[PATCH, PR66444] Handle -fipa-ra in reload_combine

[PATCH, PR66444] Handle -fipa-ra in reload_combine

[Tom de Vries]

[-- Attachment #1: Type: text/plain, Size: 1360 bytes --]

Hi,

this patch fixes PR66444, a problem with -fipa-ra in reload_combine.

The problem is that for the test-case, reload_combine combines these two 
insns:
...
(insn 13 12 14 2 (parallel [
             (set (reg/v/f:DI 37 r8 [orig:92 xD.1858 ] [92])
		(plus:DI (reg:DI 37 r8 [96])
                     (reg:DI 0 ax [orig:95 D.1884 ] [95])))
             (clobber (reg:CC 17 flags))
         ])
      (expr_list:REG_EQUAL (plus:DI (reg:DI 0 ax [orig:95 D.1884 ] [95])
             (const_int 962072674304 [0xe000000000]))
         (nil)))
(insn 14 13 15 2 (set (reg:DI 5 di)
         (reg/v/f:DI 37 r8 [orig:92 xD.1858 ] [92]))
      (nil))
...

into this insn:
...
(insn 14 12 15 2 (set (reg:DI 5 di)
	(plus:DI (reg:DI 37 r8 [96])
             (reg:DI 0 ax [orig:95 D.1884 ] [95])))
      (nil))
...

That removes the set of r8 by insn 13. And that set of r8 is used by 
insn 16:
....
(call_insn 15 ...)
(insn 16 15 17 2 (set (reg:DI 5 di)
	(reg/v/f:DI 37 r8 [orig:92 x ] [92])) test.c:33 85 {*movdi_internal}
      (nil))
...

But reload_combine doesn't acknowledge that use, because it considers 
that r8 is killed by call_insn 15.

The patch fixes the problem by using get_call_reg_set_usage to find out 
that r8 is actually not killed by the call_insn.

Bootstrapped and reg-tested on x86_64 on top of trunk.

OK for trunk and gcc-5-branch?

Thanks,
- Tom

[-- Attachment #2: 0001-Handle-fipa-ra-in-reload_combine.patch --]
[-- Type: text/x-patch, Size: 1977 bytes --]

Handle -fipa-ra in reload_combine

2015-06-08  Tom de Vries  <tom@codesourcery.com>

	PR rtl-optimization/66444
	* postreload.c (reload_combine): Use get_call_reg_set_usage instead of
	call_used_regs.

	* gcc.dg/pr66444.c: New test.
---
 gcc/postreload.c               |  5 +++-
 gcc/testsuite/gcc.dg/pr66444.c | 58 ++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 62 insertions(+), 1 deletion(-)
 create mode 100644 gcc/testsuite/gcc.dg/pr66444.c

diff --git a/gcc/postreload.c b/gcc/postreload.c
index 7ecca15..1cc7b14 100644
--- a/gcc/postreload.c
+++ b/gcc/postreload.c
@@ -1352,9 +1352,12 @@ reload_combine (void)
       if (CALL_P (insn))
 	{
 	  rtx link;
+	  HARD_REG_SET used_regs;
+
+	  get_call_reg_set_usage (insn, &used_regs, call_used_reg_set);
 
 	  for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
-	    if (call_used_regs[r])
+	    if (TEST_HARD_REG_BIT (used_regs, r))
 	      {
 		reg_state[r].use_index = RELOAD_COMBINE_MAX_USES;
 		reg_state[r].store_ruid = reload_combine_ruid;
diff --git a/gcc/testsuite/gcc.dg/pr66444.c b/gcc/testsuite/gcc.dg/pr66444.c
new file mode 100644
index 0000000..93a7644
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pr66444.c
@@ -0,0 +1,58 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -fipa-ra" } */
+
+extern void abort (void);
+
+int __attribute__((noinline, noclone))
+bar (void)
+{
+  return 1;
+}
+
+struct S
+{
+  unsigned long p, q, r;
+  void *v;
+};
+
+struct S *s1;
+struct S *s2;
+
+void __attribute__((noinline, noclone))
+fn2 (struct S *x)
+{
+  s2 = x;
+}
+
+__attribute__((noinline, noclone)) void *
+fn1 (struct S *x)
+{
+  /* Just a statement to make it a non-const function.  */
+  s1 = x;
+
+  return (void *)0;
+}
+
+int __attribute__((noinline, noclone))
+baz (void)
+{
+  struct S *x = (struct S *) 0xe0000000U;
+
+  x += bar ();
+
+  fn1 (x);
+  fn2 (x);
+
+  return 0;
+}
+
+int
+main (void)
+{
+  baz ();
+
+  if (s2 != (((struct S *) 0xe0000000U) + 1))
+    abort ();
+
+  return 0;
+}
-- 
1.9.1

