[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #25 from jamborm at gcc dot gnu dot org 2010-07-21 13:57 --- Subject: Bug 44900 Author: jamborm Date: Wed Jul 21 13:57:12 2010 New Revision: 162374 URL: http://gcc.gnu.org/viewcvs?root=gccview=revrev=162374 Log: 2010-07-21 Martin Jambor mjam...@suse.cz PR tree-optimization/44900 * tree-sra.c (load_assign_lhs_subreplacements): Updated comments. (sra_modify_assign): Move gsi to the next statmenent unconditionally. * testsuite/g++.dg/torture/pr44900.C: New test. Added: branches/gcc-4_5-branch/gcc/testsuite/g++.dg/torture/pr44900.C Modified: branches/gcc-4_5-branch/gcc/ChangeLog branches/gcc-4_5-branch/gcc/testsuite/ChangeLog branches/gcc-4_5-branch/gcc/tree-sra.c -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #26 from jamborm at gcc dot gnu dot org 2010-07-21 14:17 --- Subject: Bug 44900 Author: jamborm Date: Wed Jul 21 14:17:11 2010 New Revision: 162375 URL: http://gcc.gnu.org/viewcvs?root=gccview=revrev=162375 Log: 2010-07-21 Martin Jambor mjam...@suse.cz PR tree-optimization/44900 * tree-sra.c (load_assign_lhs_subreplacements): Updated comments. (sra_modify_assign): Move gsi to the next statmenent unconditionally. * testsuite/g++.dg/torture/pr44900.C: New test. Added: trunk/gcc/testsuite/g++.dg/torture/pr44900.C Modified: trunk/gcc/ChangeLog trunk/gcc/testsuite/ChangeLog trunk/gcc/tree-sra.c -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #27 from jamborm at gcc dot gnu dot org 2010-07-21 17:21 --- Fixed. -- jamborm at gcc dot gnu dot org changed: What|Removed |Added Status|ASSIGNED|RESOLVED Resolution||FIXED http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #24 from jamborm at gcc dot gnu dot org 2010-07-20 13:34 --- I posted a proposed fix to the mailing list: http://gcc.gnu.org/ml/gcc-patches/2010-07/msg01570.html -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
-- jamborm at gcc dot gnu dot org changed: What|Removed |Added AssignedTo|unassigned at gcc dot gnu |jamborm at gcc dot gnu dot |dot org |org Status|NEW |ASSIGNED Last reconfirmed|2010-07-11 11:25:40 |2010-07-19 13:45:37 date|| http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #15 from yottui at yahoo dot co dot jp 2010-07-14 09:22 ---
I found the similar case with gcc 4.4.4 of MacPorts and gcc 4.4.0 of MinGW.
-- begin testcase --
// g++ -O -msse2 test.cpp
typedef long long __m128i __attribute__ ((__vector_size__ (16),
__may_alias__));
struct vec
{
__m128i v;
static vec load(const int * p)
{ return (__m128i) __builtin_ia32_loaddqu((char const *)p); }
const int operator [](int i) const
{
union u {
__m128i v;
int e[4];
};
return ((const u )v).e[i];
}
vec() {}
vec(const __m128i a) : v(a) {}
};
extern C {
int printf (const char*, ...);
}
int main( int argc, char * argv[] )
{
__attribute__((aligned(16))) int data[] =
{ 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5 };
vec a = vec::load(data);
printf(v: %d, %d, %d, %d\n, a[0], a[1], a[2], a[3]);
return 0;
}
-- end testcase --
-- begin output --
v: 16, 16, 14, 14
-- end output --
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #16 from yottui at yahoo dot co dot jp 2010-07-14 12:24 ---
This is also the wrong result with MinGW gcc 3.4.5.
I'm expecting that all component of v will be 2500.
4.4.4 of MacPorts, 4.5.0 of MacPorts, 4.4.0 of MinGW and 4.5.0-1 of MinGW were
worked fine.
This time, I did not check 3.4.6 of MacPorts.
