Re: [PATCH] PR70674: S/390: Add memory barrier to stack pointer restore from fpr.

[Andreas Krebbel]

On 04/19/2016 12:54 PM, Jakub Jelinek wrote:
> Can you please:
>   rtx fpr = gen_rtx_REG (DImode, cfun_gpr_save_slot (i));
>   if (i == STACK_POINTER_REGNUM)
> insn = emit_insn (gen_stack_restore_from_fpr (fpr));
>   else
> insn = emit_move_insn (gen_rtx_REG (DImode, i), fpr);
> That way IMHO it is more nicely formatted, you avoid the ugly (
> at the end of line, it uses fewer lines anyway and additionally
> you can make it clear what the gen_rtx_REG (DImode, cfun_gpr_save_slot (i))
> means by giving it a name.  Of course, choose whatever other var
> name you prefer to describe what it is.

Right, that's better. I'll change the patch and commit it tomorrow. Thanks!

-Andreas-

Re: [PATCH] PR70674: S/390: Add memory barrier to stack pointer restore from fpr.

[Jakub Jelinek]

On Tue, Apr 19, 2016 at 11:02:34AM +0200, Andreas Krebbel wrote:
> I'll post the patches for the other two parts when gcc 7 entered stage
> 1 again.

It will not reenter stage 1 again, that happened last Friday ;)

> This needs to go into 4.9/5/6 branches.

Ok for 6, but I have formatting nit:

> +  rtx_insn *insn;
> +
> +  if (!FP_REGNO_P (cfun_gpr_save_slot (i)))
> + continue;
> +

Can you please:
rtx fpr = gen_rtx_REG (DImode, cfun_gpr_save_slot (i));
if (i == STACK_POINTER_REGNUM)
  insn = emit_insn (gen_stack_restore_from_fpr (fpr));
else
  insn = emit_move_insn (gen_rtx_REG (DImode, i), fpr);
That way IMHO it is more nicely formatted, you avoid the ugly (
at the end of line, it uses fewer lines anyway and additionally
you can make it clear what the gen_rtx_REG (DImode, cfun_gpr_save_slot (i))
means by giving it a name.  Of course, choose whatever other var
name you prefer to describe what it is.

> +  if (i == STACK_POINTER_REGNUM)
> + insn = emit_insn (gen_stack_restore_from_fpr (
> + gen_rtx_REG (DImode, cfun_gpr_save_slot (i;
> +  else
> + insn =
> +   emit_move_insn (gen_rtx_REG (DImode, i),
> +   gen_rtx_REG (DImode, cfun_gpr_save_slot (i)));

Jakub

[PATCH] PR70674: S/390: Add memory barrier to stack pointer restore from fpr.

[Andreas Krebbel]

This patches fixes a problem with stack variable accesses being
scheduled after the stack pointer restore instructions.  In the
testcase this happened with the stack variable 'a' accessed through the
frame pointer.

The existing stack_tie we have in the backend is basically useless
when trying to block stack variable accesses from being scheduled
across an insn.  The alias set of stack variables and the frame alias
set usually differ and hence aren't in conflict with each other.  The
solution appears to be a magic MEM term with a scratch register which
is handled as a full memory barrier when analyzing scheduling
dependencies.

With the patch a (clobber (mem:BLK (scratch))) is being added to the
restore instruction in order to prevent any memory operations to be
scheduled across the insn.  The patch does that only for the one case
where the stack pointer is restored from an FPR.  Theoretically this
might happen also in the case where the stack pointer gets restored
using a load multiple.  However, triggering that problem with
load-multiple appears to be much harder since the load-multiple will
restore the frame pointer as well.  So in order to see the problem a
different call-clobbered register would need to be used as temporary
stack pointer.

Another case which needs to be handled some day is the stack pointer
allocation part.  It needs to be a memory barrier as well.

I'll post the patches for the other two parts when gcc 7 entered stage
1 again.

Bootstrapped and regression tested with --with-arch z196 and z13 on
s390 and s390x.

