+/* The following multiclass tries to make sure that in code like
+ * x.store (immediate op x.load(acquire), release)
+ * and
+ * x.store (register op x.load(acquire), release)
+ * an operation directly on memory is generated instead of wasting a register.
+ * It is not automatic as atomic_store/load are only lowered to MOV instructions
+ * extremely late to prevent them from being accidentally reordered in the backend
+ * (see below the RELEASE_MOV* / ACQUIRE_MOV* pseudo-instructions)
+ */
+multiclass RELEASE_BINOP_MI<SDNode op> {
+ def NAME#8mi : I<0, Pseudo, (outs), (ins i8mem:$dst, i8imm:$src),
+ "#BINOP "#NAME#"8mi PSEUDO!",
+ [(atomic_store_8 addr:$dst, (op
+ (atomic_load_8 addr:$dst), (i8 imm:$src)))]>;
+ def NAME#8mr : I<0, Pseudo, (outs), (ins i8mem:$dst, GR8:$src),
+ "#BINOP "#NAME#"8mr PSEUDO!",
+ [(atomic_store_8 addr:$dst, (op
+ (atomic_load_8 addr:$dst), GR8:$src))]>;
+ // NAME#16 is not generated as 16-bit arithmetic instructions are considered
+ // costly and avoided as far as possible by this backend anyway
+ def NAME#32mi : I<0, Pseudo, (outs), (ins i32mem:$dst, i32imm:$src),
+ "#BINOP "#NAME#"32mi PSEUDO!",
+ [(atomic_store_32 addr:$dst, (op
+ (atomic_load_32 addr:$dst), (i32 imm:$src)))]>;
+ def NAME#32mr : I<0, Pseudo, (outs), (ins i32mem:$dst, GR32:$src),
+ "#BINOP "#NAME#"32mr PSEUDO!",
+ [(atomic_store_32 addr:$dst, (op
+ (atomic_load_32 addr:$dst), GR32:$src))]>;
+ def NAME#64mi32 : I<0, Pseudo, (outs), (ins i64mem:$dst, i64i32imm:$src),
+ "#BINOP "#NAME#"64mi32 PSEUDO!",
+ [(atomic_store_64 addr:$dst, (op
+ (atomic_load_64 addr:$dst), (i64immSExt32:$src)))]>;
+ def NAME#64mr : I<0, Pseudo, (outs), (ins i64mem:$dst, GR64:$src),
+ "#BINOP "#NAME#"64mr PSEUDO!",
+ [(atomic_store_64 addr:$dst, (op
+ (atomic_load_64 addr:$dst), GR64:$src))]>;
+}
+defm RELEASE_ADD : RELEASE_BINOP_MI<add>;
+defm RELEASE_AND : RELEASE_BINOP_MI<and>;
+defm RELEASE_OR : RELEASE_BINOP_MI<or>;
+defm RELEASE_XOR : RELEASE_BINOP_MI<xor>;
+// Note: we don't deal with sub, because substractions of constants are
+// optimized into additions before this code can run
+
+// Same as above, but for floating-point.
+// FIXME: imm version.
+// FIXME: Version that doesn't clobber $src, using AVX's VADDSS.
+// FIXME: This could also handle SIMD operations with *ps and *pd instructions.
+let usesCustomInserter = 1 in {
+multiclass RELEASE_FP_BINOP_MI<SDNode op> {
+ def NAME#32mr : I<0, Pseudo, (outs), (ins i32mem:$dst, FR32:$src),
+ "#BINOP "#NAME#"32mr PSEUDO!",
+ [(atomic_store_32 addr:$dst,
+ (i32 (bitconvert (op
+ (f32 (bitconvert (i32 (atomic_load_32 addr:$dst)))),
+ FR32:$src))))]>, Requires<[HasSSE1]>;
+ def NAME#64mr : I<0, Pseudo, (outs), (ins i64mem:$dst, FR64:$src),
+ "#BINOP "#NAME#"64mr PSEUDO!",
+ [(atomic_store_64 addr:$dst,
+ (i64 (bitconvert (op
+ (f64 (bitconvert (i64 (atomic_load_64 addr:$dst)))),
+ FR64:$src))))]>, Requires<[HasSSE2]>;
+}
+defm RELEASE_FADD : RELEASE_FP_BINOP_MI<fadd>;
+// FIXME: Add fsub, fmul, fdiv, ...
