This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [Patch ARM] Fix PR71778


Hi James,

On 12/06/17 14:57, James Greenhalgh wrote:
Hi,

PR71778 is an ICE when you pass a non-constant argument to an intrinsic
which requires a constant.

This ICE was introduced after we rewrote some of the builtin handling for
Neon intrinsics, the issue is that after throwing an error in
arm_expand_builtin_args, we return const0_rtx to indicate the expand
has failed

       if (!(*insn_data[icode].operand[opno].predicate)
	  (op[argc], mode[argc]))
	{
	  error ("%Kargument %d must be a constant immediate",
		 exp, argc + 1);
	  return const0_rtx;
	}

At this point we're safely in to invalid code, but the mid-end continues
trying to resolve the assignment, with const0_rtx on the right-hand
side. That gets in to trouble in movv2di, which sees the constant and
tries to expand through neon_make_constant, which doesn't expect to see
a const0_rtx in the assignment (it wants a vector), and we hit a
gcc_unreachable () and take the ICE.

That movv2di expander is the one in vec-common.md that ends up calling
neon_make_constant. I wonder why const0_rtx passed its predicate check
(that would require a V2DImode vector of zeroes rather than a const0_rtx).
Perhaps the midend code at this point doesn't check the operand predicate.

In the builtin expansion code that you quoted I wonder wonder if we could fail
more gracefully by returning CONST0_RTX (mode[argc]) to match the expected
mode of the operand (we've already emitted an error, so we shouldn't care
what RTL we emit as long as it doesn't cause an ICE).

There are a few moving parts in the back end, so it isn't clear to me that
the fix I've come up with is 100% in the right place. AArch64 doesn't
bother with a similar construct, expanding straight to a mov with whatever
you've given it, so I don't see a right place over there.

The change is defensible, but I don't really know the ARM back end.

Bootstrapped on arm-none-linux-gnueabihf.

OK?

Thanks,
James

---
gcc/

2017-06-12  James Greenhalgh  <james.greenhalgh@arm.com>

	PR target/71778
	* config/arm/arm.c (neon_make_constant): Return const0_rtx for
	const0_rtx input.

gcc/testsuite/

2017-06-12  James Greenhalgh  <james.greenhalgh@arm.com>

	PR target/71778
	* gcc.target/arm/pr71778.c: New.


diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index e503891..b8d59c6 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -12124,6 +12124,11 @@ neon_make_constant (rtx vals)
       if (n_const == n_elts)
 	const_vec = gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0));
     }
+  else if (vals == const0_rtx)
+    /* Something invalid, perhaps from expanding an intrinsic
+       which requires a constant argument, where a variable argument
+       was passed.  */
+     return const0_rtx;
   else
     gcc_unreachable ();

I'm not a fan of this as the function has a precondition that its argument is a PARALLEL or a CONST_VECTOR
and special-casing const0_rtx breaks that. I'd rather we tried fixing this closer to the error source.
Can you try the suggestion above instead please?

Thanks,
Kyrill


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]