[patch] PR tree-optimization/52633 - ICE due to vectorizer pattern detection collision
Richard Guenther wrote:
On Tue, 24 Apr 2012, Ulrich Weigand wrote:
However, even so, it might actually be preferable to just handle such
cases within vect_recog_widen_shift_pattern itself. Indeed, the routine
already looks for another subsequent type cast, in order to handle
unsigned shift variants. Maybe it simply ought to always look for
another cast, and detect over-widening situations itself?
Does this look reasonable? Any comments or suggestions appreciated!
Yes, getting rid of this fragile interaction by doing more work in
vect_recog_widen_shift_pattern sounds like the correct thing to do.
The following patch implements this. Tested with no regressions on
armv7l-linux-gnueabi and i686-linux-gnu; a couple of vectorizer tests
had to be adapted since now the widening-shift patttern matches
instead of the over-widening pattern (as expected).
Performance testing on ARM showed no regressions (and possible minor
improvements in some cases).
OK for mainline?
Bye,
Ulrich
ChangeLog:
gcc/
PR tree-optimization/52633
* tree-vect-patterns.c (vect_vect_recog_func_ptrs): Swap order of
vect_recog_widen_shift_pattern and vect_recog_over_widening_pattern.
(vect_recog_over_widening_pattern): Remove handling of code that was
already detected as over-widening pattern. Remove special handling
of unsigned cases. Instead, support general case of conversion
of the shift result to another type.
gcc/testsuite/
PR tree-optimization/52633
* gcc.dg/vect/vect-over-widen-1.c: Two patterns should now be
recognized as widening shifts instead of over-widening.
* gcc.dg/vect/vect-over-widen-1-big-array.c: Likewise.
* gcc.dg/vect/vect-over-widen-4.c: Likewise.
* gcc.dg/vect/vect-over-widen-4-big-array.c: Likewise.
* gcc.target/arm/pr52633.c: New test.
Index: gcc-head/gcc/tree-vect-patterns.c
===
--- gcc-head.orig/gcc/tree-vect-patterns.c 2012-05-04 14:23:38.0
+0200
+++ gcc-head/gcc/tree-vect-patterns.c 2012-05-04 14:24:06.0 +0200
@@ -63,8 +63,8 @@ static vect_recog_func_ptr vect_vect_rec
vect_recog_widen_sum_pattern,
vect_recog_dot_prod_pattern,
vect_recog_pow_pattern,
- vect_recog_over_widening_pattern,
vect_recog_widen_shift_pattern,
+ vect_recog_over_widening_pattern,
vect_recog_vector_vector_shift_pattern,
vect_recog_sdivmod_pow2_pattern,
vect_recog_mixed_size_cond_pattern,
@@ -1309,16 +1309,20 @@ vect_recog_over_widening_pattern (VEC (g
where type 'TYPE' is at least double the size of type 'type'.
- Also detect unsigned cases:
+ Also detect cases where the shift result is immediately converted
+ to another type 'result_type' that is no larger in size than 'TYPE'.
+ In those cases we perform a widen-shift that directly results in
+ 'result_type', to avoid a possible over-widening situation:
- unsigned type a_t;
- unsigned TYPE u_res_T;
+ type a_t;
TYPE a_T, res_T;
+ result_type res_result;
S1 a_t = ;
S2 a_T = (TYPE) a_t;
S3 res_T = a_T CONST;
- S4 u_res_T = (unsigned TYPE) res_T;
+ S4 res_result = (result_type) res_T;
+ '-- res_result' = a_t w CONST;
And a case when 'TYPE' is 4 times bigger than 'type'. In that case we
create an additional pattern stmt for S2 to create a variable of an
@@ -1359,60 +1363,21 @@ vect_recog_widen_shift_pattern (VEC (gim
gimple def_stmt0;
tree oprnd0, oprnd1;
tree type, half_type0;
- gimple pattern_stmt, orig_stmt = NULL;
+ gimple pattern_stmt;
tree vectype, vectype_out = NULL_TREE;
tree dummy;
tree var;
enum tree_code dummy_code;
int dummy_int;
VEC (tree, heap) * dummy_vec;
- gimple use_stmt = NULL;
- bool over_widen = false;
+ gimple use_stmt;
bool promotion;
if (!is_gimple_assign (last_stmt) || !vinfo_for_stmt (last_stmt))
return NULL;
- orig_stmt = last_stmt;
if (STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (last_stmt)))
-{
- /* This statement was also detected as over-widening operation (it can't
- be any other pattern, because only over-widening detects shifts).
- LAST_STMT is the final type demotion statement, but its related
- statement is shift. We analyze the related statement to catch cases:
-
- orig code:
- type a_t;
- itype res;
- TYPE a_T, res_T;
-
- S1 a_T = (TYPE) a_t;
- S2 res_T = a_T CONST;
- S3 res = (itype)res_T;
-
- (size of type * 2 = size of itype
- and size of itype * 2 = size of TYPE)
-
- code after over-widening pattern detection:
-
- S1 a_T = (TYPE) a_t;
- -- a_it = (itype) a_t;
- S2 res_T = a_T CONST;
- S3 res = (itype)res_T; --- LAST_STMT
- -- res = a_it CONST;
-
- after
Re: [patch] PR tree-optimization/52633 - ICE due to vectorizer pattern detection collision
On Fri, 4 May 2012, Ulrich Weigand wrote:
Richard Guenther wrote:
On Tue, 24 Apr 2012, Ulrich Weigand wrote:
However, even so, it might actually be preferable to just handle such
cases within vect_recog_widen_shift_pattern itself. Indeed, the routine
already looks for another subsequent type cast, in order to handle
unsigned shift variants. Maybe it simply ought to always look for
another cast, and detect over-widening situations itself?
