------- Comment #42 from rguenth at gcc dot gnu dot org 2007-05-04 11:45
-------
I put it on the tester. A testcase that does regress (simplified from
tramp3d):
template <class T, int D>
class Vector
{
public:
Vector()
{
for (int i = 0; i < D; ++i)
new (&x_m[i]) T();
}
T& operator[](int i) { return x_m[i]; }
private:
T x_m[D];
};
void foo(Vector<double, 3> *m)
{
Vector<double, 3> v;
v[0] = 1.0;
v[1] = 2.0;
v[3] = 3.0;
*m = v;
}
where we can no longer optimize the redundant stores to v:
<bb 2>:
v.x_m[0] = 0.0;
__asm__ __volatile__("":"=m" v.x_m[0]:"m" v.x_m[0]);
v.x_m[1] = 0.0;
__asm__ __volatile__("":"=m" v.x_m[1]:"m" v.x_m[1]);
v.x_m[2] = 0.0;
__asm__ __volatile__("":"=m" v.x_m[2]:"m" v.x_m[2]);
v.x_m[0] = 1.0e+0;
v.x_m[1] = 2.0e+0;
v.x_m[3] = 3.0e+0;
*m = v;
return;
vs.
<bb 2>:
v.x_m[2] = 0.0;
v.x_m[0] = 1.0e+0;
v.x_m[1] = 2.0e+0;
v.x_m[3] = 3.0e+0;
*m = v;
return;
and assembly:
_Z3fooP6VectorIdLi3EE:
.LFB17:
xorl %r9d, %r9d
movq %r9, -40(%rsp)
movq %r9, -32(%rsp)
movq %r9, -24(%rsp)
movabsq $4607182418800017408, %r8
movabsq $4611686018427387904, %rsi
movq -24(%rsp), %rax
movq %r8, -40(%rsp)
movq %rsi, -32(%rsp)
movq -40(%rsp), %rcx
movq -32(%rsp), %rdx
movq %rax, 16(%rdi)
movq %rcx, (%rdi)
movq %rdx, 8(%rdi)
ret
vs.
_Z3fooP6VectorIdLi3EE:
.LFB17:
movabsq $4607182418800017408, %r8
movabsq $4611686018427387904, %rsi
movq $0, -24(%rsp)
movq %r8, -40(%rsp)
movq %rsi, -32(%rsp)
movq -40(%rsp), %rcx
movq -32(%rsp), %rdx
movq -24(%rsp), %rax
movq %rcx, (%rdi)
movq %rdx, 8(%rdi)
movq %rax, 16(%rdi)
ret
this happens in hot tramp3d loops and is a quite common idiom for initializing
storage. To fix this we need to avoid creating the asm if the type of the
original storage is the same as the other.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29286
