This patch catches a couple places where the generic parts of the
compiler would transform prologue probing sequences in unsafe ways.
Changes since V2:
Testing for aarch64 showed the scheduler would perform transformations
similar to combine-stack-adjustments via its dependency breaking
transformations. So the scheduler was taught to not break dependencies
when the increment insn contains a STACK_CHECK note. A test for this
specific issue shows up in the aarch64 tests in the next patch.
* combine-stack-adj.c (combine_stack_adjustments_for_block): Do
nothing for stack adjustments with REG_STACK_CHECK.
* sched-deps.c (parse_add_or_inc): Reject insns with
REG_STACK_CHECK from dependency breaking.
* config/i386/i386.c (pro_epilogue_adjust_stack): Return insn.
(ix86_adjust_satck_and_probe_stack_clash): Add REG_STACK_NOTEs.
* reg-notes.def (STACK_CHECK): New note.
testsuite/
* gcc.target/i386/stack-check-11.c: New test.
diff --git a/gcc/combine-stack-adj.c b/gcc/combine-stack-adj.c
index 9ec14a3..82d6dba 100644
--- a/gcc/combine-stack-adj.c
+++ b/gcc/combine-stack-adj.c
@@ -508,6 +508,8 @@ combine_stack_adjustments_for_block (basic_block bb)
continue;
set = single_set_for_csa (insn);
+ if (set && find_reg_note (insn, REG_STACK_CHECK, NULL_RTX))
+ set = NULL_RTX;
if (set)
{
rtx dest = SET_DEST (set);
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index b4ddc83..69fecb9 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -13405,7 +13405,7 @@ ix86_add_queued_cfa_restore_notes (rtx insn)
zero if %r11 register is live and cannot be freely used and positive
otherwise. */
-static void
+static rtx
pro_epilogue_adjust_stack (rtx dest, rtx src, rtx offset,
int style, bool set_cfa)
{
@@ -13496,6 +13496,7 @@ pro_epilogue_adjust_stack (rtx dest, rtx src, rtx
offset,
m->fs.sp_valid = valid;
m->fs.sp_realigned = realigned;
}
+ return insn;
}
/* Find an available register to be used as dynamic realign argument
@@ -13839,9 +13840,11 @@ ix86_adjust_stack_and_probe_stack_clash (const
HOST_WIDE_INT size)
for (i = probe_interval; i <= size; i += probe_interval)
{
/* Allocate PROBE_INTERVAL bytes. */
- pro_epilogue_adjust_stack (stack_pointer_rtx, stack_pointer_rtx,
- GEN_INT (-probe_interval), -1,
- m->fs.cfa_reg == stack_pointer_rtx);
+ rtx insn
+ = pro_epilogue_adjust_stack (stack_pointer_rtx, stack_pointer_rtx,
+ GEN_INT (-PROBE_INTERVAL), -1,
+ m->fs.cfa_reg == stack_pointer_rtx);
+ add_reg_note (insn, REG_STACK_CHECK, const0_rtx);
/* And probe at *sp. */
emit_stack_probe (stack_pointer_rtx);
diff --git a/gcc/reg-notes.def b/gcc/reg-notes.def
index 8734d26..18cf7e3 100644
--- a/gcc/reg-notes.def
+++ b/gcc/reg-notes.def
@@ -223,6 +223,10 @@ REG_NOTE (ARGS_SIZE)
pseudo reg. */
REG_NOTE (RETURNED)
+/* Indicates the instruction is a stack check probe that should not
+ be combined with other stack adjustments. */
+REG_NOTE (STACK_CHECK)
+
/* Used to mark a call with the function decl called by the call.
The decl might not be available in the call due to splitting of the call
insn. This note is a SYMBOL_REF. */
diff --git a/gcc/sched-deps.c b/gcc/sched-deps.c
index 4c459e6..a661426 100644
--- a/gcc/sched-deps.c
+++ b/gcc/sched-deps.c
@@ -4713,6 +4713,11 @@ parse_add_or_inc (struct mem_inc_info *mii, rtx_insn
*insn, bool before_mem)
if (RTX_FRAME_RELATED_P (insn) || !pat)
return false;
+ /* Do not allow breaking data dependencies for insns that are marked
+ with REG_STACK_CHECK. */
+ if (find_reg_note (insn, REG_STACK_CHECK, NULL))
+ return false;
+
/* Result must be single reg. */
if (!REG_P (SET_DEST (pat)))
return false;
diff --git a/gcc/testsuite/gcc.target/i386/stack-check-11.c
b/gcc/testsuite/gcc.target/i386/stack-check-11.c
new file mode 100644
index 0000000..183103f
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/stack-check-11.c
@@ -0,0 +1,16 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fstack-clash-protection" } */
+/* { dg-require-effective-target supports_stack_clash_protection } */
+
+extern void arf (unsigned long int *, unsigned long int *);
+void
+frob ()
+{
+ unsigned long int num[859];
+ unsigned long int den[859];
+ arf (den, num);
+}
+
+/* { dg-final { scan-assembler-times "subq" 4 } } */
+/* { dg-final { scan-assembler-times "orq" 3 } } */
+