I tried to install gcc34 of MacPorts, but failed.
-- begin testcase --
// g++ -O -msse2 test.cpp
#include xmmintrin.h
#include emmintrin.h
extern C int printf (const char*, ...);
// There is no _mm_castps_si128() in gcc 3.4
inline __m128i my_castps_si128(const __m128 a)
{
//return *(const __m128i *)a; // same result
union u {
__m128 s;
__m128i i;
};
const u v = (const u )a;
return v.i;
}
inline __m128 my_castsi128_ps(const __m128i a)
{
//return *(const __m128 *)a; // same result
union u {
__m128 s;
__m128i i;
};
const u v = (const u )a;
return v.s;
}
int main( int argc, char * argv[] )
{
union u {
__m128i v;
int e[4];
};
__m128i a = _mm_set1_epi32(1250);
__m128i b = _mm_set1_epi32(2);
__m128i v0 = _mm_setzero_si128();
__m128i al = _mm_unpacklo_epi32(a, v0);
__m128i ah = _mm_unpackhi_epi32(a, v0);
__m128i bl = _mm_unpacklo_epi32(b, v0);
__m128i bh = _mm_unpackhi_epi32(b, v0);
__m128i lo = _mm_mul_epu32(al, bl);
__m128i hi = _mm_mul_epu32(ah, bh);
__m128 sl = my_castsi128_ps(lo);
__m128 sh = my_castsi128_ps(hi);
__m128 s = _mm_shuffle_ps(sl, sh, 0x88); // 2, 0, 2, 0
__m128i r = my_castps_si128(s);
u v = (u )r;
printf(v: %d, %d, %d, %d\n, v.e[0], v.e[1], v.e[2], v.e[3]);
return 0;
}
-- end testcase --
-- begin output --
v: 0, 0, 2500, 2500
-- end output --
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #17 from sezeroz at gmail dot com 2010-07-14 14:02 ---
(In reply to comment #16)
This is also the wrong result with MinGW gcc 3.4.5.
I'm expecting that all component of v will be 2500.
4.4.4 of MacPorts, 4.5.0 of MacPorts, 4.4.0 of MinGW and 4.5.0-1 of MinGW were
worked fine.
This time, I did not check 3.4.6 of MacPorts.
I tried to install gcc34 of MacPorts, but failed.
-- begin testcase --
// g++ -O -msse2 test.cpp
#include xmmintrin.h
#include emmintrin.h
extern C int printf (const char*, ...);
// There is no _mm_castps_si128() in gcc 3.4
inline __m128i my_castps_si128(const __m128 a)
{
//return *(const __m128i *)a; // same result
union u {
__m128 s;
__m128i i;
};
const u v = (const u )a;
return v.i;
}
inline __m128 my_castsi128_ps(const __m128i a)
{
//return *(const __m128 *)a; // same result
union u {
__m128 s;
__m128i i;
};
const u v = (const u )a;
return v.s;
}
int main( int argc, char * argv[] )
{
union u {
__m128i v;
int e[4];
};
__m128i a = _mm_set1_epi32(1250);
__m128i b = _mm_set1_epi32(2);
__m128i v0 = _mm_setzero_si128();
__m128i al = _mm_unpacklo_epi32(a, v0);
__m128i ah = _mm_unpackhi_epi32(a, v0);
__m128i bl = _mm_unpacklo_epi32(b, v0);
__m128i bh = _mm_unpackhi_epi32(b, v0);
__m128i lo = _mm_mul_epu32(al, bl);
__m128i hi = _mm_mul_epu32(ah, bh);
__m128 sl = my_castsi128_ps(lo);
__m128 sh = my_castsi128_ps(hi);
__m128 s = _mm_shuffle_ps(sl, sh, 0x88); // 2, 0, 2, 0
__m128i r = my_castps_si128(s);
u v = (u )r;
printf(v: %d, %d, %d, %d\n, v.e[0], v.e[1], v.e[2], v.e[3]);
return 0;
}
-- end testcase --
-- begin output --
v: 0, 0, 2500, 2500
-- end output --
--
sezeroz at gmail dot com changed:
What|Removed |Added
CC||sezeroz at gmail dot com
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #18 from sezeroz at gmail dot com 2010-07-14 14:05 --- (In reply to comment #15) I found the similar case with gcc 4.4.4 of MacPorts and gcc 4.4.0 of MinGW. This case fails with 4.4 on i686-linux too, printing 16, 16, 14, 14 instead of 16, 15, 14, 13 which 4.3 does. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #19 from hjl dot tools at gmail dot com 2010-07-14 15:52
---
(In reply to comment #15)
I found the similar case with gcc 4.4.4 of MacPorts and gcc 4.4.0 of MinGW.