+}
+
+multiclass RELEASE_UNOP<dag dag8, dag dag16, dag dag32, dag dag64> {
+ def NAME#8m : I<0, Pseudo, (outs), (ins i8mem:$dst),
+ "#UNOP "#NAME#"8m PSEUDO!",
+ [(atomic_store_8 addr:$dst, dag8)]>;
+ def NAME#16m : I<0, Pseudo, (outs), (ins i16mem:$dst),
+ "#UNOP "#NAME#"16m PSEUDO!",
+ [(atomic_store_16 addr:$dst, dag16)]>;
+ def NAME#32m : I<0, Pseudo, (outs), (ins i32mem:$dst),
+ "#UNOP "#NAME#"32m PSEUDO!",
+ [(atomic_store_32 addr:$dst, dag32)]>;
+ def NAME#64m : I<0, Pseudo, (outs), (ins i64mem:$dst),
+ "#UNOP "#NAME#"64m PSEUDO!",
+ [(atomic_store_64 addr:$dst, dag64)]>;
+}
+
+defm RELEASE_INC : RELEASE_UNOP<
+ (add (atomic_load_8 addr:$dst), (i8 1)),
+ (add (atomic_load_16 addr:$dst), (i16 1)),
+ (add (atomic_load_32 addr:$dst), (i32 1)),
+ (add (atomic_load_64 addr:$dst), (i64 1))>, Requires<[NotSlowIncDec]>;
+defm RELEASE_DEC : RELEASE_UNOP<
+ (add (atomic_load_8 addr:$dst), (i8 -1)),
+ (add (atomic_load_16 addr:$dst), (i16 -1)),
+ (add (atomic_load_32 addr:$dst), (i32 -1)),
+ (add (atomic_load_64 addr:$dst), (i64 -1))>, Requires<[NotSlowIncDec]>;
+/*
+TODO: These don't work because the type inference of TableGen fails.
+TODO: find a way to fix it.
+defm RELEASE_NEG : RELEASE_UNOP<
+ (ineg (atomic_load_8 addr:$dst)),
+ (ineg (atomic_load_16 addr:$dst)),
+ (ineg (atomic_load_32 addr:$dst)),
+ (ineg (atomic_load_64 addr:$dst))>;
+defm RELEASE_NOT : RELEASE_UNOP<
+ (not (atomic_load_8 addr:$dst)),
+ (not (atomic_load_16 addr:$dst)),
+ (not (atomic_load_32 addr:$dst)),
+ (not (atomic_load_64 addr:$dst))>;
+*/
+
+def RELEASE_MOV8mi : I<0, Pseudo, (outs), (ins i8mem:$dst, i8imm:$src),
+ "#RELEASE_MOV8mi PSEUDO!",
+ [(atomic_store_8 addr:$dst, (i8 imm:$src))]>;
+def RELEASE_MOV16mi : I<0, Pseudo, (outs), (ins i16mem:$dst, i16imm:$src),
+ "#RELEASE_MOV16mi PSEUDO!",
+ [(atomic_store_16 addr:$dst, (i16 imm:$src))]>;
+def RELEASE_MOV32mi : I<0, Pseudo, (outs), (ins i32mem:$dst, i32imm:$src),
+ "#RELEASE_MOV32mi PSEUDO!",
+ [(atomic_store_32 addr:$dst, (i32 imm:$src))]>;
+def RELEASE_MOV64mi32 : I<0, Pseudo, (outs), (ins i64mem:$dst, i64i32imm:$src),
+ "#RELEASE_MOV64mi32 PSEUDO!",
+ [(atomic_store_64 addr:$dst, i64immSExt32:$src)]>;