao_double_t.h
index 2a676ea..fd0dd6e 100644 (file)
--- a/src/atomic_ops/sysdeps/standard_ao_double_t.h
+++ b/src/atomic_ops/sysdeps/standard_ao_double_t.h
@@ -32,7 +32,7 @@
   typedef __m128 double_ptr_storage;
 #elif defined(_WIN32) && !defined(__GNUC__)
   typedef unsigned __int64 double_ptr_storage;
-#elif defined(__aarch64__)
+#elif defined(__aarch64__) && !defined(__ILP32__)
   typedef unsigned __int128 double_ptr_storage;
 #elif defined(__i386__) && defined(__GNUC__)
   typedef unsigned long long double_ptr_storage