-- begin testcase --
// g++ -O -msse2 test.cpp
typedef long long __m128i __attribute__ ((__vector_size__ (16),
__may_alias__));
struct vec
{
__m128i v;
static vec load(const int * p)
{ return (__m128i) __builtin_ia32_loaddqu((char const *)p); }
const int operator [](int i) const
{
union u {
__m128i v;
int e[4];
};
return ((const u )v).e[i];
}
vec() {}
vec(const __m128i a) : v(a) {}
};
extern C {
int printf (const char*, ...);
}
int main( int argc, char * argv[] )
{
__attribute__((aligned(16))) int data[] =
{ 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5 };
vec a = vec::load(data);
printf(v: %d, %d, %d, %d\n, a[0], a[1], a[2], a[3]);
return 0;
}
-- end testcase --
-- begin output --
v: 16, 16, 14, 14
-- end output --
This is caused by revision 134947:
http://gcc.gnu.org/ml/gcc-cvs/2008-05/msg00107.html
--
hjl dot tools at gmail dot com changed:
What|Removed |Added
CC||pinskia at gcc dot gnu dot
||org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #20 from pinskia at gcc dot gnu dot org 2010-07-14 16:41 --- (In reply to comment #15) I found the similar case with gcc 4.4.4 of MacPorts and gcc 4.4.0 of MinGW. I think the code in comment #15 is invalid and voilates C/C++ aliasing rules. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #21 from pinskia at gcc dot gnu dot org 2010-07-14 16:44 --- (In reply to comment #20) (In reply to comment #15) I found the similar case with gcc 4.4.4 of MacPorts and gcc 4.4.0 of MinGW. I think the code in comment #15 is invalid and voilates C/C++ aliasing rules. Even if it did not voilate aliasing rules, the IR looks good: D.4999_70 = VIEW_CONVERT_EXPRconst union u(D.4995_68).i; D.4863_25 = VIEW_CONVERT_EXPRunion u(D.4999_70).e[3]; D.4864_26 = VIEW_CONVERT_EXPRunion u(D.4999_70).e[2]; D.4865_27 = VIEW_CONVERT_EXPRunion u(D.4999_70).e[1]; D.4866_28 = VIEW_CONVERT_EXPRunion u(D.4999_70).e[0]; -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #22 from pinskia at gcc dot gnu dot org 2010-07-14 16:50 --- (In reply to comment #21) Even if it did not voilate aliasing rules, the IR looks good: The expansion looks incorrect though: (insn 39 38 40 t.cc:55 (set (reg:DI 107) (vec_select:DI (reg:V2DI 79 [ D.4999 ]) (parallel [ (const_int 1 [0x1]) ]))) -1 (nil)) (insn 40 39 41 t.cc:55 (set (reg:DI 108) (vec_select:DI (reg:V2DI 79 [ D.4999 ]) (parallel [ (const_int 1 [0x1]) ]))) -1 (nil)) (insn 41 40 42 t.cc:55 (set (reg:DI 109) (vec_select:DI (reg:V2DI 79 [ D.4999 ]) (parallel [ (const_int 0 [0x0]) ]))) -1 (nil)) (insn 42 41 43 t.cc:55 (set (reg:DI 110) (vec_select:DI (reg:V2DI 79 [ D.4999 ]) (parallel [ (const_int 0 [0x0]) ]))) -1 (nil)) (insn 43 42 44 t.cc:55 (set (reg:SI 37 r8) (subreg:SI (reg:DI 107) 0)) -1 (nil)) (insn 44 43 45 t.cc:55 (set (reg:SI 2 cx) (subreg:SI (reg:DI 108) 0)) -1 (nil)) (insn 45 44 46 t.cc:55 (set (reg:SI 1 dx) (subreg:SI (reg:DI 109) 0)) -1 (nil)) (insn 46 45 47 t.cc:55 (set (reg:SI 4 si) (subreg:SI (reg:DI 110) 0)) -1 (nil)) Please file that as a different bug. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #23 from yottui at yahoo dot co dot jp 2010-07-15 03:45 --- (In reply to comment #22) Please file that as a different bug. May I enter comment #15 as a new bug to Bugzilla? -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #14 from rguenth at gcc dot gnu dot org 2010-07-12 07:54 --- It indeed works with the 4.5.0 release which makes it a blocker for 4.5.1. -- rguenth at gcc dot gnu dot org changed: What|Removed |Added Priority|P3 |P1 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #6 from rguenth at gcc dot gnu dot org 2010-07-11 12:48 --- Confirmed, caused by SRA, maybe latent on the trunk. -- rguenth at gcc dot gnu dot org changed: What|Removed |Added CC||jamborm at gcc dot gnu dot ||org Component|target |tree-optimization Keywords||wrong-code Known to fail|4.5.1 |4.5.0 4.5.1 Target Milestone|--- |4.5.1 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #7 from hjl dot tools at gmail dot com 2010-07-11 20:59 --- It works with gcc 4.5.0. It is a 4.5.1 regression. -- hjl dot tools at gmail dot com changed: What|Removed |Added Known to fail|4.5.0 4.5.1 |4.5.1 Known to work|4.6.0 4.4.4 |4.6.0 4.4.4 4.5.0 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #8 from rguenth at gcc dot gnu dot org 2010-07-11 21:59 --- Hm? I checked 4.5.0 and it was broken as well, so someone should double-check (I can't at the moment). -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #9 from pinskia at gcc dot gnu dot org 2010-07-11 22:47 --- Works with: GNU C++ (GCC) version 4.5.0 20100401 (experimental) [trunk revision 157933] (x86_64-unknown-linux-gnu) compiled by GNU C version 4.5.0 20100401 (experimental) [trunk revision 157933], GMP version 4.3.2, MPFR version 2.4.2, MPC version 0.8.1 Which is before the branch of 4.5. -- pinskia at gcc dot gnu dot org changed: What|Removed |Added Known to fail|4.5.1 | Known to work|4.6.0 4.4.4 4.5.0 |4.6.0 4.4.4 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #10 from yottui at yahoo dot co dot jp 2010-07-11 23:11 --- Please use '-m32' if host is x64. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #11 from hjl dot tools at gmail dot com 2010-07-11 23:35 --- It is very likely caused by revision 158826: http://gcc.gnu.org/ml/gcc-cvs/2010-04/msg00933.html I will verify it and find out which checkin on trunk fixes/hides this bug. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #12 from hjl dot tools at gmail dot com 2010-07-12 01:05 --- It is caused by revision 158826. On trunk, it is fixed/hidden by revision 158732: http://gcc.gnu.org/ml/gcc-cvs/2010-04/msg00839.html -- hjl dot tools at gmail dot com changed: What|Removed |Added CC||hubicka at gcc dot gnu dot ||org Known to fail||4.5.1 Known to work|4.6.0 4.4.4 |4.6.0 4.4.4 4.5.0 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
[Bug tree-optimization/44900] [4.5 Regression] The variable of SSE will be broken
--- Comment #13 from hjl dot tools at gmail dot com 2010-07-12 02:03 --- If you replace --- vec2 a; a = vec2::load(p); --- with --- vec2 a = vec2::load(p); --- the testcase will pass. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44900