Re: [PATCH, PR66444] Handle -fipa-ra in reload_combine

[Jakub Jelinek]

On Mon, Jun 08, 2015 at 02:04:12PM +0200, Tom de Vries wrote:
> this patch fixes PR66444, a problem with -fipa-ra in reload_combine.
> 
> The problem is that for the test-case, reload_combine combines these two
> insns:

Please work out with Vlad whether reload_cse_move2add doesn't need similar
fix (and check other spots too).

> 2015-06-08  Tom de Vries  <tom@codesourcery.com>
> 
> 	PR rtl-optimization/66444
> 	* postreload.c (reload_combine): Use get_call_reg_set_usage instead of
> 	call_used_regs.

LGTM.

> 	* gcc.dg/pr66444.c: New test.

> +int __attribute__((noinline, noclone))
> +baz (void)
> +{
> +  struct S *x = (struct S *) 0xe0000000U;

I'm still afraid this will not really work on s390-linux (which has only
31-bit pointers) and will not work on 16-bit int targets either
(some have say 24-bit pointers etc., not really familiar with the embedded
world).
So, I'd suggest use a macro for the address, so you don't need to duplicate
it, and define it to say ((struct S *) 0x8000UL), if it reproduces
even with that change without your reload_combine fix.

Ok for trunk and 5.2 with that change.

	Jakub

Re: [PATCH, PR66444] Handle -fipa-ra in reload_combine

[Tom de Vries]

[-- Attachment #1: Type: text/plain, Size: 1607 bytes --]

On 08/06/15 17:31, Jakub Jelinek wrote:
> On Mon, Jun 08, 2015 at 02:04:12PM +0200, Tom de Vries wrote:
>> this patch fixes PR66444, a problem with -fipa-ra in reload_combine.
>>
>> The problem is that for the test-case, reload_combine combines these two
>> insns:
>
> Please work out with Vlad whether reload_cse_move2add doesn't need similar
> fix (and check other spots too).
>
>> 2015-06-08  Tom de Vries  <tom@codesourcery.com>
>>
>> 	PR rtl-optimization/66444
>> 	* postreload.c (reload_combine): Use get_call_reg_set_usage instead of
>> 	call_used_regs.
>
> LGTM.
>
>> 	* gcc.dg/pr66444.c: New test.
>
>> +int __attribute__((noinline, noclone))
>> +baz (void)
>> +{
>> +  struct S *x = (struct S *) 0xe0000000U;
>
> I'm still afraid this will not really work on s390-linux (which has only
> 31-bit pointers) and will not work on 16-bit int targets either
> (some have say 24-bit pointers etc., not really familiar with the embedded
> world).
> So, I'd suggest use a macro for the address, so you don't need to duplicate
> it,

Used a macro CONST_PTR.

> and define it to say ((struct S *) 0x8000UL), if it reproduces
> even with that change without your reload_combine fix.

Unfortunately, it didn't. So I used __SIZEOF_POINTER__ to produce a 
valid constant pointer for the 16-31 bit pointer-size cases, while still 
triggering the original problem for x86_64 -m64 case using a larger pointer.

Furthermore, I used __SIZEOF_POINTER__ to ensured a valid pointer 
constant suffix.

>
> Ok for trunk and 5.2 with that change.

Committed to trunk and backported gcc-5-branch as attached.

Thanks,
- Tom

[-- Attachment #2: 0001-Handle-fipa-ra-in-reload_combine.patch --]
[-- Type: text/x-patch, Size: 2733 bytes --]

Handle -fipa-ra in reload_combine

2015-06-08  Tom de Vries  <tom@codesourcery.com>

	PR rtl-optimization/66444
	* postreload.c (reload_combine): Use get_call_reg_set_usage instead of
	call_used_regs.

