The RISC-V port showed a handful of regressions after integrating the
final patch from HJ for volatile accesses.
The core issue is the RISC-V port has two patterns which are basically
the same RTL form with the exception of a clobber of a scratch operand.
Those patterns do differ materially in their condition in that one is
more strict than the other. The pattern with the stricter condition
also happens to be the one without the clobber. So it's clearly
stricter across both of those axis.
The stricter pattern naturally generates more efficient (loopless) code
for the relevant atomic operations. We have a handful of tests in the
RISC-V port which verify that we use the right sequences.
With HJ's patch the insn gets re-matched during combine which adds the
clobber and ultimately matches the more general pattern (which is
currently first in the MD file). So we end up with the less efficient
sequence and the relevant tests naturally fail.
This patch just puts the two patterns in the right order with the
stricter pattern coming first. I also walked through the rest of the
sync.md patterns and none obviously had the same problem.
This has been bootstrapped and regression tested on risc-v with both the
Pioneer and BPI F3 systems.
I'll let pre-commit CI chew on it overnight before the planned commit
tomorrow.
jeff
gcc/
* config/riscv/sync.md (atomic compare and set): Reorder patterns
so the stricter pattern comes first.
diff --git a/gcc/config/riscv/sync.md b/gcc/config/riscv/sync.md
index 01eab1a1ef36..bc8a562b1cda 100644
--- a/gcc/config/riscv/sync.md
+++ b/gcc/config/riscv/sync.md
@@ -528,33 +528,6 @@ (define_insn "subword_atomic_exchange_strong"
; Atomic CAS ops
-(define_insn "zalrsc_atomic_cas_value_strong<mode>"
- [(set (match_operand:GPR 0 "register_operand" "=&r")
- (match_operand:GPR 1 "memory_operand" "+A"))
- (set (match_dup 1)
- (unspec_volatile:GPR [(match_operand:GPR 2 "reg_or_0_operand" "rJ")
- (match_operand:GPR 3 "reg_or_0_operand" "rJ")
- (match_operand:SI 4 "const_int_operand") ;; mod_s
- (match_operand:SI 5 "const_int_operand")] ;; mod_f
- UNSPEC_COMPARE_AND_SWAP))
- (clobber (match_scratch:GPR 6 "=&r"))]
- "TARGET_ZALRSC"
- {
- enum memmodel model_success = (enum memmodel) INTVAL (operands[4]);
- enum memmodel model_failure = (enum memmodel) INTVAL (operands[5]);
- /* Find the union of the two memory models so we can satisfy both success
- and failure memory models. */
- operands[5] = GEN_INT (riscv_union_memmodels (model_success,
model_failure));
- return "1:\;"
- "lr.<amo>%I5\t%0,%1\;"
- "bne\t%0,%z2,1f\;"
- "sc.<amo>%J5\t%6,%z3,%1\;"
- "bnez\t%6,1b\;"
- "1:";
- }
- [(set_attr "type" "multi")
- (set (attr "length") (const_int 16))])
-
;; Implement compare_exchange with a conservative leading fence when
;; model_failure is seq_cst.
;; This allows us to be compatible with the ISA manual Table A.6 and Table A.7
@@ -588,6 +561,33 @@ (define_insn "zacas_atomic_cas_value_strong<mode>"
(symbol_ref "(is_mm_seq_cst (memmodel_from_int (INTVAL (operands[5])))
? 8
: 4)"))])
+(define_insn "zalrsc_atomic_cas_value_strong<mode>"
+ [(set (match_operand:GPR 0 "register_operand" "=&r")
+ (match_operand:GPR 1 "memory_operand" "+A"))
+ (set (match_dup 1)
+ (unspec_volatile:GPR [(match_operand:GPR 2 "reg_or_0_operand" "rJ")
+ (match_operand:GPR 3 "reg_or_0_operand" "rJ")
+ (match_operand:SI 4 "const_int_operand") ;; mod_s
+ (match_operand:SI 5 "const_int_operand")] ;; mod_f
+ UNSPEC_COMPARE_AND_SWAP))
+ (clobber (match_scratch:GPR 6 "=&r"))]
+ "TARGET_ZALRSC"
+ {
+ enum memmodel model_success = (enum memmodel) INTVAL (operands[4]);
+ enum memmodel model_failure = (enum memmodel) INTVAL (operands[5]);
+ /* Find the union of the two memory models so we can satisfy both success
+ and failure memory models. */
+ operands[5] = GEN_INT (riscv_union_memmodels (model_success,
model_failure));
+ return "1:\;"
+ "lr.<amo>%I5\t%0,%1\;"
+ "bne\t%0,%z2,1f\;"
+ "sc.<amo>%J5\t%6,%z3,%1\;"
+ "bnez\t%6,1b\;"
+ "1:";
+ }
+ [(set_attr "type" "multi")
+ (set (attr "length") (const_int 16))])
+
(define_expand "atomic_compare_and_swap<mode>"
[(match_operand:SI 0 "register_operand" "") ;; bool output
(match_operand:GPR 1 "register_operand" "") ;; val output
