The testcase from PR47086 void foo () { int n = 0; while (1) { int i[n % 1]; n++; } } reveals that while on the tree level n++ has an outgoing EH edge, RTL expansion leaves us with ;; n_5 = n_1 + 1; (insn 12 11 13 (set (reg:DI 67) (sign_extend:DI (reg/v:SI 61 [ n ]))) /space/rguenther/src/svn/trunk/gcc/testsuite/gcc.dg/pr47086.c:11 -1 (nil)) (insn 13 12 14 (set (reg:DI 4 si) (const_int 1 [0x1])) /space/rguenther/src/svn/trunk/gcc/testsuite/gcc.dg/pr47086.c:11 -1 (nil)) (insn 14 13 15 (set (reg:DI 5 di) (reg:DI 67)) /space/rguenther/src/svn/trunk/gcc/testsuite/gcc.dg/pr47086.c:11 -1 (nil)) (call_insn/u 15 14 16 (set (reg:DI 0 ax) (call (mem:QI (symbol_ref:DI ("__addvdi3") [flags 0x41]) [0 S1 A8]) (const_int 0 [0]))) /space/rguenther/src/svn/trunk/gcc/testsuite/gcc.dg/pr47086.c:11 -1 (expr_list:REG_EH_REGION (const_int -2147483648 [0xffffffff80000000]) (nil)) (expr_list:REG_DEP_TRUE (use (reg:DI 4 si)) (expr_list:REG_DEP_TRUE (use (reg:DI 5 di)) (nil)))) (insn 16 15 17 (set (reg:DI 68) (reg:DI 0 ax)) /space/rguenther/src/svn/trunk/gcc/testsuite/gcc.dg/pr47086.c:11 -1 (expr_list:REG_EQUAL (plus:DI (reg:DI 67) (const_int 1 [0x1])) (nil))) (insn 17 16 0 (set (reg/v:SI 63 [ n ]) (subreg:SI (reg:DI 68) 0)) /space/rguenther/src/svn/trunk/gcc/testsuite/gcc.dg/pr47086.c:11 -1 (nil)) note that the call is marked CONST. The EH edge is simply thrown away: ;; Basic block 4 , prev 3, next 5, loop_depth 1, count 0, freq 10000, maybe hot, flags: new reachable rtl modified. ;; Predecessors: 3 [100.0%] (fallthru) 5 [100.0%] (dfs_back) ... (call_insn/u 15 14 16 4 (set (reg:DI 0 ax) (call (mem:QI (symbol_ref:DI ("__addvdi3") [flags 0x41]) [0 S1 A8]) (const_int 0 [0]))) /space/rguenther/src/svn/trunk/gcc/testsuite/gcc.dg/pr47086.c:11 -1 (expr_list:REG_EH_REGION (const_int -2147483648 [0xffffffff80000000]) (nil)) (expr_list:REG_DEP_TRUE (use (reg:DI 4 si)) (expr_list:REG_DEP_TRUE (use (reg:DI 5 di)) (nil)))) (insn 16 15 17 4 (set (reg:DI 68) (reg:DI 0 ax)) /space/rguenther/src/svn/trunk/gcc/testsuite/gcc.dg/pr47086.c:11 -1 (expr_list:REG_EQUAL (plus:DI (reg:DI 67) (const_int 1 [0x1])) (nil))) (insn 17 16 18 4 (set (reg/v:SI 63 [ n ]) (subreg:SI (reg:DI 68) 0)) /space/rguenther/src/svn/trunk/gcc/testsuite/gcc.dg/pr47086.c:11 -1 (nil)) ;; Successors: 5 [100.0%] (fallthru) So we seem to leave around the libcall because of the EH side-effect but do not consider that call as possibly throwing (as it is marked nothrow on RTL) in find_many_sub_basic_blocks (at least) and DCE it later, too (sounds similar to PR19020). Currently generated assembly for gcc.dg/pr47086.c on x86_64 is foo: .LFB0: .cfi_startproc .cfi_personality 0x3,__gcc_personality_v0 pushq %rbx .cfi_def_cfa_offset 16 .cfi_offset 3, -16 movl $0, %ebx .L2: addl $1, %ebx jmp .L2 I'm running into this issue when removing unreachable blocks during expansion, then the landing pad unwind info expansion ICEs because the landing pad was removed. We mark the call as not throwing via emit_libcall_block because we pass it (plus (reg) 1) as EQUIV, which may not trap.
Created attachment 26732 [details] patch Patch that fixes my issue and works until rtl_dce, which still removes the call, even though it throws internally (call_insn/u 15 13 36 3 (set (reg:DI 0 ax) (call (mem:QI (symbol_ref:DI ("__addvdi3") [flags 0x41]) [0 S1 A8]) (const_int 0 [0]))) /space/rguenther/src/svn/trunk/gcc/testsuite/gcc.dg/pr47086.c:11 636 {*call_value} (expr_list:REG_DEAD (reg:DI 5 di) (expr_list:REG_DEAD (reg:DI 4 si) (expr_list:REG_UNUSED (reg:DI 0 ax) (expr_list:REG_EH_REGION (const_int 1 [0x1]) (nil))))) (expr_list:REG_DEP_TRUE (use (reg:DI 4 si)) (expr_list:REG_DEP_TRUE (use (reg:DI 5 di)) (nil)))) ;; Successors: 5 (ab,abcall,eh,loop_exit) 4 [100.0%] (fallthru) ;; lr out 3 [bx] 6 [bp] 7 [sp] 16 [argp] 20 [frame] 41 [r12]
Author: rguenth Date: Mon Mar 5 13:08:55 2012 New Revision: 184932 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=184932 Log: 2012-03-05 Richard Guenther <rguenther@suse.de> PR middle-end/52353 * optabs.h (trapv_unoptab_p): New function. (trapv_binoptab_p): Likewise. * optabs.c (expand_binop): Use emit_libcall_block_1 with a proper equiv_may_trap argument. (expand_unop): Likewise. (emit_libcall_block_1): Take extra argument whether the instruction may trap. Renamed from ... (emit_libcall_block): ... this. New wrapper. Modified: trunk/gcc/ChangeLog trunk/gcc/optabs.c trunk/gcc/optabs.h
Fixed for trunk.