Before the introduction of OP_SETFVAL, floating-point were
created via OP_SETVAL whose CSE is done by comparing the
pointer of the corresponding expression without any
interpretation of this pointer.
As consequence, even if two OP_SETVAL have two identical
expressions (value), in most cases the corresponding pointers
are not identical, completly inhibiting the CSE of OP_SETVALs.
Fix the CSE of floating-point literals by directly using
the value given by the new OP_SETFVAL.
Note: to respect some of the subtilities of floating-point,
the equality comparison of two literals is not done on
the floating-point value itself but bit-by-bit on its
binary representation (as such we can continue to make the
distinction between +0.0 & -0.0, handle NaNs, ...).
Signed-off-by: Luc Van Oostenryck <luc.vanoostenryck@xxxxxxxxx>
---
cse.c | 11 +++++++++++
validation/optim/cse-setfval.c | 12 ++++++++++++
2 files changed, 23 insertions(+)
create mode 100644 validation/optim/cse-setfval.c
diff --git a/cse.c b/cse.c
index f535636b4..cd1e8942c 100644
--- a/cse.c
+++ b/cse.c
@@ -89,6 +89,10 @@ static void clean_up_one_instruction(struct basic_block *bb, struct instruction
hash += hashval(insn->val);
break;
+ case OP_SETFVAL:
+ hash += hashval(insn->fvalue);
+ break;
+
case OP_SYMADDR:
hash += hashval(insn->symbol);
break;
@@ -178,6 +182,7 @@ static int insn_compare(const void *_i1, const void *_i2)
{
const struct instruction *i1 = _i1;
const struct instruction *i2 = _i2;
+ int diff;
if (i1->opcode != i2->opcode)
return i1->opcode < i2->opcode ? -1 : 1;
@@ -240,6 +245,12 @@ static int insn_compare(const void *_i1, const void *_i2)
return i1->val < i2->val ? -1 : 1;
break;
+ case OP_SETFVAL:
+ diff = memcmp(&i1->fvalue, &i2->fvalue, sizeof(i1->fvalue));
+ if (diff)
+ return diff;
+ break;
+
/* Other */
case OP_PHI:
return phi_list_compare(i1->phi_list, i2->phi_list);
diff --git a/validation/optim/cse-setfval.c b/validation/optim/cse-setfval.c
new file mode 100644
index 000000000..59c00a407
--- /dev/null
+++ b/validation/optim/cse-setfval.c
@@ -0,0 +1,12 @@
+int ftest(double a, double b)
+{
+ return a == 0.125 || b == 0.125;
+}
+
+/*
+ * check-name: CSE OP_SETFVAL
+ * check-command: test-linearize -Wno-decl $file
+ *
+ * check-output-ignore
+ * check-output-pattern-1-times: setfval\\.
+ */
--
2.12.0
--
To unsubscribe from this list: send the line "unsubscribe linux-sparse" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
