I stumbled over this while trying to fix one of the regressions for 4.9.
This peep2 pattern in the x86 backend is obviously broken as it does not
have a mode on the zero_extend in the resulting insn. As a result,
if/when this peep2 matches we get an unrecognized insn.
(define_peephole2
[(set (match_operand:DI 0 "register_operand")
(zero_extend:DI
(mult:SI (match_operand:SI 1 "register_operand")
(match_operand:SI 2 "const_int_operand"))))]
"TARGET_64BIT
&& exact_log2 (INTVAL (operands[2])) >= 0
&& REGNO (operands[0]) == REGNO (operands[1])
&& peep2_regno_dead_p (0, FLAGS_REG)"
[(parallel [(set (match_dup 0)
(zero_extend (ashift:SI (match_dup 1) (match_dup 2))))
(clobber (reg:CC FLAGS_REG))])]
"operands[2] = GEN_INT (exact_log2 (INTVAL (operands[2])));")
It's obvious the zero_extend should have been DImode. I searched
i386.md for other occurrences but didn't find any (remaining modeless
extensions were on the peep2 and define_split input side patterns).
Bootstrapped and regression tested on x86_64-unknown-linux-gnu.
Installed as obvious. No testcase as triggering requires local hackery.
* i386.md (simple LEA peephole2): Add missing mode to zero_extend
for zero-extended MULT simple LEA pattern.
diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index 6ac2802..ab5b33f 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -17464,7 +17464,7 @@
&& REGNO (operands[0]) == REGNO (operands[1])
&& peep2_regno_dead_p (0, FLAGS_REG)"
[(parallel [(set (match_dup 0)
- (zero_extend (ashift:SI (match_dup 1) (match_dup 2))))
+ (zero_extend:DI (ashift:SI (match_dup 1) (match_dup 2))))
(clobber (reg:CC FLAGS_REG))])]
"operands[2] = GEN_INT (exact_log2 (INTVAL (operands[2])));")