	* gcc.dg/pr66444.c: New test.
---
 gcc/postreload.c               |  5 ++-
 gcc/testsuite/gcc.dg/pr66444.c | 79 ++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 83 insertions(+), 1 deletion(-)
 create mode 100644 gcc/testsuite/gcc.dg/pr66444.c

diff --git a/gcc/postreload.c b/gcc/postreload.c
index 7ecca15..1cc7b14 100644
--- a/gcc/postreload.c
+++ b/gcc/postreload.c
@@ -1352,9 +1352,12 @@ reload_combine (void)
       if (CALL_P (insn))
 	{
 	  rtx link;
+	  HARD_REG_SET used_regs;
+
+	  get_call_reg_set_usage (insn, &used_regs, call_used_reg_set);
 
 	  for (r = 0; r < FIRST_PSEUDO_REGISTER; r++)
-	    if (call_used_regs[r])
+	    if (TEST_HARD_REG_BIT (used_regs, r))
 	      {
 		reg_state[r].use_index = RELOAD_COMBINE_MAX_USES;
 		reg_state[r].store_ruid = reload_combine_ruid;
diff --git a/gcc/testsuite/gcc.dg/pr66444.c b/gcc/testsuite/gcc.dg/pr66444.c
new file mode 100644
index 0000000..3f92a5c
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pr66444.c
@@ -0,0 +1,79 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -fipa-ra" } */
+
+extern void abort (void);
+
+#if (__SIZEOF_LONG_LONG__ == __SIZEOF_POINTER__)
+#define ADD_SUFFIX(a) a ## ULL
+#elif (__SIZEOF_LONG__ == __SIZEOF_POINTER__)
+#define ADD_SUFFIX(a) a ## UL
+#elif (__SIZEOF_INT__ == __SIZEOF_POINTER__)
+#define ADD_SUFFIX(a) a ## U
+#else
+#error Add target support here
+#endif
+
+#if __SIZEOF_POINTER__ <= 4
+/* Use a 16 bit pointer to have a valid pointer for 16-bit to 31-bit pointer
+   architectures.  Using sizeof, we cannot distinguish between 31-bit and 32-bit
+   pointer types, so we also handle the 32-bit pointer type case here.  */
+#define CONST_PTR ADD_SUFFIX (0x800)
+#else
+/* For x86_64 -m64, the problem reproduces with this 32-bit CONST_PTR, but not
+   with a 2-power below it.  */
+#define CONST_PTR ADD_SUFFIX (0x80000000)
+#endif
+
+int __attribute__((noinline, noclone))
+bar (void)
+{
+  return 1;
+}
+
+struct S
+{
+  unsigned long p, q, r;
+  void *v;
+};
+
+struct S *s1;
+struct S *s2;
+
+void __attribute__((noinline, noclone))
+fn2 (struct S *x)
+{
+  s2 = x;
+}
+
+__attribute__((noinline, noclone)) void *
+fn1 (struct S *x)
+{
+  /* Just a statement to make it a non-const function.  */
+  s1 = x;
+
+  return (void *)0;
+}
+
+int __attribute__((noinline, noclone))
+baz (void)
+{
+  struct S *x = (struct S *) CONST_PTR;
+
+  x += bar ();
+
+  fn1 (x);
+  fn2 (x);
+
+  return 0;
+}
+
+int
+main (void)
+{
+  baz ();
+
+  if (s2 != (((struct S *) CONST_PTR) + 1))
+    abort ();
+
+  return 0;
+}
-- 
1.9.1

Re: [PATCH, PR66444] Handle -fipa-ra in reload_combine

[Tom de Vries]

On 08/06/15 17:31, Jakub Jelinek wrote:
> On Mon, Jun 08, 2015 at 02:04:12PM +0200, Tom de Vries wrote:
>> >this patch fixes PR66444, a problem with -fipa-ra in reload_combine.
>> >
>> >The problem is that for the test-case, reload_combine combines these two
>> >insns:
> Please work out with Vlad whether reload_cse_move2add doesn't need similar
> fix (and check other spots too).
>

Filed PR66463 - review uses of call_used_regs and regs_invalidated_by_call.

Thanks,
- Tom