From: Matthew Malcomson <mmalcom...@nvidia.com>
Do demo implementation in AArch64 since that's the backend I'm most
familiar with.
Nothing much else to say -- nice to see that the demo implementation
seems to work as expected (being used for fetch_add, add_fetch and
sub_fetch even though it's only defined for fetch_sub).
Demo implementation ensures that I can run some execution tests with the
alternate codepath.
Demo is added behind a flag in order to be able to run the testsuite
with different variants (with the flag and without).
Ensuring that the functionality worked for both the fallback and when
this optab was implemented (also check with the two different fallbacks
of either using libatomic or inlining a CAS loop).
In order to run with this and the fallback implementation I use the
following flag in RUNTESTFLAGS:
--target_board='unix {unix/-mtesting-fp-atomics}'
Signed-off-by: Matthew Malcomson <mmalcom...@nvidia.com>
---
gcc/config/aarch64/aarch64.h | 2 ++
gcc/config/aarch64/aarch64.opt | 5 +++++
gcc/config/aarch64/atomics.md | 15 +++++++++++++++
3 files changed, 22 insertions(+)
diff --git a/gcc/config/aarch64/aarch64.h b/gcc/config/aarch64/aarch64.h
index f07b2c49f0d..16e30972aa9 100644
--- a/gcc/config/aarch64/aarch64.h
+++ b/gcc/config/aarch64/aarch64.h
@@ -121,6 +121,8 @@
of LSE instructions. */
#define TARGET_OUTLINE_ATOMICS (aarch64_flag_outline_atomics)
+#define TARGET_TESTING_FP_ATOMICS (aarch64_flag_testing_fp_atomics)
+
/* Align definitions of arrays, unions and structures so that
initializations and copies can be made more efficient. This is not
ABI-changing, so it only affects places where we can see the
diff --git a/gcc/config/aarch64/aarch64.opt b/gcc/config/aarch64/aarch64.opt
index c2c9965b062..d02d6f11839 100644
--- a/gcc/config/aarch64/aarch64.opt
+++ b/gcc/config/aarch64/aarch64.opt
@@ -338,6 +338,11 @@ moutline-atomics
Target Var(aarch64_flag_outline_atomics) Init(2) Save
Generate local calls to out-of-line atomic operations.
+mtesting-fp-atomics
+Target Var(aarch64_flag_testing_fp_atomics) Init(0) Save
+Use the demonstration implementation of atomic_fetch_sub_<mode> for floating
+point modes.
+
-param=aarch64-vect-compare-costs=
Target Joined UInteger Var(aarch64_vect_compare_costs) Init(1) IntegerRange(0,
1) Param
When vectorizing, consider using multiple different approaches and use
diff --git a/gcc/config/aarch64/atomics.md b/gcc/config/aarch64/atomics.md
index 32a0a723732..ee8fbcd6c58 100644
--- a/gcc/config/aarch64/atomics.md
+++ b/gcc/config/aarch64/atomics.md
@@ -368,6 +368,21 @@
;; However we also implement the acquire memory barrier with DMB LD,
;; and so the ST<OP> is not blocked by the barrier.
+(define_insn "atomic_fetch_sub<mode>"
+ [(set (match_operand:GPF 0 "register_operand" "=&w")
+ (match_operand:GPF 1 "aarch64_sync_memory_operand" "+Q"))
+ (set (match_dup 1)
+ (unspec_volatile:GPF
+ [(minus:GPF (match_dup 1)
+ (match_operand:GPF 2 "register_operand" "w"))
+ (match_operand:SI 3 "const_int_operand")]
+ UNSPECV_ATOMIC_LDOP_PLUS))
+ (clobber (match_scratch:GPF 4 "=w"))]
+ "TARGET_TESTING_FP_ATOMICS"
+ "// Here's your sandwich.\;ldr %<s>0, %1\;fsub %<s>4, %<s>0, %<s>2\;str
%<s>4, %1\;// END"
+)
+
+
(define_insn "aarch64_atomic_<atomic_ldoptab><mode>_lse"
[(set (match_operand:ALLI 0 "aarch64_sync_memory_operand" "+Q")
(unspec_volatile:ALLI
--
2.43.0
