| Message ID | 5593D276.6090504@mentor.com |
|---|---|
| State | New |
| Headers | show |
On Wed, Jul 1, 2015 at 1:43 PM, Tom de Vries <Tom_deVries@mentor.com> wrote: > Hi, > > I. > > When running test libgomp.c/appendix-a/a.29.1.c with '--target_board > unix/-O2/-g', we run into this failure: > ... > FAIL: libgomp.c/appendix-a/a.29.1.c (test for excess errors) > Excess errors: > src/libgomp/testsuite/libgomp.c/appendix-a/a.29.1.c:6:1: error: type > mismatch between an SSA_NAME and its symbol > ... > > Without -g, the testcase passes. > > > II. > > The scenario for the failure is as follows: > > At fnsplit, we split off f.part.0 from f, which at source level looks like > this: > ... > void > f (int n, int B[n][n], int C[]) > { > int D[2][2] = { 1, 2, 3, 4 }; > int E[n][n]; > assert (n >= 2); > E[1][1] = 4; > #pragma omp parallel firstprivate(B, C, D, E) > { > assert (sizeof (B) == sizeof (int (*)[n])); > assert (sizeof (C) == sizeof (int *)); > assert (sizeof (D) == 4 * sizeof (int)); > assert (sizeof (E) == n * n * sizeof (int)); > /* Private B and C have values of original B and C. */ > assert (&B[1][1] == &A[1][1]); > assert (&C[3] == &A[1][1]); > assert (D[1][1] == 4); > assert (E[1][1] == 4); > } > } > ... > > The split introduces a debug_insn and ssa-name that references param B in f: > ... > # DEBUG D#4ptD.0 => B_3(D) > .. > > And a debug_insn that references param B in f.part.0: > ... > # DEBUG D#7ptD.0 s=> BD.1846 > ... > > At this point, the type of the ssa name and the param are the same. With the same PARM_DECL? I think that's the bug. > Then at inline, we decide to inline f.part.0 back into function f. > > For inlining, we rewrite the body of inlined function f.part.0. While > rewriting the debug insn mentioned above, we hit this code for param B: > ... > if (TREE_CODE (*tp) != OMP_CLAUSE) > TREE_TYPE (*tp) = remap_type (TREE_TYPE (*tp), id); > ... > And since it's a variable-sized type, the type of param is changed. The param of the inlined body but that should be unrelated to that refered to by # DEBUG D#4ptD.0 => B_3(D) in f. > Now the type of the ssa name and the param are no longer the same, and a bit > later we hit the ICE. > > > III. > > Attached patch fixes the ICE by handling PARM_DECL in remap_gimple_op_r. > [ I'm not confident though this is the right fix ] > > Bootstrapped and reg-tested on x86_64. > > OK for trunk, or to be fixed otherwise? > > Thanks, > - Tom
Handle PARM_DECL in remap_gimple_op_r 2015-06-30 Tom de Vries <tom@codesourcery.com> PR middle-end/66432 * tree-inline.c (remap_gimple_op_r): Handle PARM_DECL in remap_gimple_op_r. * testsuite/libgomp.c/appendix-a/a.29.1-g.c: New test. --- gcc/tree-inline.c | 10 ++++++-- libgomp/testsuite/libgomp.c/appendix-a/a.29.1-g.c | 31 +++++++++++++++++++++++ 2 files changed, 39 insertions(+), 2 deletions(-) create mode 100644 libgomp/testsuite/libgomp.c/appendix-a/a.29.1-g.c diff --git a/gcc/tree-inline.c b/gcc/tree-inline.c index 6f138ed..422fd33 100644 --- a/gcc/tree-inline.c +++ b/gcc/tree-inline.c @@ -919,10 +919,16 @@ remap_gimple_op_r (tree *tp, int *walk_subtrees, void *data) || decl_function_context (*tp) == id->src_fn)) /* These may need to be remapped for EH handling. */ *tp = remap_decl (*tp, id); - else if (TREE_CODE (*tp) == FIELD_DECL) + else if (TREE_CODE (*tp) == FIELD_DECL + || TREE_CODE (*tp) == PARM_DECL) { /* If the enclosing record type is variably_modified_type_p, the field - has already been