We saw a bug in dse2 after porting the patch https://gcc.gnu.org/ml/gcc-patches/2014-10/msg01209.html from gcc-4_9 to google-4_9 branch. From the analysis below, I think the problem exists but is hidden in trunk and gcc-4_9 too. I cannot extract a small testcase to show it independently without turning on some optimization in google-4_9, so I just described it here: We have such IR in a case: The IR before dse2: (insn/f 67 4 68 2 (set (mem:DI (pre_dec:DI (reg/f:DI 7 sp)) [0 S8 A8]) (reg/f:DI 6 bp)) contentads/adx/mixer/auction/candidate.cc:14 -1 (nil)) (insn/f 68 67 69 2 (set (reg/f:DI 6 bp) (reg/f:DI 7 sp)) contentads/adx/mixer/auction/candidate.cc:14 -1 (nil)) (insn/f 70 69 71 2 (parallel [ (set (reg/f:DI 7 sp) (plus:DI (reg/f:DI 7 sp) (const_int -24 [0xffffffffffffffe8]))) (clobber (reg:CC 17 flags)) (clobber (mem:BLK (scratch) [0 A8])) ]) contentads/adx/mixer/auction/candidate.cc:14 -1 (nil)) (note 71 70 2 2 NOTE_INSN_PROLOGUE_END) .... (insn 7 3 9 2 (set (mem/c:SI (reg/f:DI 7 sp) [0 MEM[(void *)&D.3507754]+0 S4 A128]) (const_int 0 [0])) ./ads/base/money.h:67 90 {*movsi_internal} (nil)) (insn 9 7 10 2 (set (mem/c:HI (reg/f:DI 7 sp) [0 MEM[(void *)&D.3507754]+0 S2 A128]) (const_int 21333 [0x5355])) ./ads/base/money.h:68 92 {*movhi_internal} (nil)) (insn 10 9 11 2 (set (mem/c:QI (plus:DI (reg/f:DI 7 sp) (const_int 2 [0x2])) [0 MEM[(void *)&D.3507754]+2 S1 A16]) (const_int 68 [0x44])) ./ads/base/money.h:68 93 {*movqi_internal} (nil)) (insn 11 10 12 2 (set (reg:SI 0 ax [orig:87 D.3507754 ] [87]) (mem/c:SI (reg/f:DI 7 sp) [0 D.3507754+0 S4 A128])) ./ads/base/money.h:302 90 {*movsi_internal} (expr_list:REG_EQUIV (mem/c:SI (plus:DI (reg/f:DI 20 frame) (const_int -16 [0xfffffffffffffff0])) [0 D.3507754+0 S4 A128]) (nil))) ... (insn 15 13 17 2 (set (mem/c:SI (reg/f:DI 7 sp) [0 MEM[(void *)&D.3507756]+0 S4 A128]) (const_int 0 [0])) ./ads/base/money.h:67 90 {*movsi_internal} (nil)) The IR after dse2: The store in insn 10 is deleted. The other part is the same as above. (mem/c:QI (plus:DI (reg/f:DI 7 sp) (const_int 2 [0x2])) in insn10 is regarded to have no alias with (mem/c:SI (reg/f:DI 7 sp) in insn11, which is wrong. This is because with the applied patch, get_addr is used to extract original addresses for x_addr and mem_addr before they are used to find_base_term and used in base_alias_check. See the description of x_addr and mem_addr below: x is (mem/c:SI (reg/f:DI 7 sp) x_addr before calling get_addr is: (value:DI 4:12939 @0x84355f8/0x84354a0) x_addr after calling get_addr is: (plus:DI (value:DI 3:8637 @0x84355e8/0x8435478) (const_int -24 [0xffffffffffffffe8])) x_addr_base is: (address:DI -4) mem is (mem/c:QI (plus:DI (reg/f:DI 7 sp) (const_int 2 [0x2])) mem_addr before calling get_addr is: (plus:DI (value:DI 4:12939 @0x84355f8/0x84354a0) (const_int 2 [0x2])) mem_addr after calling get_addr is: // Notice: get_addr cannot handle plus expr, so it returns the origin expr. (plus:DI (value:DI 4:12939 @0x84355f8/0x84354a0) (const_int 2 [0x2])) mem_addr_base is: (address:DI -1) // value:DI 4:12939 @0x84355f8/0x84354a0 corresponds to reg/f:DI 7 sp // value:DI 3:8637 @0x84355e8/0x8435478 corresponds to reg/f:DI 6 bp // address:DI -1 corresponds to reg/f:DI 7 sp // address:DI -4 corresponds to reg/f:DI 6 bp x_addr_base and mem_addr_base are different, and unique_base_value_p will return true for (address:DI -1) and (address:DI -4), so base_alias_check will return 0, which is wrong. I think the root cause of the problem is get_addr can only handle VALUE but cannot handle VALUE inside an expr, like:(plus:DI (value:DI 4:12939 @0x84355f8/0x84354a0) (const_int 2 [0x2])), so find_base_term cannot know x_addr and mem_addr actually have the same base.
