diff mbox

Let target get_memmodel know about stores/barriers

Message ID 1358177908-31624-1-git-send-email-andi@firstfloor.org
State New
Headers show

Commit Message

Andi Kleen Jan. 14, 2013, 3:38 p.m. UTC
From: Andi Kleen <ak@linux.intel.com>

For HLE stores are only valid with __ATOMIC_HLE_RELEASE. The middle end
didn't know this. This adds a new parameter to the get_memmodel target
hook to distingush stores and give an warning for acquire stores.

I also did a similar check for the barriers where HLE is not useful
at all.

Fixes one todo in the earlier hle release patch.

gcc/:
2013-01-13  Andi Kleen  <ak@linux.intel.com>

	PR target/55948
	* builtins.c (get_memmodel): Pass store, barrier parameters to target
	hook.
	(expand_builtin_atomic_exchange,
         expand_builtin_atomic_compare_exchange,
         expand_builtin_atomic_load,
         expand_builtin_atomic_store,
         expand_builtin_atomic_fetch_op,
         expand_builtin_atomic_test_and_set,
         expand_builtin_atomic_test_lock,
         expand_builtin_atomic_clear,
         expand_builtin_atomic_thread_fence): Pass store, barrier
	parameters to get_memmodel.
	* config/i386/i386.c (ix86_memmodel_check): Add store, barrier warning.
	* doc/tm.texi (TARGET_GET_MEMMODEL): Add store, barrier parameters.
	* doc/tm.exit.in (TARGET_GET_MEMMODEL): Dito.
	* target.def (TARGET_GET_MEMMODEL): Dito.
---
 gcc/builtins.c         |   30 +++++++++++++++---------------
 gcc/config/i386/i386.c |   21 +++++++++++++++++++--
 gcc/doc/tm.texi        |    6 ++++--
 gcc/doc/tm.texi.in     |    4 +++-
 gcc/target.def         |    3 ++-
 5 files changed, 43 insertions(+), 21 deletions(-)

Comments

Uros Bizjak Jan. 14, 2013, 4:53 p.m. UTC | #1
On Mon, Jan 14, 2013 at 4:38 PM, Andi Kleen <andi@firstfloor.org> wrote:

> For HLE stores are only valid with __ATOMIC_HLE_RELEASE. The middle end
> didn't know this. This adds a new parameter to the get_memmodel target
> hook to distingush stores and give an warning for acquire stores.
>
> I also did a similar check for the barriers where HLE is not useful
> at all.
>
> Fixes one todo in the earlier hle release patch.
>
> gcc/:
> 2013-01-13  Andi Kleen  <ak@linux.intel.com>
>
>         PR target/55948
>         * builtins.c (get_memmodel): Pass store, barrier parameters to target
>         hook.
>         (expand_builtin_atomic_exchange,
>          expand_builtin_atomic_compare_exchange,
>          expand_builtin_atomic_load,
>          expand_builtin_atomic_store,
>          expand_builtin_atomic_fetch_op,
>          expand_builtin_atomic_test_and_set,
>          expand_builtin_atomic_test_lock,
>          expand_builtin_atomic_clear,
>          expand_builtin_atomic_thread_fence): Pass store, barrier
>         parameters to get_memmodel.
>         * config/i386/i386.c (ix86_memmodel_check): Add store, barrier warning.
>         * doc/tm.texi (TARGET_GET_MEMMODEL): Add store, barrier parameters.
>         * doc/tm.exit.in (TARGET_GET_MEMMODEL): Dito.
>         * target.def (TARGET_GET_MEMMODEL): Dito.
> ---
>  gcc/builtins.c         |   30 +++++++++++++++---------------
>  gcc/config/i386/i386.c |   21 +++++++++++++++++++--
>  gcc/doc/tm.texi        |    6 ++++--
>  gcc/doc/tm.texi.in     |    4 +++-
>  gcc/target.def         |    3 ++-
>  5 files changed, 43 insertions(+), 21 deletions(-)
>

> diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
> index 4f778c1..50b6297 100644
> --- a/gcc/config/i386/i386.c
> +++ b/gcc/config/i386/i386.c
> @@ -42076,14 +42076,31 @@ ix86_destroy_cost_data (void *data)
>    free (data);
>  }
>
> -/* Validate target specific memory model bits in VAL. */
> +/* Validate target specific memory model bits in VAL. When
> +   STORE is true this for a store.  When BARRIER is true
> +   this is for a barrier. */
>
>  static unsigned HOST_WIDE_INT
> -ix86_memmodel_check (unsigned HOST_WIDE_INT val)
> +ix86_memmodel_check (unsigned HOST_WIDE_INT val, bool store,
> +                    bool barrier)
>  {
>    unsigned HOST_WIDE_INT model = val & MEMMODEL_MASK;
>    bool strong;
>
> +  if (barrier && (val & (IX86_HLE_ACQUIRE|IX86_HLE_RELEASE)))
> +    {
> +      warning (OPT_Winvalid_memory_model,
> +              "__ATOMIC_HLE_ACQUIRE or __ATOMIC_HLE_RELEASE not valid for barriers.");

No dot at the end of warning sentence.

> +      return MEMMODEL_SEQ_CST;
> +    }
> +
> +  if (store && (val & IX86_HLE_ACQUIRE))
> +    {
> +      warning (OPT_Winvalid_memory_model,
> +              "__ATOMIC_HLE_ACQUIRE not valid for stores.");

Also here.

The target part is OK, the rest needs a review from a middle-end reviewer.

Thanks,
Uros.
diff mbox

Patch

diff --git a/gcc/builtins.c b/gcc/builtins.c
index 1d0aec6..f9c5656 100644
--- a/gcc/builtins.c
+++ b/gcc/builtins.c
@@ -5272,10 +5272,12 @@  expand_builtin_sync_lock_release (enum machine_mode mode, tree exp)
 }
 
 /* Given an integer representing an ``enum memmodel'', verify its
-   correctness and return the memory model enum.  */
+   correctness and return the memory model enum.  When STORE
+   is true this for a store only. When BARRIER is true this if for
+   a barrier.  */
 
 static enum memmodel
-get_memmodel (tree exp)
+get_memmodel (tree exp, bool store, bool barrier)
 {
   rtx op;
   unsigned HOST_WIDE_INT val;
@@ -5289,7 +5291,7 @@  get_memmodel (tree exp)
 
   val = INTVAL (op);
   if (targetm.memmodel_check)
-    val = targetm.memmodel_check (val);
+    val = targetm.memmodel_check (val, store, barrier);
   else if (val & ~MEMMODEL_MASK)
     {
       warning (OPT_Winvalid_memory_model,
@@ -5318,7 +5320,7 @@  expand_builtin_atomic_exchange (enum machine_mode mode, tree exp, rtx target)
   rtx val, mem;
   enum memmodel model;
 
-  model = get_memmodel (CALL_EXPR_ARG (exp, 2));
+  model = get_memmodel (CALL_EXPR_ARG (exp, 2), false, false);
   if ((model & MEMMODEL_MASK) == MEMMODEL_CONSUME)
     {
       error ("invalid memory model for %<__atomic_exchange%>");
@@ -5352,8 +5354,8 @@  expand_builtin_atomic_compare_exchange (enum machine_mode mode, tree exp,
   tree weak;
   bool is_weak;
 
-  success = get_memmodel (CALL_EXPR_ARG (exp, 4));
-  failure = get_memmodel (CALL_EXPR_ARG (exp, 5));
+  success = get_memmodel (CALL_EXPR_ARG (exp, 4), false, false);
+  failure = get_memmodel (CALL_EXPR_ARG (exp, 5), false, false);
 
   if ((failure & MEMMODEL_MASK) == MEMMODEL_RELEASE
       || (failure & MEMMODEL_MASK) == MEMMODEL_ACQ_REL)
@@ -5408,7 +5410,7 @@  expand_builtin_atomic_load (enum machine_mode mode, tree exp, rtx target)
   rtx mem;
   enum memmodel model;
 
