The program below doesn't exit with 0 as expected when it's compiled with -O2. It works fine with -O1 or -O2 -fno-delayed-branch. -- int *dummy; void bar (int *p) { dummy = p; } void foo (int x) { int var; bar (&var); if (x) throw 1; } int main () { try { foo (1); } catch (...) { return 0; } return 1; } -- With -O2, the function foo looks like: _Z3fooi: .LFB3: mov.l r8,@-r15 .LCFI3: mov r4,r8 mov.l r14,@-r15 .LCFI4: sts.l pr,@-r15 .LCFI5: mov.l .L11,r1 add #-4,r15 .LCFI6: mov r15,r14 .LCFI7: jsr @r1 mov r14,r4 tst r8,r8 bf/s .L10 add #4,r14 mov r14,r15 lds.l @r15+,pr mov.l @r15+,r14 rts mov.l @r15+,r8 .align 5 .L10: mov.l .L12,r0 jsr @r0 mov #4,r4 mov #1,r1 mov.l r1,@r0 mov.l .L13,r1 mov r0,r4 mov.l .L14,r5 jsr @r1 mov #0,r6 ... Thus the throw part starting with .L10 is called after the excution of "add #4,r14" in the delayed slot. It seems that this doesn't match the frame info of foo: $ readelf -a ./a.out | grep foo 99: 00400760 100 FUNC GLOBAL DEFAULT 11 _Z3fooi $ readelf --debug-dump=frames ./a.out The section .eh_frame contains: ... 0000001c 00000028 00000020 FDE cie=00000000 pc=00400760..004007c4 Augmentation data: 00 00 00 00 DW_CFA_advance_loc: 2 to 00400762 DW_CFA_def_cfa_offset: 4 DW_CFA_offset: r8 at cfa-4 DW_CFA_advance_loc: 4 to 00400766 DW_CFA_def_cfa_offset: 8 DW_CFA_advance_loc: 2 to 00400768 DW_CFA_def_cfa_offset: 12 DW_CFA_advance_loc: 4 to 0040076c DW_CFA_def_cfa_offset: 16 DW_CFA_offset: r17 at cfa-12 DW_CFA_offset: r14 at cfa-8 DW_CFA_advance_loc: 2 to 0040076e DW_CFA_def_cfa_reg: r14 DW_CFA_nop DW_CFA_nop which assumes that r14 points the bottom of the frame when the throw part is called.
We don't emit call frame information for the epilogue. If we did, than any basic blocks that happen to come after the epilogue due to block reordering would end up with incorrect cfi information, since we don't have any machinery for writing compensating cfi when we skip over an epilogue. Thus, the only way we have to make this work with the current infrastructure is to avoid scheduling the epilogue with previous basic blocks when exception handling is enabled (this also affects debugging information, but I'd say we shouldn't pessimize -O2 code to get better debugging information). There is already code in sh.c:sh_expand_epilogue to emit a blockage instruction before the stack adjustments. However, if a frame pointer is needed, the initial adjustment of the frame pointer isn't adjusted, since this doesn't matter for interrupts: if (frame_pointer_needed) { output_stack_adjust (frame_size, frame_pointer_rtx, e, &live_regs_mask); /* We must avoid moving the stack pointer adjustment past code which reads from the local frame, else an interrupt could occur after the SP adjustment and clobber data in the local frame. */ emit_insn (gen_blockage ()); emit_insn (GEN_MOV (stack_pointer_rtx, frame_pointer_rtx)); } So, if flag_exceptions is set, we should emit this blockage before the frame pointer adjustment.
I'm testing a patch you suggest with the usual bootstrap and regression test. I'll send it to gcc-patches if it passes. --- ORIG/gcc/gcc/config/sh/sh.c 2004-10-08 07:46:30.000000000 +0900 +++ LOCAL/gcc/gcc/config/sh/sh.c 2004-10-19 07:20:50.000000000 +0900 @@ -5828,6 +5828,10 @@ sh_expand_epilogue (bool sibcall_p) if (frame_pointer_needed) { + /* We must avoid scheduling the epilogue with previous basic blocks + when exception handling is enabled. See PR/18032. */ + if (flag_exceptions) + emit_insn (gen_blockage ()); output_stack_adjust (frame_size, frame_pointer_rtx, e, &live_regs_mask); /* We must avoid moving the stack pointer adjustment past code
Subject: Bug 18032 CVSROOT: /cvs/gcc Module name: gcc Changes by: kkojima@gcc.gnu.org 2004-10-20 13:57:23 Modified files: gcc : ChangeLog gcc/config/sh : sh.c Log message: PR target/18032 * config/sh/sh.c (sh_expand_epilogue): Emit a blockage insn before the frame pointer adjustment when exception handling is enabled. Patches: http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.5957&r2=2.5958 http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/config/sh/sh.c.diff?cvsroot=gcc&r1=1.306&r2=1.307
Fixed.