The experimental patch is to call get_addr for VALUE of base before plus other constant, when creating mem_addr for dependence check and for store_info. bootstrap and regression on x86_64-linux-gnu are ok. Index: dse.c =================================================================== --- dse.c (revision 219421) +++ dse.c (working copy) @@ -1564,6 +1564,7 @@ record_store (rtx body, bb_info_t bb_inf = rtx_group_vec[group_id]; mem_addr = group->canon_base_addr; } + mem_addr = get_addr (mem_addr); if (offset) mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); } @@ -2177,6 +2178,7 @@ check_mem_read_rtx (rtx *loc, bb_info_t = rtx_group_vec[group_id]; mem_addr = group->canon_base_addr; } + mem_addr = get_addr (mem_addr); if (offset) mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); }
Google ref: b/18933287
Author: wmi Date: Thu Jan 22 17:59:23 2015 New Revision: 220010 URL: https://gcc.gnu.org/viewcvs?rev=220010&root=gcc&view=rev Log: 2015-01-22 Wei Mi <wmi@google.com> PR rtl-optimization/64557 * dse.c (record_store): Call get_addr for mem_addr. (check_mem_read_rtx): Likewise. Modified: trunk/gcc/ChangeLog trunk/gcc/dse.c
Author: wmi Date: Fri Jan 23 17:55:32 2015 New Revision: 220051 URL: https://gcc.gnu.org/viewcvs?rev=220051&root=gcc&view=rev Log: Backported from trunk. 2015-01-22 Wei Mi <wmi@google.com> PR rtl-optimization/64557 * dse.c (record_store): Call get_addr for mem_addr. (check_mem_read_rtx): Likewise. Modified: branches/gcc-4_9-branch/gcc/ChangeLog branches/gcc-4_9-branch/gcc/dse.c
Author: uros Date: Fri Feb 20 12:04:21 2015 New Revision: 220854 URL: https://gcc.gnu.org/viewcvs?rev=220854&root=gcc&view=rev Log: Backport from mainline 2015-01-22 Wei Mi <wmi@google.com> PR rtl-optimization/64557 * dse.c (record_store): Call get_addr for mem_addr. (check_mem_read_rtx): Likewise. Backport from mainline 2014-10-20 Uros Bizjak <ubizjak@gmail.com> * varasm.c (const_alias_set): Remove. (init_varasm_once): Remove initialization of const_alias_set. (build_constant_desc): Do not set alias set to const_alias_set. Backport from mainline 2014-10-14 Uros Bizjak <ubizjak@gmail.com> PR rtl-optimization/63475 * alias.c (true_dependence_1): Always use get_addr to extract true address operands from x_addr and mem_addr. Use extracted address operands to check for references with alignment ANDs. Use extracted address operands with find_base_term and base_alias_check. For noncanonicalized operands call canon_rtx with extracted address operand. (write_dependence_1): Ditto. (may_alias_p): Ditto. Remove unused calls to canon_rtx. Backport from mainline 2014-10-10 Uros Bizjak <ubizjak@gmail.com> PR rtl-optimization/63483 * alias.c (true_dependence_1): Do not exit early for MEM_READONLY_P references when alignment ANDs are involved. (write_dependence_p): Ditto. (may_alias_p): Ditto. Backport from mainline 2013-03-26 Richard Biener <rguenther@suse.de> * alias.c (find_base_term): Avoid redundant and not used recursion. (base_alias_check): Get the initial base term from the caller. (true_dependence_1): Compute and pass base terms to base_alias_check. (write_dependence_p): Likewise. (may_alias_p): Likewise. Modified: branches/gcc-4_8-branch/gcc/ChangeLog branches/gcc-4_8-branch/gcc/alias.c branches/gcc-4_8-branch/gcc/dse.c branches/gcc-4_8-branch/gcc/varasm.c
Fixed everywhere.
This fix causes a regression on s390. Please refer to bug 67443: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67443
Author: jakub Date: Tue Jan 19 12:34:45 2016 New Revision: 232554 URL: https://gcc.gnu.org/viewcvs?rev=232554&root=gcc&view=rev Log: PR rtl-optimization/68955 PR rtl-optimization/64557 * dse.c (record_store, check_mem_read_rtx): Don't call get_addr here. Fix up formatting. * alias.c (get_addr): Handle VALUE +/- CONST_SCALAR_INT_P. * gcc.dg/torture/pr68955.c: New test. Added: trunk/gcc/testsuite/gcc.dg/torture/pr68955.c Modified: trunk/gcc/ChangeLog trunk/gcc/alias.c trunk/gcc/dse.c trunk/gcc/testsuite/ChangeLog
Author: jakub Date: Wed Feb 10 18:34:30 2016 New Revision: 233293 URL: https://gcc.gnu.org/viewcvs?rev=233293&root=gcc&view=rev Log: Backported from mainline 2016-01-19 Jakub Jelinek <jakub@redhat.com> PR rtl-optimization/68955 PR rtl-optimization/64557 * dse.c (record_store, check_mem_read_rtx): Don't call get_addr here. Fix up formatting. * alias.c (get_addr): Handle VALUE +/- CONST_SCALAR_INT_P. * gcc.dg/torture/pr68955.c: New test. Added: branches/gcc-5-branch/gcc/testsuite/gcc.dg/torture/pr68955.c Modified: branches/gcc-5-branch/gcc/ChangeLog branches/gcc-5-branch/gcc/alias.c branches/gcc-5-branch/gcc/dse.c branches/gcc-5-branch/gcc/testsuite/ChangeLog
Author: jakub Date: Thu Feb 11 09:23:06 2016 New Revision: 233331 URL: https://gcc.gnu.org/viewcvs?rev=233331&root=gcc&view=rev Log: Backported from mainline 2016-01-19 Jakub Jelinek <jakub@redhat.com> PR rtl-optimization/68955 PR rtl-optimization/64557 * dse.c (record_store, check_mem_read_rtx): Don't call get_addr here. Fix up formatting. * alias.c (get_addr): Handle VALUE +/- CONST_SCALAR_INT_P. * gcc.dg/torture/pr68955.c: New test. Added: branches/gcc-4_9-branch/gcc/testsuite/gcc.dg/torture/pr68955.c Modified: branches/gcc-4_9-branch/gcc/ChangeLog branches/gcc-4_9-branch/gcc/alias.c branches/gcc-4_9-branch/gcc/dse.c branches/gcc-4_9-branch/gcc/testsuite/ChangeLog