powerpc: Another umaddditi4 fix [PR108862]

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.

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.

diff --git a/gcc/config/rs6000/rs6000.md b/gcc/config/rs6000/rs6000.md
index 162340b5581ac73d304fe3aba5d6c34249e525cc..81bffb04ceb0c237feec10176acc972a444c7a38 100644 (file)
--- a/gcc/config/rs6000/rs6000.md
+++ b/gcc/config/rs6000/rs6000.md
@@ -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;

diff --git a/gcc/testsuite/gcc.dg/pr108862.c b/gcc/testsuite/gcc.dg/pr108862.c
new file mode 100644 (file)
index 0000000..065099e
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pr108862.c
@@ -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;
+}

diff --git a/gcc/testsuite/gcc.target/powerpc/pr108862.c b/gcc/testsuite/gcc.target/powerpc/pr108862.c
new file mode 100644 (file)
index 0000000..89bbb08
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/pr108862.c
@@ -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"