Fix postreload_combine miscompilation (PR 69941)

Fix postreload_combine miscompilation (PR 69941)

[Bernd Schmidt]

[-- Attachment #1: Type: text/plain, Size: 416 bytes --]

This is a transformation which looks for

(set reg1 const)
(set reg2 (plus reg2 reg1))

and tries to replace all further uses of reg2 with (plus reg2 reg1). The 
problem here is that one of the uses is in a narrower mode, inside a 
zero_extend, so we produce wrong results.

The fix seems rather straightforward, verifying all uses have a mode 
matching the set. Bootstrapped and tested on x86_64-linux, ok?


Bernd

[-- Attachment #2: rcom-modes.diff --]
[-- Type: text/x-patch, Size: 2373 bytes --]

	PR rtl-optimization/69941
	* postreload.c (reload_combine_recognize_pattern): Ensure all uses of
	the reg share its mode.

testsuite/
	PR rtl-optimization/69941
	* gcc.dg/torture/pr69941.c: New test.

Index: gcc/postreload.c
===================================================================
--- gcc/postreload.c	(revision 233451)
+++ gcc/postreload.c	(working copy)
@@ -1057,7 +1057,6 @@ static bool
 reload_combine_recognize_pattern (rtx_insn *insn)
 {
   rtx set, reg, src;
-  unsigned int regno;
 
   set = single_set (insn);
   if (set == NULL_RTX)
@@ -1068,7 +1067,20 @@ reload_combine_recognize_pattern (rtx_in
   if (!REG_P (reg) || REG_NREGS (reg) != 1)
     return false;
 
-  regno = REGNO (reg);
+  unsigned int regno = REGNO (reg);
+  machine_mode mode = GET_MODE (reg);
+
+  if (reg_state[regno].use_index < 0
+      || reg_state[regno].use_index >= RELOAD_COMBINE_MAX_USES)
+    return false;
+
+  for (int i = reg_state[regno].use_index;
+       i < RELOAD_COMBINE_MAX_USES; i++)
+    {
+      struct reg_use *use = reg_state[regno].reg_use + i;
+      if (GET_MODE (*use->usep) != mode)
+	return false;
+    }
 
   /* Look for (set (REGX) (CONST_INT))
      (set (REGX) (PLUS (REGX) (REGY)))
@@ -1090,8 +1102,6 @@ reload_combine_recognize_pattern (rtx_in
       && REG_P (XEXP (src, 1))
       && rtx_equal_p (XEXP (src, 0), reg)
       && !rtx_equal_p (XEXP (src, 1), reg)
-      && reg_state[regno].use_index >= 0
-      && reg_state[regno].use_index < RELOAD_COMBINE_MAX_USES
       && last_label_ruid < reg_state[regno].use_ruid)
     {
       rtx base = XEXP (src, 1);
Index: gcc/testsuite/gcc.dg/torture/pr69941.c
===================================================================
--- gcc/testsuite/gcc.dg/torture/pr69941.c	(revision 0)
+++ gcc/testsuite/gcc.dg/torture/pr69941.c	(working copy)
@@ -0,0 +1,30 @@
+/* { dg-do run } */
+ 
+int a = 0;
+int b = 0;
+int c = 0;
+int e = 0;
+int f = 0;
+int *g = &e;
+ 
+int fn1() { return b ? a : b; }
+ 
+int main() {
+  int h = fn1() <= 0x8000000000000000ULL; // h = 1;
+ 
+  int k = f; // k = 0;
+ 
+  long i = h ? k : k / h; // i = 0;
+ 
+  long l = (unsigned short)(i - 0x1800); // l = 0xe800
+ 
+  i = l ? l : c; // i = 0xe800;
+ 
+  *g = i; // *g = 0xe800; e = 0xe800;
+ 
+  unsigned char result = e >> 9; // result = 0x74;
+
+  if ((int)result != 0x74)
+    __builtin_abort ();
+  return 0;
+}

Re: Fix postreload_combine miscompilation (PR 69941)

[Jeff Law]

On 03/04/2016 04:57 PM, Bernd Schmidt wrote:
> This is a transformation which looks for
>
> (set reg1 const)
> (set reg2 (plus reg2 reg1))
>
> and tries to replace all further uses of reg2 with (plus reg2 reg1). The
> problem here is that one of the uses is in a narrower mode, inside a
> zero_extend, so we produce wrong results.
>
> The fix seems rather straightforward, verifying all uses have a mode
> matching the set. Bootstrapped and tested on x86_64-linux, ok?
>
>
> Bernd
>
> rcom-modes.diff
>
>
> 	PR rtl-optimization/69941
> 	* postreload.c (reload_combine_recognize_pattern): Ensure all uses of
> 	the reg share its mode.
>
> testsuite/
> 	PR rtl-optimization/69941
> 	* gcc.dg/torture/pr69941.c: New test.
OK.

FWIW, based on my reading of the BZ, I think this counts as a regression 
-- it just happens to come and go.  I strongly suspect that's due to 
changing register assignments or some such.

I'm going to go ahead and push this to the trunk.

Jeff

Re: Fix postreload_combine miscompilation (PR 69941)

[Bernd Schmidt]

On 03/05/2016 06:27 AM, Jeff Law wrote:
>>     PR rtl-optimization/69941
>>     * postreload.c (reload_combine_recognize_pattern): Ensure all uses of
>>     the reg share its mode.
>>
>> testsuite/
>>     PR rtl-optimization/69941
>>     * gcc.dg/torture/pr69941.c: New test.
> OK.

For branches as well?


Bernd

Re: Fix postreload_combine miscompilation (PR 69941)

[Jeff Law]

On 03/07/2016 05:37 AM, Bernd Schmidt wrote:
> On 03/05/2016 06:27 AM, Jeff Law wrote:
>>>     PR rtl-optimization/69941
>>>     * postreload.c (reload_combine_recognize_pattern): Ensure all
>>> uses of
>>>     the reg share its mode.
>>>
>>> testsuite/
>>>     PR rtl-optimization/69941
>>>     * gcc.dg/torture/pr69941.c: New test.
>> OK.
>
> For branches as well?
Yes.

jeff