-  model = get_memmodel (CALL_EXPR_ARG (exp, 1));
+  model = get_memmodel (CALL_EXPR_ARG (exp, 1), false, false);
   if ((model & MEMMODEL_MASK) == MEMMODEL_RELEASE
       || (model & MEMMODEL_MASK) == MEMMODEL_ACQ_REL)
     {
@@ -5437,7 +5439,7 @@  expand_builtin_atomic_store (enum machine_mode mode, tree exp)
   rtx mem, val;
   enum memmodel model;
 
-  model = get_memmodel (CALL_EXPR_ARG (exp, 2));
+  model = get_memmodel (CALL_EXPR_ARG (exp, 2), true, false);
   if ((model & MEMMODEL_MASK) != MEMMODEL_RELAXED
       && (model & MEMMODEL_MASK) != MEMMODEL_SEQ_CST
       && (model & MEMMODEL_MASK) != MEMMODEL_RELEASE)
@@ -5477,7 +5479,7 @@  expand_builtin_atomic_fetch_op (enum machine_mode mode, tree exp, rtx target,
   tree fndecl;
   tree addr;
 
-  model = get_memmodel (CALL_EXPR_ARG (exp, 2));
+  model = get_memmodel (CALL_EXPR_ARG (exp, 2), false, false);
 
   /* Expand the operands.  */
   mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
@@ -5544,7 +5546,7 @@  expand_builtin_atomic_clear (tree exp)
 
   mode = mode_for_size (BOOL_TYPE_SIZE, MODE_INT, 0);
   mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
-  model = get_memmodel (CALL_EXPR_ARG (exp, 1));
+  model = get_memmodel (CALL_EXPR_ARG (exp, 1), true, false);
 
   if ((model & MEMMODEL_MASK) == MEMMODEL_ACQUIRE
       || (model & MEMMODEL_MASK) == MEMMODEL_ACQ_REL)
@@ -5553,8 +5555,6 @@  expand_builtin_atomic_clear (tree exp)
       return const0_rtx;
     }
 
-  /* need target hook there to check for not hle acquire */
-
   if (HAVE_atomic_clear)
     {
       emit_insn (gen_atomic_clear (mem, model));
@@ -5585,7 +5585,7 @@  expand_builtin_atomic_test_and_set (tree exp, rtx target)
 
   mode = mode_for_size (BOOL_TYPE_SIZE, MODE_INT, 0);
   mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
-  model = get_memmodel (CALL_EXPR_ARG (exp, 1));
+  model = get_memmodel (CALL_EXPR_ARG (exp, 1), false, false);
 
   return expand_atomic_test_and_set (target, mem, model);
 }
@@ -5724,7 +5724,7 @@  expand_builtin_atomic_is_lock_free (tree exp)
 static void
 expand_builtin_atomic_thread_fence (tree exp)
 {
-  enum memmodel model = get_memmodel (CALL_EXPR_ARG (exp, 0));
+  enum memmodel model = get_memmodel (CALL_EXPR_ARG (exp, 0), false, true);
   expand_mem_thread_fence (model);
 }
 
@@ -5735,7 +5735,7 @@  expand_builtin_atomic_thread_fence (tree exp)
 static void
 expand_builtin_atomic_signal_fence (tree exp)
 {
-  enum memmodel model = get_memmodel (CALL_EXPR_ARG (exp, 0));
+  enum memmodel model = get_memmodel (CALL_EXPR_ARG (exp, 0), false, true);
   expand_mem_signal_fence (model);
 }
 
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 4f778c1..50b6297 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -42076,14 +42076,31 @@  ix86_destroy_cost_data (void *data)
   free (data);
 }
 
