[committed] powerpc: Another umaddditi4 fix [PR108862]

[committed] powerpc: Another umaddditi4 fix [PR108862]

[Jakub Jelinek]

Hi!

The following testcase is miscompiled on powerpc64le-linux with
-O2 -mcpu=power9.  The problem is that gen_umaddditi4 is called with
the same TImode register for both op0 and op3, and maddlddi4
overwrites the low half of op0 before the low half of op3 is read,
so when they are the same register it reads the result of maddlddi4.

The following patch fixes that by swapping maddlddi4 and
umadddi4_highpart{,_le} during expansion, as the latter writes into
a temporary pseudo and so can't change anything maddlddi4 depends on.

Bootstrapped/regtested on powerpc64-linux (power7, tested -m32/-m64),
powerpc64le-linux (power8 and another on power9 with
--with-cpu-64=power9 --with-tune-64=power9), preapproved by Segher on IRC,
committed to trunk.

2023-02-20  Jakub Jelinek  <jakub2redhat.com>

	PR target/108862
	* config/rs6000/rs6000.md (umaddditi4): Swap gen_maddlddi4 with
	gen_umadddi4_highpart{,_le}.

	* gcc.dg/pr108862.c: New test.
	* gcc.target/powerpc/pr108862.c: New test.

--- gcc/config/rs6000/rs6000.md.jj	2023-02-15 10:51:12.745802021 +0100
+++ gcc/config/rs6000/rs6000.md	2023-02-20 16:01:02.929027764 +0100
@@ -3249,8 +3249,6 @@
   rtx op3_hi = gen_rtx_SUBREG (DImode, operands[3], BYTES_BIG_ENDIAN ? 0 : 8);
   rtx hi_temp = gen_reg_rtx (DImode);
 
-  emit_insn (gen_maddlddi4 (op0_lo, operands[1], operands[2], op3_lo));
-
   if (BYTES_BIG_ENDIAN)
     emit_insn (gen_umadddi4_highpart (hi_temp, operands[1], operands[2],
 				      op3_lo));
@@ -3258,6 +3256,8 @@
     emit_insn (gen_umadddi4_highpart_le (hi_temp, operands[1], operands[2],
 					 op3_lo));
 
+  emit_insn (gen_maddlddi4 (op0_lo, operands[1], operands[2], op3_lo));
+
   emit_insn (gen_adddi3 (op0_hi, hi_temp, op3_hi));
 
   DONE;
--- gcc/testsuite/gcc.dg/pr108862.c.jj	2023-02-20 15:52:20.570619215 +0100
+++ gcc/testsuite/gcc.dg/pr108862.c	2023-02-20 15:51:52.363029125 +0100
@@ -0,0 +1,27 @@
+/* PR target/108862 */
+/* { dg-do run { target int128 } } */
+/* { dg-options "-O2" } */
+
+unsigned long long a[2] = { 0x04a13945d898c296ULL, 0x0000100000000fffULL };
+unsigned long long b[4] = { 0x04a13945d898c296ULL, 0, 0, 0x0000100000000fffULL };
+
+__attribute__((noipa)) unsigned __int128
+foo (int x, unsigned long long *y, unsigned long long *z)
+{
+  unsigned __int128 w = 0;
+  for (int i = 0; i < x; i++)
+    w += (unsigned __int128)*y++ * (unsigned __int128)*z--;
+  return w;
+}
+
+int
+main ()
+{
+  unsigned __int128 x = foo (1, &a[0], &a[1]);
+  unsigned __int128 y = foo (2, &b[0], &b[3]);
+  if ((unsigned long long) (x >> 64) != 0x0000004a13945dd3ULL
+      || (unsigned long long) x != 0x9b1c8443b3909d6aULL
+      || x != y)
+    __builtin_abort ();
+  return 0;
+}
--- gcc/testsuite/gcc.target/powerpc/pr108862.c.jj	2023-02-20 15:52:51.374171586 +0100
+++ gcc/testsuite/gcc.target/powerpc/pr108862.c	2023-02-20 15:53:04.497980869 +0100
@@ -0,0 +1,6 @@
+/* PR target/108862 */
+/* { dg-do run { target int128 } } */
+/* { dg-require-effective-target p9vector_hw } */
+/* { dg-options "-O2 -mdejagnu-cpu=power9" } */
+
+#include "../../gcc.dg/pr108862.c"

	Jakub