[PATCH, rtl-optimization]: Fix PR77452, ICE: in plus_constant, at explow.c

[PATCH, rtl-optimization]: Fix PR77452, ICE: in plus_constant, at explow.c

[Uros Bizjak]

[-- Attachment #1: Type: text/plain, Size: 822 bytes --]

Hello!

As shown in the PR [1], combine is able to simplify lowpart
CONST_VECTOR constant pool reference to its inner-mode reference.
However, plus_constant was not able to extract the constant from
narrowed access.

Attached patch teaches plus_constant how to handle this situation.

2016-09-04  Uros Bizjak  <ubizjak@gmail.com>

    PR rtl-optimization/77452
    * explow.c (plus_constant) <case MEM>: Extract scalar constant from
    inner-mode reference to a CONST_VECTOR constant in the constant pool.

2016-09-04  Uros Bizjak  <ubizjak@gmail.com>

testsuite/ChangeLog:

    PR rtl-optimization/77452
    * gcc.target/i386/pr77452.c: New test.

Patch was bootstrapped and regression tested on x86_64-linux-gnu {,-m32}.

OK for mainline and gcc-6 branch?

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77452

Uros.

[-- Attachment #2: p.diff.txt --]
[-- Type: text/plain, Size: 1441 bytes --]

Index: explow.c
===================================================================
--- explow.c	(revision 239975)
+++ explow.c	(working copy)
@@ -106,7 +106,15 @@ plus_constant (machine_mode mode, rtx x, HOST_WIDE
       if (GET_CODE (XEXP (x, 0)) == SYMBOL_REF
 	  && CONSTANT_POOL_ADDRESS_P (XEXP (x, 0)))
 	{
-	  tem = plus_constant (mode, get_pool_constant (XEXP (x, 0)), c);
+	  rtx cst = get_pool_constant (XEXP (x, 0));
+
+	  if (GET_CODE (cst) == CONST_VECTOR
+	      && GET_MODE_INNER (GET_MODE (cst)) == mode)
+	    {
+	      cst = gen_lowpart (mode, cst);
+	      gcc_assert (cst);
+	    }
+	  tem = plus_constant (mode, cst, c);
 	  tem = force_const_mem (GET_MODE (x), tem);
 	  /* Targets may disallow some constants in the constant pool, thus
 	     force_const_mem may return NULL_RTX.  */
Index: testsuite/gcc.target/i386/pr77452.c
===================================================================
--- testsuite/gcc.target/i386/pr77452.c	(nonexistent)
+++ testsuite/gcc.target/i386/pr77452.c	(working copy)
@@ -0,0 +1,13 @@
+/* { dg-do compile { target int128 } } */
+/* { dg-options "-O -mavx512f -fno-split-wide-types --param max-combine-insns=2" } */
+
+typedef unsigned      int U __attribute__((vector_size(64)));
+typedef unsigned __int128 V __attribute__((vector_size(64)));
+
+V
+foo(V v)
+{
+  v[0] = 1u << ((  ((V)(U){1, 1, v[0]})[0]) & 0xf)
+            >> ((-~((V)(U){1, 1, v[0]})[0]) & 0xf);
+  return v;
+}

Re: [PATCH, rtl-optimization]: Fix PR77452, ICE: in plus_constant, at explow.c

[Bernd Schmidt]

On 09/04/2016 02:12 PM, Uros Bizjak wrote:
> As shown in the PR [1], combine is able to simplify lowpart
> CONST_VECTOR constant pool reference to its inner-mode reference.
> However, plus_constant was not able to extract the constant from
> narrowed access.
>
> Attached patch teaches plus_constant how to handle this situation.

Ok. It looks like accesses to non-lowpart subregs would be handled by 
not optimizing them as well rather than crashing.


Bernd

Re: [PATCH, rtl-optimization]: Fix PR77452, ICE: in plus_constant, at explow.c

[Richard Biener]

On September 5, 2016 12:11:47 PM GMT+02:00, Bernd Schmidt <bschmidt@redhat.com> wrote:
>On 09/04/2016 02:12 PM, Uros Bizjak wrote:
>> As shown in the PR [1], combine is able to simplify lowpart
>> CONST_VECTOR constant pool reference to its inner-mode reference.
>> However, plus_constant was not able to extract the constant from
>> narrowed access.
>>
>> Attached patch teaches plus_constant how to handle this situation.
>
>Ok. It looks like accesses to non-lowpart subregs would be handled by 
>not optimizing them as well rather than crashing.

Are we missing similar handling for complex integer constants?

Richard.

>
>Bernd