c-family, c++: Fix up handling of types which may have padding in __atomic_{compare_...

On Fri, Feb 16, 2024 at 01:51:54PM +0000, Jonathan Wakely wrote:
> Ah, although __atomic_compare_exchange only takes pointers, the
> compiler replaces that with a call to __atomic_compare_exchange_n
> which takes the newval by value, which presumably uses an 80-bit FP
> register and so the padding bits become indeterminate again.

The problem is that __atomic_{,compare_}exchange lowering if it has
a supported atomic 1/2/4/8/16 size emits code like:
  _3 = *p2;
  _4 = VIEW_CONVERT_EXPR<I_type> (_3);
so if long double or some small struct etc. has some carefully filled
padding bits, those bits can be lost on the assignment.  The library call
for __atomic_{,compare_}exchange would actually work because it woiuld
load the value from memory using integral type or memcpy.
E.g. on
void
foo (long double *a, long double *b, long double *c)
{
  __atomic_compare_exchange (a, b, c, false, __ATOMIC_RELAXED, __ATOMIC_RELAXED);
}
we end up with -O0 with:
        fldt    (%rax)
        fstpt   -48(%rbp)
        movq    -48(%rbp), %rax
        movq    -40(%rbp), %rdx
i.e. load *c from memory into 387 register, store it back to uninitialized
stack slot (the padding bits are now random in there) and then load a
__uint128_t (pair of GPR regs).  The problem is that we first load it using
whatever type the pointer points to and then VIEW_CONVERT_EXPR that value:
  p2 = build_indirect_ref (loc, p2, RO_UNARY_STAR);
  p2 = build1 (VIEW_CONVERT_EXPR, I_type, p2);
The following patch fixes that by creating a MEM_REF instead, with the
I_type type, but with the original pointer type on the second argument for
aliasing purposes, so we actually preserve the padding bits that way.
With this patch instead of the above assembly we emit
        movq    8(%rax), %rdx
        movq    (%rax), %rax
I had to add support for MEM_REF in pt.cc, though with the assumption
that it has been already originally created with non-dependent
types/operands (which is the case here for the __atomic*exchange lowering).

2024-03-08  Jakub Jelinek  <jakub@redhat.com>

gcc/c-family/
	* c-common.cc (resolve_overloaded_atomic_exchange): Instead of setting
	p1 to VIEW_CONVERT_EXPR<I_type> (*p1), set it to MEM_REF with p1 and
	(typeof (p1)) 0 operands and I_type type.
	(resolve_overloaded_atomic_compare_exchange): Similarly for p2.
gcc/cp/
	* pt.cc (tsubst_expr): Handle MEM_REF.
gcc/testsuite/
	* g++.dg/ext/atomic-5.C: New test.

diff --git a/gcc/c-family/c-common.cc b/gcc/c-family/c-common.cc
index e15eff698dfd245a21995af4d5c62339153a5d07..48844b17f77a64805feae60177f4b6765edfc607 100644 (file)
--- a/gcc/c-family/c-common.cc
+++ b/gcc/c-family/c-common.cc
@@ -7793,9 +7793,14 @@ resolve_overloaded_atomic_exchange (location_t loc, tree function,
   /* Convert object pointer to required type.  */
   p0 = build1 (VIEW_CONVERT_EXPR, I_type_ptr, p0);
   (*params)[0] = p0; 
-  /* Convert new value to required type, and dereference it.  */
-  p1 = build_indirect_ref (loc, p1, RO_UNARY_STAR);
-  p1 = build1 (VIEW_CONVERT_EXPR, I_type, p1);
+  /* Convert new value to required type, and dereference it.
+     If *p1 type can have padding or may involve floating point which
+     could e.g. be promoted to wider precision and demoted afterwards,
+     state of padding bits might not be preserved.  */
+  build_indirect_ref (loc, p1, RO_UNARY_STAR);
+  p1 = build2_loc (loc, MEM_REF, I_type,
+                  build1 (VIEW_CONVERT_EXPR, I_type_ptr, p1),
+                  build_zero_cst (TREE_TYPE (p1)));
   (*params)[1] = p1;
 
