[PATCH] explow, aarch64: Fix force-Pmode-to-mem for ILP32 [PR97269]

[PATCH] explow, aarch64: Fix force-Pmode-to-mem for ILP32 [PR97269]

[Richard Sandiford]

This patch fixes a mode/rtx mismatch for ILP32 targets in:

	  mem = force_const_mem (ptr_mode, imm);

where imm can be Pmode rather than ptr_mode.

The patch uses convert_memory_address to convert the Pmode address
to ptr_mode before the call.  However, immediate addresses can in
general contain unspecs, and convert_memory_address wasn't set up
to handle those.

The patch therefore adds some generic unspec handling to
convert_memory_address_addr_space_1.  As the comment says, we can add
a target hook if this behaviour turns out to be wrong for some targets.
But I think what the patch does is a strict improvement over the status
quo: without it, we would try to force the unspec into a register,
but nevertheless wrap the result in a (const ...).  That in turn
would be invalid rtl and seems bound to generate an ICE later.

I tested the explow.c part using -fstack-protector with local hacks
to force SYMBOL_FORCE_TO_MEM for UNSPEC_SALT_ADDR.

Fixes c-c++-common/torture/pr57945.c and various other tests.

Tested on aarch64-linux-gnu, aarch64_be-elf and x86_64-linux-gnu.
OK for the explow bits?

Richard


gcc/
	PR target/97269
	* explow.c (convert_memory_address_addr_space_1): Handle UNSPECs
	nested in CONSTs.
	* config/aarch64/aarch64.c (aarch64_expand_mov_immediate): Use
	convert_memory_address to convert symbolic immediates to ptr_mode
	before forcing them to memory.
---
 gcc/config/aarch64/aarch64.c |  5 ++++-
 gcc/explow.c                 | 20 ++++++++++++++++++++
 2 files changed, 24 insertions(+), 1 deletion(-)

diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index ed09fe0c6f8..d3ca6dc037e 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -5224,8 +5224,11 @@ aarch64_expand_mov_immediate (rtx dest, rtx imm)
       switch (sty)
 	{
 	case SYMBOL_FORCE_TO_MEM:
+	  if (int_mode != ptr_mode)
+	    imm = convert_memory_address (ptr_mode, imm);
+
 	  if (const_offset != 0
-	      && targetm.cannot_force_const_mem (int_mode, imm))
+	      && targetm.cannot_force_const_mem (ptr_mode, imm))
 	    {
 	      gcc_assert (can_create_pseudo_p ());
 	      base = aarch64_force_temporary (int_mode, dest, base);
diff --git a/gcc/explow.c b/gcc/explow.c
index 65f904a3d14..b8c9fbed03e 100644
--- a/gcc/explow.c
+++ b/gcc/explow.c
@@ -378,6 +378,26 @@ convert_memory_address_addr_space_1 (scalar_int_mode to_mode ATTRIBUTE_UNUSED,
 	}
       break;
 
+    case UNSPEC:
+      /* Assume that all UNSPECs in a constant address can be converted
+	 operand-by-operand.  We could add a target hook if some targets
+	 require different behavior.  */
+      if (in_const && GET_MODE (x) == from_mode)
+	{
+	  unsigned int n = XVECLEN (x, 0);
+	  rtvec v = gen_rtvec (n);
+	  for (unsigned int i = 0; i < n; ++i)
+	    {
+	      rtx op = XVECEXP (x, 0, i);
+	      if (GET_MODE (op) == from_mode)
+		op = convert_memory_address_addr_space_1 (to_mode, op, as,
+							  in_const, no_emit);
+	      RTVEC_ELT (v, i) = op;
+	    }
+	  return gen_rtx_UNSPEC (to_mode, v, XINT (x, 1));
+	}
+      break;
+
     default:
       break;
     }

Re: [PATCH] explow, aarch64: Fix force-Pmode-to-mem for ILP32 [PR97269]

[Jeff Law]

On 1/4/21 4:46 AM, Richard Sandiford via Gcc-patches wrote:
> This patch fixes a mode/rtx mismatch for ILP32 targets in:
>
> 	  mem = force_const_mem (ptr_mode, imm);
>
> where imm can be Pmode rather than ptr_mode.
>
> The patch uses convert_memory_address to convert the Pmode address
> to ptr_mode before the call.  However, immediate addresses can in
> general contain unspecs, and convert_memory_address wasn't set up
> to handle those.
>
> The patch therefore adds some generic unspec handling to
> convert_memory_address_addr_space_1.  As the comment says, we can add
> a target hook if this behaviour turns out to be wrong for some targets.
> But I think what the patch does is a strict improvement over the status
> quo: without it, we would try to force the unspec into a register,
> but nevertheless wrap the result in a (const ...).  That in turn
> would be invalid rtl and seems bound to generate an ICE later.
>
> I tested the explow.c part using -fstack-protector with local hacks
> to force SYMBOL_FORCE_TO_MEM for UNSPEC_SALT_ADDR.
>
> Fixes c-c++-common/torture/pr57945.c and various other tests.
>
> Tested on aarch64-linux-gnu, aarch64_be-elf and x86_64-linux-gnu.
> OK for the explow bits?
>
> Richard
>
>
> gcc/
> 	PR target/97269
> 	* explow.c (convert_memory_address_addr_space_1): Handle UNSPECs
> 	nested in CONSTs.
> 	* config/aarch64/aarch64.c (aarch64_expand_mov_immediate): Use
> 	convert_memory_address to convert symbolic immediates to ptr_mode
> 	before forcing them to memory.
OK.    I can't think of a case where the operand-by-operand conversion
wouldn't work, but as we both know, targets sometimes to "odd" things.  
So a conditional ACK and we'll see how the various targets in the tester
respond.

jeff