Our friends at Altera spotted a bug in the Nios II backend that was
causing different code to be generated randomly for functions involving
taking the address of a symbol. The root of the problem was that in
r255492 I goofed and introduced some uses of INTVAL that were not
guarded by CONST_INT_P so they were randomly grabbing contents of a
SYMBOL_REF rtx instead. AFAICT the generated code was not functionally
incorrect, just different; the bug caused the movsi_internal splitter to
fail to fire when it was supposed to, but I think LRA was subsequently
generating the %hi/%lo pair via a different code path.
The fix is pretty obvious.... add the missing guard. :-P
-Sandra
2018-03-15 Sandra Loosemore <san...@codesourcery.com>
gcc/
* config/nios2/nios2.md (movsi_internal): Fix thinko in
split predicate.
Index: gcc/config/nios2/nios2.md
===================================================================
--- gcc/config/nios2/nios2.md (revision 482298)
+++ gcc/config/nios2/nios2.md (working copy)
@@ -298,9 +298,10 @@
}
"(nios2_large_constant_memory_operand_p (operands[0])
|| nios2_large_constant_memory_operand_p (operands[1])
- || (nios2_large_constant_p (operands[1])
- && !SMALL_INT_UNSIGNED (INTVAL (operands[1]))
- && !UPPER16_INT (INTVAL (operands[1]))))"
+ || (nios2_large_constant_p (operands[1])
+ && !(CONST_INT_P (operands[1])
+ && (SMALL_INT_UNSIGNED (INTVAL (operands[1]))
+ || UPPER16_INT (INTVAL (operands[1]))))))"
[(set (match_dup 0) (match_dup 1))]
{
if (nios2_large_constant_memory_operand_p (operands[0]))