   /* Move memory model to the 3rd position, and end param list.  */
@@ -7873,9 +7878,14 @@ resolve_overloaded_atomic_compare_exchange (location_t loc, tree function,
   p1 = build1 (VIEW_CONVERT_EXPR, I_type_ptr, p1);
   (*params)[1] = p1;
 
-  /* Convert desired value to required type, and dereference it.  */
-  p2 = build_indirect_ref (loc, p2, RO_UNARY_STAR);
-  p2 = build1 (VIEW_CONVERT_EXPR, I_type, p2);
+  /* Convert desired value to required type, and dereference it.
+     If *p2 type can have padding or may involve floating point which
+     could e.g. be promoted to wider precision and demoted afterwards,
+     state of padding bits might not be preserved.  */
+  build_indirect_ref (loc, p2, RO_UNARY_STAR);
+  p2 = build2_loc (loc, MEM_REF, I_type,
+                  build1 (VIEW_CONVERT_EXPR, I_type_ptr, p2),
+                  build_zero_cst (TREE_TYPE (p2)));
   (*params)[2] = p2;
 
   /* The rest of the parameters are fine. NULL means no special return value

diff --git a/gcc/cp/pt.cc b/gcc/cp/pt.cc
index 514e8029d030a8eed574ca806c38f2f0427ace09..8cf0d5b7a8d83ee87822a796f749911dbdb90a53 100644 (file)
--- a/gcc/cp/pt.cc
+++ b/gcc/cp/pt.cc
@@ -20090,6 +20090,14 @@ tsubst_expr (tree t, tree args, tsubst_flags_t complain, tree in_decl)
        RETURN (r);
       }
 
+    case MEM_REF:
+      {
+       tree op0 = RECUR (TREE_OPERAND (t, 0));
+       tree op1 = RECUR (TREE_OPERAND (t, 0));
+       tree new_type = tsubst (TREE_TYPE (t), args, complain, in_decl);
+       RETURN (build2_loc (EXPR_LOCATION (t), MEM_REF, new_type, op0, op1));
+      }
+
     case NOP_EXPR:
       {
        tree type = tsubst (TREE_TYPE (t), args, complain, in_decl);

diff --git a/gcc/testsuite/g++.dg/ext/atomic-5.C b/gcc/testsuite/g++.dg/ext/atomic-5.C
new file mode 100644 (file)
index 0000000..fa9cda8
--- /dev/null
+++ b/gcc/testsuite/g++.dg/ext/atomic-5.C
@@ -0,0 +1,42 @@
+// { dg-do compile { target c++14 } }
+
+template <int N>
+void
+foo (long double *ptr, long double *val, long double *ret)
+{
+  __atomic_exchange (ptr, val, ret, __ATOMIC_RELAXED);
+}
+
+template <int N>
+bool
+bar (long double *ptr, long double *exp, long double *des)
+{
+  return __atomic_compare_exchange (ptr, exp, des, false,
+                                   __ATOMIC_RELAXED, __ATOMIC_RELAXED);
+}
+
+bool
+baz (long double *p, long double *q, long double *r)
+{
+  foo<0> (p, q, r);
+  foo<1> (p + 1, q + 1, r + 1);
+  return bar<0> (p + 2, q + 2, r + 2) || bar<1> (p + 3, q + 3, r + 3);
+}
+
+constexpr int
+qux (long double *ptr, long double *val, long double *ret)
+{
+  __atomic_exchange (ptr, val, ret, __ATOMIC_RELAXED);
+  return 0;
+}
+
+constexpr bool
+corge (long double *ptr, long double *exp, long double *des)
+{
+  return __atomic_compare_exchange (ptr, exp, des, false,
+                                   __ATOMIC_RELAXED, __ATOMIC_RELAXED);
+}
+
+long double a[6];
+const int b = qux (a, a + 1, a + 2);
+const bool c = corge (a + 3, a + 4, a + 5);