The following short testcase gets vectorized with 4.1.1 and doesn't with 4.2.0
revision 114610
============================================================
template <class T>
class vec
{
public:
vec(unsigned int n) : size_(n)
{
data_ = new T[n];
}
vec& multiply(const vec& other)
{
const T* op=other.data_;
for (unsigned int i=0; i<size_; ++i) {
data_[i] *= op[i];
}
return *this;
}
private:
unsigned int size_;
T* data_;
};
template class vec<float>;
============================================================
/usr/local/4.2/bin/g++4.2.0 -O3 -ftree-vectorize -ftree-vectorizer-verbose=7
-march=pentium-m -c vectorizer.cpp
vectorizer.cpp:16: note: ===== analyze_loop_nest =====
vectorizer.cpp:16: note: === vect_analyze_loop_form ===
vectorizer.cpp:16: note: split exit edge.
vectorizer.cpp:16: note: === get_loop_niters ===
vectorizer.cpp:16: note: ==> get_loop_niters:D.2376_16
vectorizer.cpp:16: note: Symbolic number of iterations is D.2376_16
vectorizer.cpp:16: note: === vect_analyze_data_refs ===
vectorizer.cpp:16: note: get vectype with 4 units of type float
vectorizer.cpp:16: note: vectype: vector float
vectorizer.cpp:16: note: get vectype with 4 units of type const float
vectorizer.cpp:16: note: vectype: const vector float
vectorizer.cpp:16: note: get vectype with 4 units of type float
vectorizer.cpp:16: note: vectype: vector float
vectorizer.cpp:16: note: === vect_analyze_scalar_cycles ===
vectorizer.cpp:16: note: Analyze phi: SMT.6_28 = PHI <SMT.6_27(5),
SMT.6_26(3)>;
vectorizer.cpp:16: note: virtual phi. skip.
vectorizer.cpp:16: note: Analyze phi: i_4 = PHI <i_23(5), 0(3)>;
vectorizer.cpp:16: note: Access function of PHI: {0, +, 1}_1
vectorizer.cpp:16: note: step: 1, init: 0
vectorizer.cpp:16: note: Detected induction.
vectorizer.cpp:16: note: === vect_pattern_recog ===
vectorizer.cpp:16: note: === vect_mark_stmts_to_be_vectorized ===
vectorizer.cpp:16: note: init: phi relevant? SMT.6_28 = PHI <SMT.6_27(5),
SMT.6_26(3)>;
vectorizer.cpp:16: note: init: phi relevant? i_4 = PHI <i_23(5), 0(3)>;
vectorizer.cpp:16: note: init: stmt relevant? <L0>:
vectorizer.cpp:16: note: init: stmt relevant? D.2378_9 = pretmp.24_1
vectorizer.cpp:16: note: init: stmt relevant? D.2379_10 = i_4 * 4
vectorizer.cpp:16: note: init: stmt relevant? D.2380_11 = (float *) D.2379_10
vectorizer.cpp:16: note: init: stmt relevant? D.2381_12 = pretmp.24_1 +
D.2380_11
vectorizer.cpp:16: note: init: stmt relevant? D.2382_17 = *D.2381_12
vectorizer.cpp:16: note: init: stmt relevant? D.2383_19 = (const float *)
D.2379_10
vectorizer.cpp:16: note: init: stmt relevant? D.2384_20 = D.2383_19 + op_3
vectorizer.cpp:16: note: init: stmt relevant? D.2385_21 = *D.2384_20
vectorizer.cpp:16: note: init: stmt relevant? D.2386_22 = D.2382_17 * D.2385_21
vectorizer.cpp:16: note: init: stmt relevant? *D.2381_12 = D.2386_22
vectorizer.cpp:16: note: vec_stmt_relevant_p: stmt has vdefs.
vectorizer.cpp:16: note: mark relevant 1, live 0.
vectorizer.cpp:16: note: init: stmt relevant? i_23 = i_4 + 1
vectorizer.cpp:16: note: init: stmt relevant? if (D.2376_16 > i_23) goto <L9>;
else goto <L12>;
vectorizer.cpp:16: note: init: stmt relevant? <L9>:
vectorizer.cpp:16: note: worklist: examine stmt: *D.2381_12 = D.2386_22
vectorizer.cpp:16: note: vect_is_simple_use: operand D.2386_22
vectorizer.cpp:16: note: def_stmt: D.2386_22 = D.2382_17 * D.2385_21
vectorizer.cpp:16: note: type of def: 2.
vectorizer.cpp:16: note: worklist: examine use 2: D.2386_22
vectorizer.cpp:16: note: mark relevant 1, live 0.
vectorizer.cpp:16: note: worklist: examine stmt: D.2386_22 = D.2382_17 *
D.2385_21
vectorizer.cpp:16: note: vect_is_simple_use: operand D.2382_17
vectorizer.cpp:16: note: def_stmt: D.2382_17 = *D.2381_12
vectorizer.cpp:16: note: type of def: 2.
