def DS_MIN_F32 : DS_1A1D_NORET <0x12, "DS_MIN_F32", VReg_32>;
def DS_MAX_F32 : DS_1A1D_NORET <0x13, "DS_MAX_F32", VReg_32>;
-def DS_ADD_RTN_U32 : DS_1A1D_RET <0x20, "DS_ADD_RTN_U32", VReg_32>;
-def DS_SUB_RTN_U32 : DS_1A1D_RET <0x21, "DS_SUB_RTN_U32", VReg_32>;
-def DS_RSUB_RTN_U32 : DS_1A1D_RET <0x22, "DS_RSUB_RTN_U32", VReg_32>;
-def DS_INC_RTN_U32 : DS_1A1D_RET <0x23, "DS_INC_RTN_U32", VReg_32>;
-def DS_DEC_RTN_U32 : DS_1A1D_RET <0x24, "DS_DEC_RTN_U32", VReg_32>;
-def DS_MIN_RTN_I32 : DS_1A1D_RET <0x25, "DS_MIN_RTN_I32", VReg_32>;
-def DS_MAX_RTN_I32 : DS_1A1D_RET <0x26, "DS_MAX_RTN_I32", VReg_32>;
-def DS_MIN_RTN_U32 : DS_1A1D_RET <0x27, "DS_MIN_RTN_U32", VReg_32>;
-def DS_MAX_RTN_U32 : DS_1A1D_RET <0x28, "DS_MAX_RTN_U32", VReg_32>;
-def DS_AND_RTN_B32 : DS_1A1D_RET <0x29, "DS_AND_RTN_B32", VReg_32>;
-def DS_OR_RTN_B32 : DS_1A1D_RET <0x2a, "DS_OR_RTN_B32", VReg_32>;
-def DS_XOR_RTN_B32 : DS_1A1D_RET <0x2b, "DS_XOR_RTN_B32", VReg_32>;
-def DS_MSKOR_RTN_B32 : DS_1A1D_RET <0x2c, "DS_MSKOR_RTN_B32", VReg_32>;
+def DS_ADD_RTN_U32 : DS_1A1D_RET <0x20, "DS_ADD_RTN_U32", VReg_32, "DS_ADD_U32">;
+def DS_SUB_RTN_U32 : DS_1A1D_RET <0x21, "DS_SUB_RTN_U32", VReg_32, "DS_SUB_U32">;
+def DS_RSUB_RTN_U32 : DS_1A1D_RET <0x22, "DS_RSUB_RTN_U32", VReg_32, "DS_RSUB_U32">;
+def DS_INC_RTN_U32 : DS_1A1D_RET <0x23, "DS_INC_RTN_U32", VReg_32, "DS_INC_U32">;
+def DS_DEC_RTN_U32 : DS_1A1D_RET <0x24, "DS_DEC_RTN_U32", VReg_32, "DS_DEC_U32">;
+def DS_MIN_RTN_I32 : DS_1A1D_RET <0x25, "DS_MIN_RTN_I32", VReg_32, "DS_MIN_I32">;
+def DS_MAX_RTN_I32 : DS_1A1D_RET <0x26, "DS_MAX_RTN_I32", VReg_32, "DS_MAX_I32">;
+def DS_MIN_RTN_U32 : DS_1A1D_RET <0x27, "DS_MIN_RTN_U32", VReg_32, "DS_MIN_U32">;
+def DS_MAX_RTN_U32 : DS_1A1D_RET <0x28, "DS_MAX_RTN_U32", VReg_32, "DS_MAX_U32">;
+def DS_AND_RTN_B32 : DS_1A1D_RET <0x29, "DS_AND_RTN_B32", VReg_32, "DS_AND_B32">;
+def DS_OR_RTN_B32 : DS_1A1D_RET <0x2a, "DS_OR_RTN_B32", VReg_32, "DS_OR_B32">;
+def DS_XOR_RTN_B32 : DS_1A1D_RET <0x2b, "DS_XOR_RTN_B32", VReg_32, "DS_XOR_B32">;
+def DS_MSKOR_RTN_B32 : DS_1A1D_RET <0x2c, "DS_MSKOR_RTN_B32", VReg_32, "DS_MSKOR_B32">;
def DS_WRXCHG_RTN_B32 : DS_1A1D_RET <0x2d, "DS_WRXCHG_RTN_B32", VReg_32>;
-//def DS_WRXCHG2_RTN_B32 : DS_2A0D_RET <0x2e, "DS_WRXCHG2_RTN_B32", VReg_32>;
-//def DS_WRXCHG2ST64_RTN_B32 : DS_2A0D_RET <0x2f, "DS_WRXCHG2_RTN_B32", VReg_32>;
-def DS_CMPST_RTN_B32 : DS_1A2D_RET <0x30, "DS_CMPST_RTN_B32", VReg_32>;
-def DS_CMPST_RTN_F32 : DS_1A2D_RET <0x31, "DS_CMPST_RTN_F32", VReg_32>;
-def DS_MIN_RTN_F32 : DS_1A1D_RET <0x32, "DS_MIN_RTN_F32", VReg_32>;
-def DS_MAX_RTN_F32 : DS_1A1D_RET <0x33, "DS_MAX_RTN_F32", VReg_32>;
+//def DS_WRXCHG2_RTN_B32 : DS_2A0D_RET <0x2e, "DS_WRXCHG2_RTN_B32", VReg_32, "DS_WRXCHG2_B32">;
+//def DS_WRXCHG2ST64_RTN_B32 : DS_2A0D_RET <0x2f, "DS_WRXCHG2_RTN_B32", VReg_32, "DS_WRXCHG2ST64_B32">;
