@@ -369,6 +369,10 @@ extern void (*arm_lang_output_object_attributes_hook)(void);
/* Nonzero if this chip supports ldrex and strex */
#define TARGET_HAVE_LDREX ((arm_arch6 && TARGET_ARM) || arm_arch7)
+/* Nonzero if this chip supports LPAE. */
+#define TARGET_HAVE_LPAE \
+ (arm_arch7 && ((insn_flags & FL_FOR_ARCH7VE) == FL_FOR_ARCH7VE))
+
/* Nonzero if this chip supports ldrex{bh} and strex{bh}. */
#define TARGET_HAVE_LDREXBH ((arm_arch6k && TARGET_ARM) || arm_arch7)
@@ -99,32 +99,62 @@ (define_insn "atomic_store<mode>"
[(set_attr "predicable" "yes")
(set_attr "predicable_short_it" "no")])
-;; Note that ldrd and vldr are *not* guaranteed to be single-copy atomic,
-;; even for a 64-bit aligned address. Instead we use a ldrexd unparied
-;; with a store.
+;; An LDRD instruction usable by the atomic_loaddi expander on LPAE targets
+
+(define_insn "arm_atomic_loaddi2_ldrd"
+ [(set (match_operand:DI 0 "register_operand" "=r")
+ (unspec_volatile:DI
+ [(match_operand:DI 1 "arm_sync_memory_operand" "Q")]
+ VUNSPEC_LDRD_ATOMIC))]
+ "ARM_DOUBLEWORD_ALIGN && TARGET_HAVE_LPAE"
+ "ldrd%?\t%0, %H0, %C1"
+ [(set_attr "predicable" "yes")
+ (set_attr "predicable_short_it" "no")])
+
+;; There are three ways to expand this depending on the architecture
+;; features available. As for the barriers, a load needs a barrier
+;; after it on all non-relaxed memory models except when the load
+;; has acquire semantics (for ARMv8-A).
+
(define_expand "atomic_loaddi"
[(match_operand:DI 0 "s_register_operand") ;; val out
(match_operand:DI 1 "mem_noofs_operand") ;; memory
(match_operand:SI 2 "const_int_operand")] ;; model
- "TARGET_HAVE_LDREXD && ARM_DOUBLEWORD_ALIGN"
+ "(TARGET_HAVE_LDREXD || TARGET_HAVE_LPAE || TARGET_HAVE_LDACQ)
+ && ARM_DOUBLEWORD_ALIGN"
{
- enum memmodel model = memmodel_from_int (INTVAL (operands[2]));
- expand_mem_thread_fence (model);
- emit_insn (gen_atomic_loaddi_1 (operands[0], operands[1]));
- if (is_mm_seq_cst (model))
+ memmodel model = memmodel_from_int (INTVAL (operands[2]));
+
+ /* For ARMv8-A we can use an LDAEXD to atomically load two 32-bit registers
+ when acquire or stronger semantics are needed. When the relaxed model is
+ used this can be relaxed to a normal LDRD. */
+ if (TARGET_HAVE_LDACQ)
+ {
+ if (is_mm_relaxed (model))
+ emit_insn (gen_arm_atomic_loaddi2_ldrd (operands[0], operands[1]));
+ else
+ emit_insn (gen_arm_load_acquire_exclusivedi (operands[0], operands[1]));
+
+ DONE;
+ }
+
+ /* On LPAE targets LDRD and STRD accesses to 64-bit aligned
+ locations are 64-bit single-copy atomic. We still need barriers in the
+ appropriate places to implement the ordering constraints. */
+ if (TARGET_HAVE_LPAE)
+ emit_insn (gen_arm_atomic_loaddi2_ldrd (operands[0], operands[1]));
+ else
+ emit_insn (gen_arm_load_exclusivedi (operands[0], operands[1]));
+
+
+ /* All non-relaxed models need a barrier after the load when load-acquire
+ instructions are not available. */
+ if (!is_mm_relaxed (model))
expand_mem_thread_fence (model);
+
DONE;
})
-(define_insn "atomic_loaddi_1"
- [(set (match_operand:DI 0 "s_register_operand" "=r")
- (unspec:DI [(match_operand:DI 1 "mem_noofs_operand" "Ua")]
- UNSPEC_LL))]
- "TARGET_HAVE_LDREXD && ARM_DOUBLEWORD_ALIGN"
- "ldrexd%?\t%0, %H0, %C1"
- [(set_attr "predicable" "yes")
- (set_attr "predicable_short_it" "no")])
-
(define_expand "atomic_compare_and_swap<mode>"
[(match_operand:SI 0 "s_register_operand" "") ;; bool out
(match_operand:QHSD 1 "s_register_operand" "") ;; val out
@@ -138,6 +138,7 @@ (define_c_enum "unspecv" [
VUNSPEC_ATOMIC_XCHG ; Represent an atomic exchange.
VUNSPEC_ATOMIC_OP ; Represent an atomic operation.
VUNSPEC_LL ; Represent a load-register-exclusive.
+ VUNSPEC_LDRD_ATOMIC ; Represent an LDRD used as an atomic DImode load.
VUNSPEC_SC ; Represent a store-register-exclusive.
VUNSPEC_LAX ; Represent a load-register-acquire-exclusive.
