commit 14041afe24556efd5845564aa183b6451fd9d6cc Author: law Date: Thu Feb 1 16:22:56 2018 +0000 PR target/84128 * config/i386/i386.c (release_scratch_register_on_entry): Add new OFFSET and RELEASE_VIA_POP arguments. Use SP+OFFSET to restore the scratch if RELEASE_VIA_POP is false. (ix86_adjust_stack_and_probe_stack_clash): Un-constify SIZE. If we have to save a temporary register, decrement SIZE appropriately. Pass new arguments to release_scratch_register_on_entry. (ix86_adjust_stack_and_probe): Likewise. (ix86_emit_probe_stack_range): Pass new arguments to release_scratch_register_on_entry. PR target/84128 * gcc.target/i386/pr84128.c: New test. diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c index 2fe2a0c..c25d26c 100644 --- a/gcc/config/i386/i386.c +++ b/gcc/config/i386/i386.c @@ -10182,22 +10182,39 @@ get_scratch_register_on_entry (struct scratch_reg *sr) } } -/* Release a scratch register obtained from the preceding function. */ +/* Release a scratch register obtained from the preceding function. + + If RELEASE_VIA_POP is true, we just pop the register off the stack + to release it. This is what non-Linux systems use with -fstack-check. + + Otherwise we use OFFSET to locate the saved register and the + allocated stack space becomes part of the local frame and is + deallcated by the epilogue. */ static void -release_scratch_register_on_entry (struct scratch_reg *sr) +release_scratch_register_on_entry (struct scratch_reg *sr, HOST_WIDE_INT offset, + bool release_via_pop) { if (sr->saved) { - struct machine_function *m = cfun->machine; - rtx x, insn = emit_insn (gen_pop (sr->reg)); + if (release_via_pop) + { + struct machine_function *m = cfun->machine; + rtx x, insn = emit_insn (gen_pop (sr->reg)); - /* The RTX_FRAME_RELATED_P mechanism doesn't know about pop. */ - RTX_FRAME_RELATED_P (insn) = 1; - x = gen_rtx_PLUS (Pmode, stack_pointer_rtx, GEN_INT (UNITS_PER_WORD)); - x = gen_rtx_SET (VOIDmode, stack_pointer_rtx, x); - add_reg_note (insn, REG_FRAME_RELATED_EXPR, x); - m->fs.sp_offset -= UNITS_PER_WORD; + /* The RTX FRAME_RELATED_P mechanism doesn't know about pop. */ + RTX_FRAME_RELATED_P (insn) = 1; + x = gen_rtx_PLUS (Pmode, stack_pointer_rtx, GEN_INT (UNITS_PER_WORD)); + x = gen_rtx_SET (VOIDmode, stack_pointer_rtx, x); + add_reg_note (insn, REG_FRAME_RELATED_EXPR, x); + m->fs.sp_offset -= UNITS_PER_WORD; + } + else + { + rtx x = gen_rtx_PLUS (Pmode, stack_pointer_rtx, GEN_INT (offset)); + x = gen_rtx_SET (VOIDmode, sr->reg, gen_rtx_MEM (word_mode, x)); + emit_insn (x); + } } } @@ -10212,7 +10229,7 @@ release_scratch_register_on_entry (struct scratch_reg *sr) pushed on the stack. */ static void -ix86_adjust_stack_and_probe_stack_clash (const HOST_WIDE_INT size, +ix86_adjust_stack_and_probe_stack_clash (HOST_WIDE_INT size, const bool int_registers_saved) { struct machine_function *m = cfun->machine; @@ -10331,6 +10348,12 @@ ix86_adjust_stack_and_probe_stack_clash (const HOST_WIDE_INT size, struct scratch_reg sr; get_scratch_register_on_entry (&sr); + /* If we needed to save a register, then account for any space + that was pushed (we are not going to pop the register when + we do the restore). */ + if (sr.saved) + size -= UNITS_PER_WORD; + /* Step 1: round SIZE down to a multiple of the interval. */ HOST_WIDE_INT rounded_size = size & -probe_interval; @@ -10379,7 +10402,9 @@ ix86_adjust_stack_and_probe_stack_clash (const HOST_WIDE_INT size, m->fs.cfa_reg == stack_pointer_rtx); dump_stack_clash_frame_info (PROBE_LOOP, size != rounded_size); - release_scratch_register_on_entry (&sr); + /* This does not deallocate the space reserved for the scratch + register. That will be deallocated in the epilogue. */ + release_scratch_register_on_entry (&sr, size, false); } /* Make sure nothing is scheduled before we are done. */ @@ -10392,7 +10417,7 @@ ix86_adjust_stack_and_probe_stack_clash (const HOST_WIDE_INT size, pushed on the stack. */ static void -ix86_adjust_stack_and_probe (const HOST_WIDE_INT size, +ix86_adjust_stack_and_probe (HOST_WIDE_INT size, const bool int_registers_saved) { /* We skip the probe for the first interval + a small dope of 4 words and @@ -10465,6 +10490,11 @@ ix86_adjust_stack_and_probe (const HOST_WIDE_INT size, get_scratch_register_on_entry (&sr); + /* If we needed to save a register, then account for any space + that was pushed (we are not going to pop the register when + we do the restore). */ + if (sr.saved) + size -= UNITS_PER_WORD; /* Step 1: round SIZE to the previous multiple of the interval. */ @@ -10516,7 +10546,9 @@ ix86_adjust_stack_and_probe (const HOST_WIDE_INT size, (get_probe_interval () + dope)))); - release_scratch_register_on_entry (&sr); + /* This does not deallocate the space reserved for the scratch + register. That will be deallocated in the epilogue. */ + release_scratch_register_on_entry (&sr, size, false); } gcc_assert (cfun->machine->fs.cfa_reg != stack_pointer_rtx); @@ -10669,7 +10701,7 @@ ix86_emit_probe_stack_range (HOST_WIDE_INT first, HOST_WIDE_INT size, sr.reg), rounded_size - size)); - release_scratch_register_on_entry (&sr); + release_scratch_register_on_entry (&sr, size, true); } /* Make sure nothing is scheduled before we are done. */ diff --git a/gcc/testsuite/gcc.target/i386/pr84128.c b/gcc/testsuite/gcc.target/i386/pr84128.c new file mode 100644 index 0000000..a8323fd --- /dev/null +++ b/gcc/testsuite/gcc.target/i386/pr84128.c @@ -0,0 +1,30 @@ +/* { dg-do run } */ +/* { dg-options "-O2 -march=i686 -mtune=generic -fstack-clash-protection" } */ +/* { dg-require-effective-target ia32 } */ + +__attribute__ ((noinline, noclone, weak, regparm (3))) +int +f1 (long arg0, int (*pf) (long, void *)) +{ + unsigned char buf[32768]; + return pf (arg0, buf); +} + +__attribute__ ((noinline, noclone, weak)) +int +f2 (long arg0, void *ignored) +{ + if (arg0 != 17) + __builtin_abort (); + return 19; +} + +int +main (void) +{ + if (f1 (17, f2) != 19) + __builtin_abort (); + return 0; +} + +