[PATCH] regrename: Fix -fcompare-debug issue in check_new_reg_p [PR105041]

[PATCH] regrename: Fix -fcompare-debug issue in check_new_reg_p [PR105041]

[Surya Kumari Jangala]

regrename: Fix -fcompare-debug issue in check_new_reg_p [PR105041]

In check_new_reg_p, the nregs of a du chain is computed by obtaining the MODE
of the first element in the chain, and then calling hard_regno_nregs() with the
MODE. But the first element of the chain can be a DEBUG_INSN whose mode need
not be the same as the rest of the elements in the du chain. This
was resulting in fcompare-debug failure as check_new_reg_p was returning a
different result with -g for the same candidate register. We can instead obtain
nregs from the du chain itself.

2022-06-10  Surya Kumari Jangala  <jskumari@linux.ibm.com>

gcc/
	PR rtl-optimization/105041
	* regrename.cc (check_new_reg_p): Use nregs value from du chain.

gcc/testsuite/
	PR rtl-optimization/105041
	* gcc.target/powerpc/pr105041.c: New test.


diff --git a/gcc/regrename.cc b/gcc/regrename.cc
index 10271e1..f651351 100644
--- a/gcc/regrename.cc
+++ b/gcc/regrename.cc
@@ -324,8 +324,7 @@ static bool
 check_new_reg_p (int reg ATTRIBUTE_UNUSED, int new_reg,
 		 class du_head *this_head, HARD_REG_SET this_unavailable)
 {
-  machine_mode mode = GET_MODE (*this_head->first->loc);
-  int nregs = hard_regno_nregs (new_reg, mode);
+  int nregs = this_head->nregs;
   int i;
   struct du_chain *tmp;
 
diff --git a/gcc/testsuite/gcc.target/powerpc/pr105041.c b/gcc/testsuite/gcc.target/powerpc/pr105041.c
new file mode 100644
index 0000000..89eed1c
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/pr105041.c
@@ -0,0 +1,24 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target be } */
+/* { dg-options "-m32 -mdejagnu-cpu=power4 -O2 -fcompare-debug -fharden-compares -frename-registers" } */
+
+double m;
+int n;
+
+unsigned int
+foo (unsigned int x, int y)
+{
+  long long int a = y, b = !a;
+  int c = 0;
+
+  if (b != x)
+    while ((int) m == a)
+      {
+        c = a;
+        a = 0;
+      }
+
+  n = b = y;
+
+  return x + c;
+}

Re: [PATCH] regrename: Fix -fcompare-debug issue in check_new_reg_p [PR105041]

[Segher Boessenkool]

Hi!

On Fri, Jun 10, 2022 at 07:52:57PM +0530, Surya Kumari Jangala wrote:
> In check_new_reg_p, the nregs of a du chain is computed by obtaining the MODE
> of the first element in the chain, and then calling hard_regno_nregs() with the
> MODE. But the first element of the chain can be a DEBUG_INSN whose mode need
> not be the same as the rest of the elements in the du chain. This
> was resulting in fcompare-debug failure as check_new_reg_p was returning a
> different result with -g for the same candidate register. We can instead obtain
> nregs from the du chain itself.

Great, thanks for finding and fixing this!  I cannot approve it, you'll
have to wait for someone who can.  It looks fine to me, but that does
not mean so much in regrename.c :-)

> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/powerpc/pr105041.c
> @@ -0,0 +1,24 @@
> +/* { dg-do compile } */

Please delete this line, it is the default.

> +/* { dg-require-effective-target be } */

Is there a reason to not test this on LE?  If not, please remove this
line as well.

> +/* { dg-options "-m32 -mdejagnu-cpu=power4 -O2 -fcompare-debug -fharden-compares -frename-registers" } */

Aha.  You check for LE because you use -m32 in the test?  Don't, then!
Instead, test with -m32 in your RUNTESTFLAGS, like
  make check-gcc-c RUNTESTFLAGS="--target_board=unix'{-m64,-m32}' powerpc.exp=pr105041.c"
or similar.

It's a good idea to add a comment a la
/* PR rtl-optimization/105041: This test failed with -m32.  */

Thanks again for the patch!


Segher

Re: [PATCH] regrename: Fix -fcompare-debug issue in check_new_reg_p [PR105041]

[Jeff Law]

On 6/10/2022 9:40 AM, Segher Boessenkool wrote:
> Hi!
>
> On Fri, Jun 10, 2022 at 07:52:57PM +0530, Surya Kumari Jangala wrote:
>> In check_new_reg_p, the nregs of a du chain is computed by obtaining the MODE
>> of the first element in the chain, and then calling hard_regno_nregs() with the
>> MODE. But the first element of the chain can be a DEBUG_INSN whose mode need
>> not be the same as the rest of the elements in the du chain. This
>> was resulting in fcompare-debug failure as check_new_reg_p was returning a
>> different result with -g for the same candidate register. We can instead obtain
>> nregs from the du chain itself.
> Great, thanks for finding and fixing this!  I cannot approve it, you'll
> have to wait for someone who can.  It looks fine to me, but that does
> not mean so much in regrename.c :-)
I'll go ahead and ACK the regrename bits.  So as soon as you're happy 
with the testsuite bits, this is good to go.

jeff