[PATCH] Fix combine's make_extraction (PR rtl-optimization/78378)

[PATCH] Fix combine's make_extraction (PR rtl-optimization/78378)

[Jakub Jelinek]

Hi!

If inner is a MEM, make_extraction requires that pos is a multiple of bytes
and deals with offsetting it.  Or otherwise requires that pos is a multiple
of BITS_PER_WORD and for REG inner it handles that too.  But if inner
is something different, it calls just force_to_mode to the target mode,
which only really works if pos is 0.

Thus the following patch restricts it to that case.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2016-11-16  Jakub Jelinek  <jakub@redhat.com>

	PR rtl-optimization/78378
	* combine.c (make_extraction): Use force_to_mode for non-{REG,MEM}
	inner only if pos is 0.  Fix up formatting.

	* gcc.c-torture/execute/pr78378.c: New test.

--- gcc/combine.c.jj	2016-11-16 18:51:58.000000000 +0100
+++ gcc/combine.c	2016-11-16 19:52:12.735674000 +0100
@@ -7382,6 +7382,7 @@ make_extraction (machine_mode mode, rtx
   if (tmode != BLKmode
       && ((pos_rtx == 0 && (pos % BITS_PER_WORD) == 0
 	   && !MEM_P (inner)
+	   && (pos == 0 || REG_P (inner))
 	   && (inner_mode == tmode
 	       || !REG_P (inner)
 	       || TRULY_NOOP_TRUNCATION_MODES_P (tmode, inner_mode)
@@ -7458,10 +7459,9 @@ make_extraction (machine_mode mode, rtx
 	}
       else
 	new_rtx = force_to_mode (inner, tmode,
-			     len >= HOST_BITS_PER_WIDE_INT
-			     ? HOST_WIDE_INT_M1U
-			     : (HOST_WIDE_INT_1U << len) - 1,
-			     0);
+				 len >= HOST_BITS_PER_WIDE_INT
+				 ? HOST_WIDE_INT_M1U
+				 : (HOST_WIDE_INT_1U << len) - 1, 0);
 
       /* If this extraction is going into the destination of a SET,
 	 make a STRICT_LOW_PART unless we made a MEM.  */
--- gcc/testsuite/gcc.c-torture/execute/pr78378.c.jj	2016-11-16 20:02:16.975248597 +0100
+++ gcc/testsuite/gcc.c-torture/execute/pr78378.c	2016-11-16 20:02:03.000000000 +0100
@@ -0,0 +1,18 @@
+/* PR rtl-optimization/78378 */
+
+unsigned long long __attribute__ ((noinline, noclone))
+foo (unsigned long long x)
+{
+  x <<= 41;
+  x /= 232;
+  return 1 + (unsigned short) x;
+}
+
+int
+main ()
+{
+  unsigned long long x = foo (1);
+  if (x != 0x2c24)
+    __builtin_abort();
+  return 0;
+}

	Jakub

Re: [PATCH] Fix combine's make_extraction (PR rtl-optimization/78378)

[Segher Boessenkool]

On Wed, Nov 16, 2016 at 10:07:23PM +0100, Jakub Jelinek wrote:
> If inner is a MEM, make_extraction requires that pos is a multiple of bytes
> and deals with offsetting it.  Or otherwise requires that pos is a multiple
> of BITS_PER_WORD and for REG inner it handles that too.  But if inner
> is something different, it calls just force_to_mode to the target mode,
> which only really works if pos is 0.
> 
> Thus the following patch restricts it to that case.
> 
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

Yes, thanks!


Segher

Re: [PATCH] Fix combine's make_extraction (PR rtl-optimization/78378)

[Michael Matz]

Hi,

On Wed, 16 Nov 2016, Jakub Jelinek wrote:

> If inner is a MEM, make_extraction requires that pos is a multiple of 
> bytes and deals with offsetting it.  Or otherwise requires that pos is a 
> multiple of BITS_PER_WORD and for REG inner it handles that too.  But if 
> inner is something different, it calls just force_to_mode to the target 
> mode, which only really works if pos is 0.

This should also fix the aarch64 fail for gcc.c-torture/execute/cbrt.c .  
(At least I came up with the same patch in PR78390)


Ciao,
Michael.