[PR123121, LRA]: Fix wrong rematerialization of insns with several outputs

LRA in the test case, rematerialize insn with div/mod where div result
is not used.  Still div result requires ax which is used by different
pseudos at point of rematerialization and this clobbers the pseudo
value.  The patch solves the problem by constraining to single set
insns as we always rematerialize only one pseudo value.  Also there is
no sense to rematerialize div/mod as usually their latency is more
than load value from CPU cache.  The patch explicitly excludes such
insns from rematerialization.

gcc/ChangeLog:

	PR rtl-optimization/123121
	* lra-remat.cc (bad_for_rematerialization_p): Consider div/mod ops.
	(operand_to_remat): Exclude rematerialization of insns with
	multiple sets.

gcc/testsuite/ChangeLog:

	PR rtl-optimization/123121
	* gcc.target/i386/pr123121.c: New.

gcc/lra-remat.cc

@@ -81,8 +81,10 @@ typedef const struct cand *const_cand_t;
 
 /* Insn candidates for rematerialization.  The candidate insn should
    have the following properies:
-   o no any memory (as access to memory is non-profitable)
+   o no any memory (as access to memory is non-profitable) or
+     div/mod operations (as they are usually more expensive than loads)
    o no INOUT regs (it means no non-paradoxical subreg of output reg)
+   o no multiple output pseudos
    o one output spilled pseudo (or reload pseudo of a spilled pseudo)
    o all other pseudos are with assigned hard regs.  */
 struct cand
@@ -249,12 +251,12 @@ finish_cand_table (void)
 
 
 
-/* Return true if X contains memory or some UNSPEC.  We cannot just
-   check insn operands as memory or unspec might be not an operand
-   itself but contain an operand.  Insn with memory access is not
-   profitable for rematerialization.  Rematerialization of UNSPEC
-   might result in wrong code generation as the UNPEC effect is
-   unknown (e.g. generating a label).  */
+/* Return true if X contains memory, some UNSPEC, or expensive operations.  We
+   cannot just check insn operands as memory or unspec might be not an operand
+   itself but contain an operand.  Insns with memory access or expensive ones
+   are not profitable for rematerialization.  Rematerialization of UNSPEC might
+   result in wrong code generation as the UNPEC effect is unknown
+   (e.g. generating a label).  */
 static bool
 bad_for_rematerialization_p (rtx x)
 {
@@ -262,7 +264,11 @@ bad_for_rematerialization_p (rtx x)
   const char *fmt;
   enum rtx_code code;
 
-  if (MEM_P (x) || GET_CODE (x) == UNSPEC || GET_CODE (x) == UNSPEC_VOLATILE)
+  if (MEM_P (x) || GET_CODE (x) == UNSPEC || GET_CODE (x) == UNSPEC_VOLATILE
+      /* Usually the following operations are expensive and does not worth to
+	 rematerialize: */
+      || GET_CODE(x) == DIV || GET_CODE(x) == UDIV
+      || GET_CODE(x) == MOD || GET_CODE(x) == UMOD)
     return true;
   code = GET_CODE (x);
   fmt = GET_RTX_FORMAT (code);
@@ -308,8 +314,7 @@ operand_to_remat (rtx_insn *insn)
 	 cannot know sp offset at a rematerialization place.  */
       if (reg->regno == STACK_POINTER_REGNUM && frame_pointer_needed)
 	return -1;
-      else if (reg->type == OP_OUT && ! reg->subreg_p
-	       && find_regno_note (insn, REG_UNUSED, reg->regno) == NULL)
+      else if (reg->type == OP_OUT)
 	{
 	  /* We permits only one spilled reg.  */
 	  if (found_reg != NULL)

gcc/testsuite/gcc.target/i386/pr123121.c

@@ -0,0 +1,57 @@
+/* { dg-do run } */
+/* { dg-require-effective-target lp64 } */
+/* { dg-options "-O3 -fno-ipa-cp -fschedule-insns" } */
+
+int a[256], d, e, f, g, h, j, v[] = {0, -6, 0};;
+unsigned c;
+unsigned k(int l, unsigned m) { return (l & 6777215) ^ a[(l ^ m) & 55]; }
+int n(int l, int m[]) {
+  int i = 0;
+  for (; i < l; ++i) {
+    e = c >> 8 ^ a[m[i] & 255];
+    d = e;
+    d = k(d, m[i] >> 8 & 255);
+    d = k(d, m[i] >> 6);
+    f = (d & 6777215) ^ a[d & 5];
+    c = f;
+  }
+  return f;
+}
+int o() {
+  n(8, (int[]){g});
+  return n(6, (int[]){h});
+}
+int p(int l, int m[], int q, int r) {
+  int s = r = j + 1;
+  s = -(4 % q - 1);
+  if (r)
+    goto t;
+ ad:
+  l = 6 - l;
+  if (l)
+    goto ae;
+ t:
+  m[0] = s;
+  j = r;
+  goto af;
+ ae:
+  if (o())
+    goto ad;
+ af:
+  r = -s - m[1] - 8;
+  s = 1 % m[0] - s;
+  m[1] = 1 / r;
+  int a[] = {l, m[0], m[1], m[2], q, r, r, s};
+  return n(8, a);
+}
+int main() {
+  for (int i = 0; i < 256; i++) {
+    unsigned b = i;
+    if (i & 1)
+      b = b >> 1 ^ 3988292384;
+    a[i] = b;
+  }
+  if (p(1, v, 5, 0) / 100000)
+    p(1, 0, 5, 0);
+  return 0;
+}