> To commit 2), I would like you to run a wider set of tests (e.g., the LLVM
> test suite). If this passes successful, we should give a headsup on the GCC
> mailing list and ask other people to try the new isl support.
> If now bugs have found, we switch.
I've tested the modified version of Graphite using the LLVM test suite
and found out regression only in two test cases:
/llvm-test-suite/MultiSource/Benchmarks/ASC_Sequoia/CrystalMk/Crystal_Cholesky.c,
/llvm-test suite/MultiSource/Benchmarks/Bullet
Compilation of
/llvm-test-suite/MultiSource/Benchmarks/ASC_Sequoia/CrystalMk/Crystal_Cholesky.c
produces the following error message:
/home/roman/llvm-test-suite/MultiSource/Benchmarks/ASC_Sequoia/CrystalMk/Crystal_Cholesky.c
/home/roman/llvm-test-suite/MultiSource/Benchmarks/Bullet
Crystal_Cholesky.c: In function ‘Crystal_Cholesky’:
Crystal_Cholesky.c:16:6: internal compiler error: in operator[], at vec.h:736
void Crystal_Cholesky(int nSlip,
^
0xf1d080 vec<tree_node*, va_heap, vl_embed>::operator[](unsigned int)
/home/roman/sec_trunk/gcc/gcc/vec.h:736
0xf1d080 vec<tree_node*, va_heap, vl_ptr>::operator[](unsigned int)
/home/roman/sec_trunk/gcc/gcc/vec.h:1202
0xf1d080 build_iv_mapping
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:608
0xf1d080 translate_isl_ast_node_user
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:644
0xf1d080 translate_isl_ast
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:734
0xf1c7d0 translate_isl_ast_for_loop
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:441
0xf1c7d0 translate_isl_ast_node_for
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:580
0xf1c7d0 translate_isl_ast
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:727
0xf1c3e0 translate_isl_ast_node_block
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:669
0xf1c3e0 translate_isl_ast
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:738
0xf1c3e0 translate_isl_ast_node_block
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:669
0xf1c3e0 translate_isl_ast
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:738
0xf1c3e0 translate_isl_ast_node_block
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:669
0xf1c3e0 translate_isl_ast
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:738
0xf1c3e0 translate_isl_ast_node_block
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:669
0xf1c3e0 translate_isl_ast
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:738
0xf1c3e0 translate_isl_ast_node_block
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:669
0xf1c3e0 translate_isl_ast
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:738
0xf1c7d0 translate_isl_ast_for_loop
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:441
0xf1c7d0 translate_isl_ast_node_for
/home/roman/sec_trunk/gcc/gcc/graphite-isl-ast-to-gimple.c:580
It is caused by the wrong index of the iv_map (its value is 12), which
is greater than its size (its value is equal to loop->num = 8 plus
one). I think that we should use the value of number_of_loops for the
size of the iv_map (the attached patch implements this), because it is
proven to be robust. What do you think about this?
The second error is produced by the executable, which is successfully
compiled by /llvm-test-suite/MultiSource/Benchmarks/Bullet:
“Segmentation fault (core dumped)”. I've used gdb to print stack
trace:
Segmentation fault (core dumped)
ore was generated by `./bullet.simple'.
Program terminated with signal 11, Segmentation fault.
#0 upcast (colObj=0x0)
at
/home/roman/llvm-test-suite/MultiSource/Benchmarks/Bullet/include/BulletDynamics/Dynamics/btRigidBody.h:170
170 if (colObj->getInternalType()==btCollisionObject::CO_RIGID_BODY)
(gdb) bt
#0 upcast (colObj=0x0)
at
/home/roman/llvm-test-suite/MultiSource/Benchmarks/Bullet/include/BulletDynamics/Dynamics/btRigidBody.h:170
#1 saveKinematicState (timeStep=<optimized out>, this=<optimized out>)
at
/home/roman/llvm-test-suite/MultiSource/Benchmarks/Bullet/btDiscreteDynamicsWorld.cpp:113
#2 btDiscreteDynamicsWorld::stepSimulation (this=0x14a6ca0,
timeStep=0.0166666675, maxSubSteps=1, fixedTimeStep=0.0166666675)
at
/home/roman/llvm-test-suite/MultiSource/Benchmarks/Bullet/btDiscreteDynamicsWorld.cpp:333
#3 0x0000000000401ed8 in BenchmarkDemo::clientMoveAndDisplay (
this=0x7fff90475840)
at
/home/roman/llvm-test-suite/MultiSource/Benchmarks/Bullet/BenchmarkDemo.cpp:232
#4 0x0000000000401bd2 in main (argc=<optimized out>, argv=<optimized out>)
at /home/roman/llvm-test-suite/MultiSource/Benchmarks/Bullet/main.cpp:63
The executable is produced by compilation of 317 files. That's why I
think that it is be better to temporary postpone the consideration.