vectorizer.cpp:16: note: worklist: examine use 2: D.2382_17
vectorizer.cpp:16: note: mark relevant 1, live 0.
vectorizer.cpp:16: note: vect_is_simple_use: operand D.2385_21
vectorizer.cpp:16: note: def_stmt: D.2385_21 = *D.2384_20
vectorizer.cpp:16: note: type of def: 2.
vectorizer.cpp:16: note: worklist: examine use 2: D.2385_21
vectorizer.cpp:16: note: mark relevant 1, live 0.
vectorizer.cpp:16: note: worklist: examine stmt: D.2385_21 = *D.2384_20
vectorizer.cpp:16: note: worklist: examine stmt: D.2382_17 = *D.2381_12
vectorizer.cpp:16: note: === vect_analyze_data_refs_alignment ===
vectorizer.cpp:16: note: vect_compute_data_ref_alignment:
vectorizer.cpp:16: note: Unknown alignment for access: *pretmp.24_1
vectorizer.cpp:16: note: vect_compute_data_ref_alignment:
vectorizer.cpp:16: note: Unknown alignment for access: *op_3
vectorizer.cpp:16: note: vect_compute_data_ref_alignment:
vectorizer.cpp:16: note: Unknown alignment for access: *pretmp.24_1
vectorizer.cpp:16: note: === vect_determine_vectorization_factor ===
vectorizer.cpp:16: note: ==> examining statement: <L0>:
vectorizer.cpp:16: note: skip.
vectorizer.cpp:16: note: ==> examining statement: D.2378_9 = pretmp.24_1
vectorizer.cpp:16: note: skip.
vectorizer.cpp:16: note: ==> examining statement: D.2379_10 = i_4 * 4
vectorizer.cpp:16: note: skip.
vectorizer.cpp:16: note: ==> examining statement: D.2380_11 = (float *)
D.2379_10
vectorizer.cpp:16: note: skip.
vectorizer.cpp:16: note: ==> examining statement: D.2381_12 = pretmp.24_1 +
D.2380_11
vectorizer.cpp:16: note: skip.
vectorizer.cpp:16: note: ==> examining statement: D.2382_17 = *D.2381_12
vectorizer.cpp:16: note: vectype: vector float
vectorizer.cpp:16: note: nunits = 4
vectorizer.cpp:16: note: ==> examining statement: D.2383_19 = (const float *)
D.2379_10
vectorizer.cpp:16: note: skip.
vectorizer.cpp:16: note: ==> examining statement: D.2384_20 = D.2383_19 + op_3
vectorizer.cpp:16: note: skip.
vectorizer.cpp:16: note: ==> examining statement: D.2385_21 = *D.2384_20
vectorizer.cpp:16: note: vectype: const vector float
vectorizer.cpp:16: note: nunits = 4
vectorizer.cpp:16: note: ==> examining statement: D.2386_22 = D.2382_17 *
D.2385_21
vectorizer.cpp:16: note: get vectype for scalar type: float
vectorizer.cpp:16: note: get vectype with 4 units of type float
vectorizer.cpp:16: note: vectype: vector float
vectorizer.cpp:16: note: vectype: vector float
vectorizer.cpp:16: note: nunits = 4
vectorizer.cpp:16: note: ==> examining statement: *D.2381_12 = D.2386_22
vectorizer.cpp:16: note: vectype: vector float
vectorizer.cpp:16: note: nunits = 4
vectorizer.cpp:16: note: ==> examining statement: i_23 = i_4 + 1
vectorizer.cpp:16: note: skip.
vectorizer.cpp:16: note: ==> examining statement: if (D.2376_16 > i_23) goto
<L9>; else goto <L12>;
vectorizer.cpp:16: note: skip.
vectorizer.cpp:16: note: ==> examining statement: <L9>:
vectorizer.cpp:16: note: skip.
vectorizer.cpp:16: note: === vect_analyze_dependences ===
vectorizer.cpp:16: note: dependence distance = 0.
vectorizer.cpp:16: note: accesses have the same alignment.
vectorizer.cpp:16: note: dependence distance modulo vf == 0 between *D.2381_12
and *D.2381_12
vectorizer.cpp:16: note: not vectorized: can't determine dependence between
*D.2384_20 and *D.2381_12
vectorizer.cpp:16: note: bad data dependence.
vectorizer.cpp:16: note: vectorized 0 loops in function.
The workaround with "op" is not needed with the current autovect-branch BTW.
--
Summary: [4.2 regression] missed optimization with -ftree-
vectorize
Product: gcc
Version: tree-ssa
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: gcc at pdoerfler dot com
GCC build triplet: i686-pc-linux-gnu
GCC host triplet: i686-pc-linux-gnu
GCC target triplet: i686-pc-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28029