-/* Validate target specific memory model bits in VAL. */
+/* Validate target specific memory model bits in VAL. When
+   STORE is true this for a store.  When BARRIER is true
+   this is for a barrier. */
 
 static unsigned HOST_WIDE_INT
-ix86_memmodel_check (unsigned HOST_WIDE_INT val)
+ix86_memmodel_check (unsigned HOST_WIDE_INT val, bool store,
+		     bool barrier)
 {
   unsigned HOST_WIDE_INT model = val & MEMMODEL_MASK;
   bool strong;
 
+  if (barrier && (val & (IX86_HLE_ACQUIRE|IX86_HLE_RELEASE)))
+    {
+      warning (OPT_Winvalid_memory_model,
+	       "__ATOMIC_HLE_ACQUIRE or __ATOMIC_HLE_RELEASE not valid for barriers.");
+      return MEMMODEL_SEQ_CST;
+    }
+
+  if (store && (val & IX86_HLE_ACQUIRE))
+    {
+      warning (OPT_Winvalid_memory_model,
+	       "__ATOMIC_HLE_ACQUIRE not valid for stores.");
+      return MEMMODEL_SEQ_CST;
+    }
+
   if (val & ~(unsigned HOST_WIDE_INT)(IX86_HLE_ACQUIRE|IX86_HLE_RELEASE
 				      |MEMMODEL_MASK)
       || ((val & IX86_HLE_ACQUIRE) && (val & IX86_HLE_RELEASE)))
diff --git a/gcc/doc/tm.texi b/gcc/doc/tm.texi
index 9d6f6bc..01bc562 100644
--- a/gcc/doc/tm.texi
+++ b/gcc/doc/tm.texi
@@ -11338,9 +11338,11 @@  Address Sanitizer shadow memory address.  NULL if Address Sanitizer is not
 supported by the target.
 @end deftypefn
 
-@deftypefn {Target Hook} {unsigned HOST_WIDE_INT} TARGET_MEMMODEL_CHECK (unsigned HOST_WIDE_INT @var{val})
+@deftypefn {Target Hook} {unsigned HOST_WIDE_INT} TARGET_MEMMODEL_CHECK (unsigned HOST_WIDE_INT @var{val}, bool @var{store}, bool @var{barrier})
 Validate target specific memory model mask bits. When NULL no target specific
-memory model bits are allowed.
+memory model bits are allowed.  When @code{STORE} is true this is for
+__atomic_clear or __atomic_store. When @code{BARRIER} is true this is for a 
+barrier.
 @end deftypefn
 
 @deftypevr {Target Hook} {unsigned char} TARGET_ATOMIC_TEST_AND_SET_TRUEVAL
diff --git a/gcc/doc/tm.texi.in b/gcc/doc/tm.texi.in
index 3668882..87f654a 100644
--- a/gcc/doc/tm.texi.in
+++ b/gcc/doc/tm.texi.in
@@ -11178,7 +11178,9 @@  is zero, which disables this optimization.
 
 @hook TARGET_MEMMODEL_CHECK
 Validate target specific memory model mask bits. When NULL no target specific
-memory model bits are allowed.
+memory model bits are allowed.  When @code{STORE} is true this is for
+__atomic_clear or __atomic_store. When @code{BARRIER} is true this is for a 
+barrier.
 @end deftypefn
 
 @hook TARGET_ATOMIC_TEST_AND_SET_TRUEVAL
diff --git a/gcc/target.def b/gcc/target.def
index 8627923..46957e9 100644
--- a/gcc/target.def
+++ b/gcc/target.def
@@ -2059,7 +2059,8 @@  DEFHOOKPOD
 DEFHOOK
 (memmodel_check,
  "",
- unsigned HOST_WIDE_INT, (unsigned HOST_WIDE_INT val), NULL)
+ unsigned HOST_WIDE_INT, (unsigned HOST_WIDE_INT val, bool store, bool barrier),
+ NULL)
 
 /* Defines an offset bitwise ored into shifted address to get corresponding
    Address Sanitizer shadow address, or -1 if Address Sanitizer is not