Does this look reasonable? Any comments or suggestions appreciated!
Yes, getting rid of this fragile interaction by doing more work in
vect_recog_widen_shift_pattern sounds like the correct thing to do.
The following patch implements this. Tested with no regressions on
armv7l-linux-gnueabi and i686-linux-gnu; a couple of vectorizer tests
had to be adapted since now the widening-shift patttern matches
instead of the over-widening pattern (as expected).
Performance testing on ARM showed no regressions (and possible minor
improvements in some cases).
OK for mainline?
Ok.
Thanks,
Richard.
Bye,
Ulrich
ChangeLog:
gcc/
PR tree-optimization/52633
* tree-vect-patterns.c (vect_vect_recog_func_ptrs): Swap order of
vect_recog_widen_shift_pattern and vect_recog_over_widening_pattern.
(vect_recog_over_widening_pattern): Remove handling of code that was
already detected as over-widening pattern. Remove special handling
of unsigned cases. Instead, support general case of conversion
of the shift result to another type.
gcc/testsuite/
PR tree-optimization/52633
* gcc.dg/vect/vect-over-widen-1.c: Two patterns should now be
recognized as widening shifts instead of over-widening.
* gcc.dg/vect/vect-over-widen-1-big-array.c: Likewise.
* gcc.dg/vect/vect-over-widen-4.c: Likewise.
* gcc.dg/vect/vect-over-widen-4-big-array.c: Likewise.
* gcc.target/arm/pr52633.c: New test.
Index: gcc-head/gcc/tree-vect-patterns.c
===
--- gcc-head.orig/gcc/tree-vect-patterns.c2012-05-04 14:23:38.0
+0200
+++ gcc-head/gcc/tree-vect-patterns.c 2012-05-04 14:24:06.0 +0200
@@ -63,8 +63,8 @@ static vect_recog_func_ptr vect_vect_rec
vect_recog_widen_sum_pattern,
vect_recog_dot_prod_pattern,
vect_recog_pow_pattern,
- vect_recog_over_widening_pattern,
vect_recog_widen_shift_pattern,
+ vect_recog_over_widening_pattern,
vect_recog_vector_vector_shift_pattern,
vect_recog_sdivmod_pow2_pattern,
vect_recog_mixed_size_cond_pattern,
@@ -1309,16 +1309,20 @@ vect_recog_over_widening_pattern (VEC (g
where type 'TYPE' is at least double the size of type 'type'.
- Also detect unsigned cases:
+ Also detect cases where the shift result is immediately converted
+ to another type 'result_type' that is no larger in size than 'TYPE'.
+ In those cases we perform a widen-shift that directly results in
+ 'result_type', to avoid a possible over-widening situation:
- unsigned type a_t;
- unsigned TYPE u_res_T;
+ type a_t;
TYPE a_T, res_T;
+ result_type res_result;
S1 a_t = ;
S2 a_T = (TYPE) a_t;
S3 res_T = a_T CONST;
- S4 u_res_T = (unsigned TYPE) res_T;
+ S4 res_result = (result_type) res_T;
+ '-- res_result' = a_t w CONST;
And a case when 'TYPE' is 4 times bigger than 'type'. In that case we
create an additional pattern stmt for S2 to create a variable of an
@@ -1359,60 +1363,21 @@ vect_recog_widen_shift_pattern (VEC (gim
gimple def_stmt0;
tree oprnd0, oprnd1;
tree type, half_type0;
- gimple pattern_stmt, orig_stmt = NULL;
+ gimple pattern_stmt;
tree vectype, vectype_out = NULL_TREE;
tree dummy;
tree var;
enum tree_code dummy_code;
int dummy_int;
VEC (tree, heap) * dummy_vec;
- gimple use_stmt = NULL;
- bool over_widen = false;
+ gimple use_stmt;
bool promotion;
if (!is_gimple_assign (last_stmt) || !vinfo_for_stmt (last_stmt))
return NULL;
- orig_stmt = last_stmt;
if (STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (last_stmt)))
-{
- /* This statement was also detected as over-widening operation (it
can't
- be any other pattern, because only over-widening detects shifts).
- LAST_STMT is the final type demotion statement, but its related
- statement is shift. We analyze the related statement to catch
cases:
-
- orig code:
- type a_t;
- itype res;
- TYPE a_T, res_T;
-
- S1 a_T = (TYPE) a_t;
- S2 res_T = a_T CONST;
- S3 res = (itype)res_T;
-
- (size of type * 2 = size of itype
- and size of itype * 2 = size of TYPE)
-
- code after over-widening pattern detection:
-
- S1 a_T = (TYPE) a_t;
- -- a_it = (itype) a_t;
