https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108248
Bug ID: 108248
Summary: Some insns in the risc-v backend do not have mappings
to functional units
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: law at gcc dot gnu.org
CC: rzinsly at ventanamicro dot com
Target Milestone: ---
Target: risc-v
If you review scheduling dumps for risc-v you will see various insns with no
reservations. For example:
;; 11--> b 3: i 58 r184=smax(r154,r183) :nothing
;; 3--> b 3: i 55 r182=r180<<0x3+r160 :nothing
;; 8--> b 0: i 16 r167=r193<<0x3+r193 :nothing
;; 13--> b 0: i 19 r155=r167<<0x3+r188 :nothing
;; 0--> b 0: i 26 r172=r164<<0x3+r164 :nothing
;; 2--> b 0: i 23 r144=r161<<0x3+r159 :nothing
It's worth noting these all seem to be coming from the bitmanip extensions.
However, there well could be others.
It would be advisable to ensure that all insns have appropriate mappings to
functional units.
This patch from Jivan would help, but I think a thorough review to ensure
everything has a mapping would be advisable:
diff --git a/gcc/config/riscv/generic.md b/gcc/config/riscv/generic.md
index 1a209dcb997..296c4e24923 100644
--- a/gcc/config/riscv/generic.md
+++ b/gcc/config/riscv/generic.md
@@ -88,3 +88,8 @@
(and (eq_attr "tune" "generic")
(eq_attr "type" "fsqrt"))
"fdivsqrt*25")
+
+(define_insn_reservation "generic_smax" 1
+ (and (eq_attr "tune" "generic")
+ (eq_attr "type" "bitmanip"))
+ "alu")
