Bug 88001

My solution, on a side branch, is this patch, but we need it to run also for mainline gcc as we cannot build glibc or uclibc toolchains. Any help is appreciated.

---
 gcc/function.c | 34 +++++++++++++++++++++++++++++++++-
 1 file changed, 33 insertions(+), 1 deletion(-)

diff --git a/gcc/function.c b/gcc/function.c
index 302438323c8..36227f77074 100644
--- a/gcc/function.c
+++ b/gcc/function.c
@@ -6374,6 +6374,37 @@ make_pass_thread_prologue_and_epilogue (gcc::context *ctxt)
 }
 
 
+/* Helper match_asm_constraints_1.  */
+static int
+constant_overlap_mentioned_p (const_rtx x, const_rtx in)
+{
+  const char *fmt;
+  int i, j;
+
+  if (CONST_INT_P (in))
+    return 0;
+
+  if (!CONSTANT_P (in))
+    return 0;
+
+  if (x == 0)
+    return 0;
+
+  if (x == in)
+    return 1;
+
+  fmt = GET_RTX_FORMAT (GET_CODE (x));
+  for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
+    {
+      if (fmt[i] == 'e')
+	 return constant_overlap_mentioned_p (XEXP (x, i), in);
+      else if (fmt[i] == 'E')
+	for (j = XVECLEN (x, i) - 1; j >= 0; j--)
+	  return constant_overlap_mentioned_p (XVECEXP (x, i, j), in);
+    }
+  return 0;
+}
+
 /* This mini-pass fixes fall-out from SSA in asm statements that have
    in-out constraints.  Say you start with
 
@@ -6509,7 +6540,8 @@ match_asm_constraints_1 (rtx_insn *insn, rtx *p_sets, int noutputs)
 	  SET_DEST (p_sets[j]) = replace_rtx (SET_DEST (p_sets[j]),
 					      input, output);
       for (j = 0; j < ninputs; j++)
-	if (reg_overlap_mentioned_p (input, RTVEC_ELT (inputs, j)))
+	if (reg_overlap_mentioned_p (input, RTVEC_ELT (inputs, j))
+	    || constant_overlap_mentioned_p (RTVEC_ELT (inputs, j), input))
 	  RTVEC_ELT (inputs, j) = replace_rtx (RTVEC_ELT (inputs, j),
 					       input, output);
 
-- 
2.19.1

Ah yes, I had release checking there.  Confirmed.

Created attachment 45210 [details]
proposed patch

Claudiu: could you test that patch please?  (On the real thing, not just
this testcase :-) )

Hi, I'm the ARC glibc maintainer (and desperate for this fix).

The proposed (one liner) patch seems to fix the compiler assert for sue. I'm still running more tests to make sure it is functional but looks promising.

The dejagnu tests for ARC look alright. No extra failures.

Patch posted at https://gcc.gnu.org/ml/gcc-patches/2018-12/msg00865.html .

Author: segher
Date: Fri Dec 14 08:29:34 2018
New Revision: 267122

URL: https://gcc.gnu.org/viewcvs?rev=267122&root=gcc&view=rev
Log:
match_asm_constraints: Use copy_rtx where needed (PR88001)

The new insn here (temporarily) illegally shares RTL.  This fixes it.


	PR rtl-optimization/88001
	* function.c (match_asm_constraints_1): Don't invalidly share RTL.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/function.c

Can this be stable backported to gcc-8-branch as well.
glibc folks use that branch for their regular smoke testing and without that ARC tools don't even build.

Sure this can be backported...  But can you fill in known-to-{work,fail}
then please?  Thanks.

Sure, but how can I ? if i click the "known to work" field it takes me to help.

The issue certainly with gcc-8-branch for ARC and presumably also with tip/trunk.

Author: segher
Date: Sat Dec 15 12:05:08 2018
New Revision: 267171

URL: https://gcc.gnu.org/viewcvs?rev=267171&root=gcc&view=rev
Log:
	Backport from trunk
	2018-12-14  Segher Boessenkool  <segher@kernel.crashing.org>

	PR rtl-optimization/88001
	* function.c (match_asm_constraints_1): Don't invalidly share RTL.

Modified:
    branches/gcc-8-branch/gcc/ChangeLog
    branches/gcc-8-branch/gcc/function.c

Author: segher
Date: Sat Dec 15 12:07:42 2018
New Revision: 267172

URL: https://gcc.gnu.org/viewcvs?rev=267172&root=gcc&view=rev
Log:
	Backport from trunk
	2018-12-14  Segher Boessenkool  <segher@kernel.crashing.org>

	PR rtl-optimization/88001
	* function.c (match_asm_constraints_1): Don't invalidly share RTL.

Modified:
    branches/gcc-7-branch/gcc/ChangeLog
    branches/gcc-7-branch/gcc/function.c

Ah, you don't have permission to edit the field.  I backported it to GCC 8
and GCC 7; closing now.