VUNSPEC_SLX ; Represent a store-register-release-exclusive.
new file mode 100644
@@ -0,0 +1,9 @@
+/* { dg-do compile } */
+/* { dg-options "-std=c11 -O" } */
+/* { dg-require-effective-target arm_arch_v7a_ok } */
+/* { dg-add-options arm_arch_v7a } */
+
+#include "atomic_loaddi_acquire.x"
+
+/* { dg-final { scan-assembler-times "ldrexd\tr\[0-9\]+, r\[0-9\]+, \\\[r\[0-9\]+\\\]" 1 } } */
+/* { dg-final { scan-assembler-times "dmb\tsy" 1 } } */
new file mode 100644
@@ -0,0 +1,9 @@
+/* { dg-do compile } */
+/* { dg-options "-std=c11 -O" } */
+/* { dg-require-effective-target arm_arch_v7ve_ok } */
+/* { dg-add-options arm_arch_v7ve } */
+
+#include "atomic_loaddi_acquire.x"
+
+/* { dg-final { scan-assembler-times "ldrd\tr\[0-9\]+, r\[0-9\]+, \\\[r\[0-9\]+\\\]" 1 } } */
+/* { dg-final { scan-assembler-times "dmb\tsy" 1 } } */
new file mode 100644
@@ -0,0 +1,9 @@
+/* { dg-do compile } */
+/* { dg-options "-std=c11 -O" } */
+/* { dg-require-effective-target arm_arch_v8a_ok } */
+/* { dg-add-options arm_arch_v8a } */
+
+#include "atomic_loaddi_acquire.x"
+
+/* { dg-final { scan-assembler-times "ldaexd\tr\[0-9\]+, r\[0-9\]+, \\\[r\[0-9\]+\\\]" 1 } } */
+/* { dg-final { scan-assembler-not "dmb\tsy" } } */
new file mode 100644
@@ -0,0 +1,9 @@
+/* { dg-do compile } */
+/* { dg-options "-std=c11 -O" } */
+/* { dg-require-effective-target arm_arch_v7a_ok } */
+/* { dg-add-options arm_arch_v7a } */
+
+#include "atomic_loaddi_relaxed.x"
+
+/* { dg-final { scan-assembler-times "ldrexd\tr\[0-9\]+, r\[0-9\]+, \\\[r\[0-9\]+\\\]" 1 } } */
+/* { dg-final { scan-assembler-not "dmb\tsy" } } */
new file mode 100644
@@ -0,0 +1,9 @@
+/* { dg-do compile } */
+/* { dg-options "-std=c11 -O" } */
+/* { dg-require-effective-target arm_arch_v7ve_ok } */
+/* { dg-add-options arm_arch_v7ve } */
+
+#include "atomic_loaddi_relaxed.x"
+
+/* { dg-final { scan-assembler-times "ldrd\tr\[0-9\]+, r\[0-9\]+, \\\[r\[0-9\]+\\\]" 1 } } */
+/* { dg-final { scan-assembler-not "dmb\tsy" } } */
\ No newline at end of file
new file mode 100644
@@ -0,0 +1,9 @@
+/* { dg-do compile } */
+/* { dg-options "-std=c11 -O" } */
+/* { dg-require-effective-target arm_arch_v8a_ok } */
+/* { dg-add-options arm_arch_v8a } */
+
+#include "atomic_loaddi_relaxed.x"
+
+/* { dg-final { scan-assembler-times "ldrd\tr\[0-9\]+, r\[0-9\]+, \\\[r\[0-9\]+\\\]" 1 } } */
+/* { dg-final { scan-assembler-not "dmb\tsy" } } */
\ No newline at end of file
new file mode 100644
@@ -0,0 +1,9 @@
+/* { dg-do compile } */
+/* { dg-options "-std=c11 -O" } */
+/* { dg-require-effective-target arm_arch_v7a_ok } */
+/* { dg-add-options arm_arch_v7a } */
+
+#include "atomic_loaddi_seq_cst.x"
+
+/* { dg-final { scan-assembler-times "ldrexd\tr\[0-9\]+, r\[0-9\]+, \\\[r\[0-9\]+\\\]" 1 } } */
+/* { dg-final { scan-assembler-times "dmb\tsy" 1 } } */
new file mode 100644
@@ -0,0 +1,9 @@
+/* { dg-do compile } */
+/* { dg-options "-std=c11 -O" } */
+/* { dg-require-effective-target arm_arch_v7ve_ok } */
+/* { dg-add-options arm_arch_v7ve } */
+
+#include "atomic_loaddi_seq_cst.x"
+
+/* { dg-final { scan-assembler-times "ldrd\tr\[0-9\]+, r\[0-9\]+, \\\[r\[0-9\]+\\\]" 1 } } */
+/* { dg-final { scan-assembler-times "dmb\tsy" 1 } } */
new file mode 100644
@@ -0,0 +1,9 @@
+/* { dg-do compile } */
+/* { dg-options "-std=c11 -O" } */
+/* { dg-require-effective-target arm_arch_v8a_ok } */
+/* { dg-add-options arm_arch_v8a } */
+
+#include "atomic_loaddi_seq_cst.x"
+
+/* { dg-final { scan-assembler-times "ldaexd\tr\[0-9\]+, r\[0-9\]+, \\\[r\[0-9\]+\\\]" 1 } } */
+/* { dg-final { scan-assembler-not "dmb\tsy" } } */
new file mode 100644
@@ -0,0 +1,11 @@
+#include <stdatomic.h>
+
+atomic_ullong foo;
+int glob;
+
+int
+main (void)
+{
+ atomic_load_explicit (&foo, memory_order_acquire);
+ return glob;
+}
new file mode 100644
@@ -0,0 +1,11 @@
+#include <stdatomic.h>
+
+atomic_ullong foo;
+int glob;
+
+int
+main (void)
+{
+ atomic_load_explicit (&foo, memory_order_relaxed);
+ return glob;
+}
new file mode 100644
@@ -0,0 +1,11 @@
+#include <stdatomic.h>
+
+atomic_ullong foo;
+int glob;
+
+int
+main (void)
+{
+ atomic_load_explicit (&foo, memory_order_seq_cst);
+ return glob;
+}