What do you think about this?
--
Cheers, Roman Gareev.
2014-08-07 Roman Gareev <gareevro...@gmail.com>
[gcc/]
* graphite-isl-ast-to-gimple.c:
(translate_isl_ast_node_user): Use nb_loops instead of loop->num + 1.
[gcc/testsuite]
* gcc.dg/graphite/isl-ast-gen-user-1.c: New testcase.
Index: gcc/graphite-isl-ast-to-gimple.c
===================================================================
--- gcc/graphite-isl-ast-to-gimple.c (revision 213690)
+++ gcc/graphite-isl-ast-to-gimple.c (working copy)
@@ -632,9 +632,9 @@
gcc_assert (GBB_BB (gbb) != ENTRY_BLOCK_PTR_FOR_FN (cfun) &&
"The entry block should not even appear within a scop");
- loop_p loop = gbb_loop (gbb);
- iv_map.create (loop->num + 1);
- iv_map.safe_grow_cleared (loop->num + 1);
+ int nb_loops = number_of_loops (cfun);
+ iv_map.create (nb_loops);
+ iv_map.safe_grow_cleared (nb_loops);
build_iv_mapping (iv_map, gbb, user_expr, ip, SCOP_REGION (pbb->scop));
isl_ast_expr_free (user_expr);
Index: gcc/testsuite/gcc.dg/graphite/isl-ast-gen-user-1.c
===================================================================
--- gcc/testsuite/gcc.dg/graphite/isl-ast-gen-user-1.c (revision 0)
+++ gcc/testsuite/gcc.dg/graphite/isl-ast-gen-user-1.c (working copy)
@@ -0,0 +1,63 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fgraphite-identity -fgraphite-code-generator=isl" } */
+
+#include <stdio.h>
+#include <stdlib.h>
+
+static const int MS_XTAL_NSLIP_MAX = 12;
+
+//--------------
+// test Cholesky solver on matrix
+//--------------
+void Crystal_Cholesky(int nSlip,
+ double a[MS_XTAL_NSLIP_MAX][MS_XTAL_NSLIP_MAX],
+ double r[MS_XTAL_NSLIP_MAX],
+ double g[MS_XTAL_NSLIP_MAX])
+{
+ int i, j, k;
+ double fdot;
+
+
+ /* transfer rhs to solution vector to preserve rhs */
+ for ( i = 0; i < nSlip; i++) g[i] = r[i];
+
+ /* matrix reduction */
+ for ( i = 1; i < nSlip; i++)
+ a[i][0] = a[i][0] / a[0][0];
+
+
+ for ( i = 1; i < nSlip; i++){
+ fdot = 0.0;
+ for ( k = 0; k < i; k++)
+ fdot += a[i][k] * a[k][i];
+ a[i][i] = a[i][i] - fdot;
+ for ( j = i+1; j < nSlip; j++){
+ fdot = 0.0;
+ for ( k = 0; k < i; k++)
+ fdot += a[i][k] * a[k][j];
+ a[i][j] = a[i][j] - fdot;
+ fdot = 0.0;
+ for ( k = 0; k < i; k++)
+ fdot += a[j][k] * a[k][i];
+ a[j][i] = ( a[j][i] - fdot) / a[i][i];
+ }
+ }
+
+
+ /* forward reduction of RHS */
+ for ( i = 1; i < nSlip; i++ ){
+ for ( k = 0; k < i; k++)
+ g[i] = g[i] - a[i][k] * g[k];
+ }
+
+ /* back substitution */
+ g[nSlip-1] = g[nSlip-1] / a[nSlip-1][nSlip-1];
+ for ( i = nSlip - 2; i >= 0; i=i-1){
+ for ( k = i+1; k < nSlip; k++)
+ g[i] = g[i] - a[i][k]*g[k];
+ g[i] = g[i] / a[i][i];
+ }
+}
+
+
+
