[Bug tree-optimization/49419] [4.6/4.7 regression] gcc -O2 miscompiles gp2c
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49419
--- Comment #9 from Jakub Jelinek 2011-06-16
18:18:55 UTC ---
Created attachment 24548
--> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24548
gcc47-vrp.patch
So, I've tried this patch for extract_range_from_assert.
It bootstrapped/regtested on x86_64-linux and i686-linux, but regressed
FAIL: g++.dg/tree-ssa/pr18178.C scan-tree-dump-times vrp1 "if " 1
(just the vrp1 dump checking and code removal, the assembly is the same at the
end).
The VR is just different (- is vanilla, + is with the patch).
Value ranges after VRP:
-i_1: [0, +INF(OVF)]
+i_1: [0, +INF]
a_3(D): VARYING
D.2084_4: VARYING
D.2090_7: VARYING
@@ -681,10 +927,10 @@ D.2094_11: VARYING
i_12: [1, +INF(OVF)]
D.2077_13: VARYING
.MEM_14: VARYING
-i_19: [-INF, D.2084_4 + -1] EQUIVALENCES: { i_1 } (1 elements)
-i_20: [0, +INF] EQUIVALENCES: { i_1 i_19 } (2 elements)
+i_19: [0, +INF] EQUIVALENCES: { } (0 elements)
+i_20: [0, +INF] EQUIVALENCES: { i_19 } (1 elements)
a_21: ~[0B, 0B] EQUIVALENCES: { a_3(D) } (1 elements)
-D.2084_22: [i_19 + 1, +INF] EQUIVALENCES: { D.2084_4 } (1 elements)
+D.2084_22: [1, +INF] EQUIVALENCES: { D.2084_4 } (1 elements)
D.2077_23: ~[0B, 0B] EQUIVALENCES: { D.2077_13 } (1 elements)
but resulted in:
Folding statement: if (i_20 >= D.2084_22)
-Folding predicate i_20 >= D.2084_22 to 0
-Folded into: if (0 != 0)
-
+Not folded
So, not sure if I should push this anyway (with adjusting/removing the
testcase), or not.
[Bug tree-optimization/49419] [4.6/4.7 regression] gcc -O2 miscompiles gp2c
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49419 Jakub Jelinek changed: What|Removed |Added Status|NEW |RESOLVED Resolution||FIXED --- Comment #8 from Jakub Jelinek 2011-06-16 08:17:46 UTC --- The bug should be fixed now, though VRP should be cerrtainly improved in various ways, including choosing better vr during extract_range_from_assert, trying to derive numeric ranges from symbolic ranges x - 1 maximums or x + 1 minimums, and IMHO adjust_range_for_scev could use number_of_latch_executions to see if the number is simple enough and a smaller iteration count can be derived from its value ranges or, e.g. if it is the init SSA_NAME itself or that plus/minus a constant, it could determine it doesn't decrease until -INF but only until 0 +- constant.
[Bug tree-optimization/49419] [4.6/4.7 regression] gcc -O2 miscompiles gp2c
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49419 --- Comment #7 from Jakub Jelinek 2011-06-16 07:54:46 UTC --- Author: jakub Date: Thu Jun 16 07:54:43 2011 New Revision: 175095 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=175095 Log: PR tree-optimization/49419 * tree-vrp.c (execute_vrp): Call init_range_assertions before estimate_numbers_of_iterations, call free_number_of_iterations_estimates before calling remove_range_assertions. * gcc.c-torture/execute/pr49419.c: New test. Added: branches/gcc-4_6-branch/gcc/testsuite/gcc.c-torture/execute/pr49419.c Modified: branches/gcc-4_6-branch/gcc/ChangeLog branches/gcc-4_6-branch/gcc/testsuite/ChangeLog branches/gcc-4_6-branch/gcc/tree-vrp.c
[Bug tree-optimization/49419] [4.6/4.7 regression] gcc -O2 miscompiles gp2c
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49419 --- Comment #6 from Jakub Jelinek 2011-06-16 07:45:22 UTC --- Author: jakub Date: Thu Jun 16 07:45:17 2011 New Revision: 175092 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=175092 Log: PR tree-optimization/49419 * tree-vrp.c (execute_vrp): Call init_range_assertions before estimate_numbers_of_iterations, call free_number_of_iterations_estimates before calling remove_range_assertions. * gcc.c-torture/execute/pr49419.c: New test. Added: trunk/gcc/testsuite/gcc.c-torture/execute/pr49419.c Modified: trunk/gcc/ChangeLog trunk/gcc/testsuite/ChangeLog trunk/gcc/tree-vrp.c
[Bug tree-optimization/49419] [4.6/4.7 regression] gcc -O2 miscompiles gp2c
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49419 Jakub Jelinek changed: What|Removed |Added Status|UNCONFIRMED |NEW Last reconfirmed||2011.06.15 16:02:59 Component|target |tree-optimization Ever Confirmed|0 |1