This needs to go into 4.9/5/6 branches.

-Andreas-

gcc/ChangeLog:

2016-04-19  Andreas Krebbel  

PR target/70674
* config/s390/s390.c (s390_restore_gprs_from_fprs): Pick the new
stack_restore_from_fpr pattern when restoring r15.
(s390_optimize_prologue): Strip away the memory barrier in the
parallel when trying to get rid of restore insns.
* config/s390/s390.md ("stack_restore_from_fpr"): New insn
definition for loading the stack pointer from an FPR.  Compared to
the normal move insn this pattern includes a full memory barrier.

gcc/testsuite/ChangeLog:

2016-04-19  Andreas Krebbel  

PR target/70674
* gcc.target/s390/pr70674.c: New test.
---
 gcc/config/s390/s390.c  | 91 +++--
 gcc/config/s390/s390.md | 10 
 gcc/testsuite/gcc.target/s390/pr70674.c | 13 +
 3 files changed, 76 insertions(+), 38 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/s390/pr70674.c

diff --git a/gcc/config/s390/s390.c b/gcc/config/s390/s390.c
index 1134d0f..e969542 100644
--- a/gcc/config/s390/s390.c
+++ b/gcc/config/s390/s390.c
@@ -10538,19 +10538,25 @@ s390_restore_gprs_from_fprs (void)
 
   for (i = 6; i < 16; i++)
 {
-  if (FP_REGNO_P (cfun_gpr_save_slot (i)))
-   {
- rtx_insn *insn =
-   emit_move_insn (gen_rtx_REG (DImode, i),
-   gen_rtx_REG (DImode, cfun_gpr_save_slot (i)));
- df_set_regs_ever_live (i, true);
- add_reg_note (insn, REG_CFA_RESTORE, gen_rtx_REG (DImode, i));
- if (i == STACK_POINTER_REGNUM)
-   add_reg_note (insn, REG_CFA_DEF_CFA,
- plus_constant (Pmode, stack_pointer_rtx,
-STACK_POINTER_OFFSET));
- RTX_FRAME_RELATED_P (insn) = 1;
-   }
+  rtx_insn *insn;
+
+  if (!FP_REGNO_P (cfun_gpr_save_slot (i)))
+   continue;
+
+  if (i == STACK_POINTER_REGNUM)
+   insn = emit_insn (gen_stack_restore_from_fpr (
+   gen_rtx_REG (DImode, cfun_gpr_save_slot (i;
+  else
+   insn =
+ emit_move_insn (gen_rtx_REG (DImode, i),
+ gen_rtx_REG (DImode, cfun_gpr_save_slot (i)));
+  df_set_regs_ever_live (i, true);
+  add_reg_note (insn, REG_CFA_RESTORE, gen_rtx_REG (DImode, i));
+  if (i == STACK_POINTER_REGNUM)
+   add_reg_note (insn, REG_CFA_DEF_CFA,
+ plus_constant (Pmode, stack_pointer_rtx,
+STACK_POINTER_OFFSET));
+  RTX_FRAME_RELATED_P (insn) = 1;
 }
 }
 
@@ -13032,37 +13038,46 @@ s390_optimize_prologue (void)
 
   /* Remove ldgr/lgdr instructions used for saving and restore
 GPRs if possible.  */
-  if (TARGET_Z10
- && GET_CODE (pat) == SET
- && GET_MODE (SET_SRC (pat)) == DImode
- && REG_P (SET_SRC (pat))
- && REG_P (SET_DEST (pat)))
+  if (TARGET_Z10)
{
- int src_regno = REGNO (SET_SRC (pat));
- int dest_regno = REGNO (SET_DEST (pat));
- int gpr_regno;
- int fpr_regno;
+ rtx tmp_pat = pat;
 
- if (!((GENERAL_REGNO_P (src_regno) && FP_REGNO_P (dest_regno))
-   || (FP_REGNO_P (src_regno) && GENERAL_REGNO_P (dest_regno
-   continue;