rs6000: Don't touch below the stack pointer (PR77687)

With the 32-bit SVR4 ABI we don't have a red zone, so we have to restore
the callee-saved registers before we restore the stack pointer.

The previous fix for this PR failed in two ways, for huge frames: first,
we use a negative offset from r11 in that case, so the (mem:BLK 11) access
does no good; second, sched does not handle accesses to mem:BLK correctly
in this case (does not make dependencies).

This patch fixes it by doing a store to (mem:BLK (scratch)) instead.
This means no unrelated (not to stack) loads/stores can be moved over the
stack restore either, but so be it.

	PR target/77687
	* config/rs6000/rs6000.md (stack_restore_tie): Store to a scratch
	address instead of to r1 and r11.

gcc/testsuite/
	PR target/77687
	* gcc.target/powerpc/pr77687.c: New testcase.

From-SVN: r253033

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 1bbb3bf6eb43f1267eeabaa5994101382af78c8e..371b7902f3772e7ab0cc8d474d21caa263137081 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,9 @@
+2017-09-20  Segher Boessenkool  <segher@kernel.crashing.org>
+
+       PR target/77687
+       * config/rs6000/rs6000.md (stack_restore_tie): Store to a scratch
+       address instead of to r1 and r11.
+
 2017-09-20  Sebastian Peryt  <sebastian.peryt@intel.com>
 
        * config.gcc: Support "knm".

diff --git a/gcc/config/rs6000/rs6000.md b/gcc/config/rs6000/rs6000.md
index 7f17628387223d2ea29a17aa56c96afee1d13646..13ba743812476e6587403b2c80c80ba6d9ee5d8b 100644 (file)
--- a/gcc/config/rs6000/rs6000.md
+++ b/gcc/config/rs6000/rs6000.md
@@ -13161,14 +13161,12 @@
 
 ; Some 32-bit ABIs do not have a red zone, so the stack deallocation has to
 ; stay behind all restores from the stack, it cannot be reordered to before
-; one.  See PR77687.  This insn is an add or mr, and a stack_tie on the
-; operands of that.
+; one.  See PR77687.  This insn is an add or mr, and a memory clobber.
 (define_insn "stack_restore_tie"
   [(set (match_operand:SI 0 "gpc_reg_operand" "=r,r")
        (plus:SI (match_operand:SI 1 "gpc_reg_operand" "r,r")
                 (match_operand:SI 2 "reg_or_cint_operand" "O,rI")))
-   (set (mem:BLK (match_dup 0)) (const_int 0))
-   (set (mem:BLK (match_dup 1)) (const_int 0))]
+   (set (mem:BLK (scratch)) (const_int 0))]
   "TARGET_32BIT"
   "@
    mr %0,%1

diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 801adf0def9c4d1f1ed25f9ef5f891a5b90d1a95..eebb33d0b96d272093caa52b570c14c3bb4fcac0 100644 (file)
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2017-09-19  Segher Boessenkool  <segher@kernel.crashing.org>
+
+       PR target/77687
+       * gcc.target/powerpc/pr77687.c: New testcase.
+
 2017-09-20  Jakub Jelinek  <jakub@redhat.com>
 
        P0409R2 - allow lambda capture [=, this]

diff --git a/gcc/testsuite/gcc.target/powerpc/pr77687.c b/gcc/testsuite/gcc.target/powerpc/pr77687.c
new file mode 100644 (file)
index 0000000..95b27ae
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/pr77687.c
@@ -0,0 +1,20 @@
+/* { dg-options "-std=gnu99 -O2" } */
+/* { dg-final { scan-assembler-not {\mmr r?1,r?11\M.*11.*\mblr\M} } } */
+
+/* PR77687: We used to do stack accesses (via r11) after restoring r1.  */
+
+void g(int, char *);
+const char * dum = "hello";
+
+void f(int x)
+{
+       char big[200000];
+ start:
+       g(x, big);
+       g(x, big);
+       register void *p asm("r11") = &&start;
+       asm("" : : "r"(p));
+       asm("" : : :"r28");
+       asm("" : : :"r29");
+       asm("" : : :"r30");
+}