remapped. Otherwise, it need not be. */ + has already been remapped. Otherwise, it need not be. + Likewise, if the PARM_DECL is a param of the inlined function, it will + already have been remapped. Otherwise, it might be a param of the + inlining function. This can happen if a code block in a function is + first split off by fnsplit, and then inlined back into the same + function. In that case, we don't remap. */ tree *n = id->decl_map->get (*tp); if (n) *tp = *n; diff --git a/libgomp/testsuite/libgomp.c/appendix-a/a.29.1-g.c b/libgomp/testsuite/libgomp.c/appendix-a/a.29.1-g.c new file mode 100644 index 0000000..140f89f --- /dev/null +++ b/libgomp/testsuite/libgomp.c/appendix-a/a.29.1-g.c @@ -0,0 +1,31 @@ +/* { dg-do run } */ +/* { dg-additional-options "-g" } */ + +#include <assert.h> +int A[2][2] = { 1, 2, 3, 4 }; +void +f (int n, int B[n][n], int C[]) +{ + int D[2][2] = { 1, 2, 3, 4 }; + int E[n][n]; + assert (n >= 2); + E[1][1] = 4; +#pragma omp parallel firstprivate(B, C, D, E) + { + assert (sizeof (B) == sizeof (int (*)[n])); /* { dg-warning "on array function parameter" } */ + assert (sizeof (C) == sizeof (int *)); /* { dg-warning "on array function parameter" } */ + assert (sizeof (D) == 4 * sizeof (int)); + assert (sizeof (E) == n * n * sizeof (int)); + /* Private B and C have values of original B and C. */ + assert (&B[1][1] == &A[1][1]); + assert (&C[3] == &A[1][1]); + assert (D[1][1] == 4); + assert (E[1][1] == 4); + } +} +int +main () +{ + f (2, A, A[0]); + return 0; +} -- 1.9.1
Hi, I. When running test libgomp.c/appendix-a/a.29.1.c with '--target_board unix/-O2/-g', we run into this failure: ... FAIL: libgomp.c/appendix-a/a.29.1.c (test for excess errors) Excess errors: src/libgomp/testsuite/libgomp.c/appendix-a/a.29.1.c:6:1: error: type mismatch between an SSA_NAME and its symbol ... Without -g, the testcase passes. II. The scenario for the failure is as follows: At fnsplit, we split off f.part.0 from f, which at source level looks like this: ... void f (int n, int B[n][n], int C[]) { int D[2][2] = { 1, 2, 3, 4 }; int E[n][n]; assert (n >= 2); E[1][1] = 4; #pragma omp parallel firstprivate(B, C, D, E) { assert (sizeof (B) == sizeof (int (*)[n])); assert (sizeof (C) == sizeof (int *)); assert (sizeof (D) == 4 * sizeof (int)); assert (sizeof (E) == n * n * sizeof (int)); /* Private B and C have values of original B and C. */ assert (&B[1][1] == &A[1][1]); assert (&C[3] == &A[1][1]); assert (D[1][1] == 4); assert (E[1][1] == 4); } } ... The split introduces a debug_insn and ssa-name that references param B in f: ... # DEBUG D#4ptD.0 => B_3(D) .. And a debug_insn that references param B in f.part.0: ... # DEBUG D#7ptD.0 s=> BD.1846 ... At this point, the type of the ssa name and the param are the same. Then at inline, we decide to inline f.part.0 back into function f. For inlining, we rewrite the body of inlined function f.part.0. While rewriting the debug insn mentioned above, we hit this code for param B: ... if (TREE_CODE (*tp) != OMP_CLAUSE) TREE_TYPE (*tp) = remap_type (TREE_TYPE (*tp), id); ... And since it's a variable-sized type, the type of param is changed. Now the type of the ssa name and the param are no longer the same, and a bit later we hit the ICE. III. Attached patch fixes the ICE by handling PARM_DECL in remap_gimple_op_r. [ I'm not confident though this is the right fix ] Bootstrapped and reg-tested on x86_64. OK for trunk, or to be fixed otherwise? Thanks, - Tom