+def DS_CMPST_RTN_B32 : DS_1A2D_RET <0x30, "DS_CMPST_RTN_B32", VReg_32, "DS_CMPST_B32">;
+def DS_CMPST_RTN_F32 : DS_1A2D_RET <0x31, "DS_CMPST_RTN_F32", VReg_32, "DS_CMPST_F32">;
+def DS_MIN_RTN_F32 : DS_1A1D_RET <0x32, "DS_MIN_RTN_F32", VReg_32, "DS_MIN_F32">;
+def DS_MAX_RTN_F32 : DS_1A1D_RET <0x33, "DS_MAX_RTN_F32", VReg_32, "DS_MAX_F32">;
let SubtargetPredicate = isCI in {
-def DS_WRAP_RTN_F32 : DS_1A1D_RET <0x34, "DS_WRAP_RTN_F32", VReg_32>;
+def DS_WRAP_RTN_F32 : DS_1A1D_RET <0x34, "DS_WRAP_RTN_F32", VReg_32, "DS_WRAP_F32">;
} // End isCI
-def DS_ADD_U64 : DS_1A1D_NORET <0x40, "DS_ADD_U64", VReg_32>;
-def DS_SUB_U64 : DS_1A1D_NORET <0x41, "DS_SUB_U64", VReg_32>;
-def DS_RSUB_U64 : DS_1A1D_NORET <0x42, "DS_RSUB_U64", VReg_32>;
-def DS_INC_U64 : DS_1A1D_NORET <0x43, "DS_INC_U64", VReg_32>;
-def DS_DEC_U64 : DS_1A1D_NORET <0x44, "DS_DEC_U64", VReg_32>;
+def DS_ADD_U64 : DS_1A1D_NORET <0x40, "DS_ADD_U64", VReg_64>;
+def DS_SUB_U64 : DS_1A1D_NORET <0x41, "DS_SUB_U64", VReg_64>;
+def DS_RSUB_U64 : DS_1A1D_NORET <0x42, "DS_RSUB_U64", VReg_64>;
+def DS_INC_U64 : DS_1A1D_NORET <0x43, "DS_INC_U64", VReg_64>;
+def DS_DEC_U64 : DS_1A1D_NORET <0x44, "DS_DEC_U64", VReg_64>;
def DS_MIN_I64 : DS_1A1D_NORET <0x45, "DS_MIN_I64", VReg_64>;
def DS_MAX_I64 : DS_1A1D_NORET <0x46, "DS_MAX_I64", VReg_64>;
def DS_MIN_U64 : DS_1A1D_NORET <0x47, "DS_MIN_U64", VReg_64>;
def DS_MIN_F64 : DS_1A1D_NORET <0x52, "DS_MIN_F64", VReg_64>;
def DS_MAX_F64 : DS_1A1D_NORET <0x53, "DS_MAX_F64", VReg_64>;
-def DS_ADD_RTN_U64 : DS_1A1D_RET <0x60, "DS_ADD_RTN_U64", VReg_64>;
-def DS_SUB_RTN_U64 : DS_1A1D_RET <0x61, "DS_SUB_RTN_U64", VReg_64>;
-def DS_RSUB_RTN_U64 : DS_1A1D_RET <0x62, "DS_RSUB_RTN_U64", VReg_64>;
-def DS_INC_RTN_U64 : DS_1A1D_RET <0x63, "DS_INC_RTN_U64", VReg_64>;
-def DS_DEC_RTN_U64 : DS_1A1D_RET <0x64, "DS_DEC_RTN_U64", VReg_64>;
-def DS_MIN_RTN_I64 : DS_1A1D_RET <0x65, "DS_MIN_RTN_I64", VReg_64>;
-def DS_MAX_RTN_I64 : DS_1A1D_RET <0x66, "DS_MAX_RTN_I64", VReg_64>;
-def DS_MIN_RTN_U64 : DS_1A1D_RET <0x67, "DS_MIN_RTN_U64", VReg_64>;
-def DS_MAX_RTN_U64 : DS_1A1D_RET <0x68, "DS_MAX_RTN_U64", VReg_64>;
-def DS_AND_RTN_B64 : DS_1A1D_RET <0x69, "DS_AND_RTN_B64", VReg_64>;
-def DS_OR_RTN_B64 : DS_1A1D_RET <0x6a, "DS_OR_RTN_B64", VReg_64>;
-def DS_XOR_RTN_B64 : DS_1A1D_RET <0x6b, "DS_XOR_RTN_B64", VReg_64>;
-def DS_MSKOR_RTN_B64 : DS_1A1D_RET <0x6c, "DS_MSKOR_RTN_B64", VReg_64>;
-def DS_WRXCHG_RTN_B64 : DS_1A1D_RET <0x6d, "DS_WRXCHG_RTN_B64", VReg_64>;
-//def DS_WRXCHG2_RTN_B64 : DS_2A0D_RET <0x6e, "DS_WRXCHG2_RTN_B64", VReg_64>;
-//def DS_WRXCHG2ST64_RTN_B64 : DS_2A0D_RET <0x6f, "DS_WRXCHG2_RTN_B64", VReg_64>;
-def DS_CMPST_RTN_B64 : DS_1A2D_RET <0x70, "DS_CMPST_RTN_B64", VReg_64>;
-def DS_CMPST_RTN_F64 : DS_1A2D_RET <0x71, "DS_CMPST_RTN_F64", VReg_64>;
-def DS_MIN_RTN_F64 : DS_1A1D_RET <0x72, "DS_MIN_F64", VReg_64>;
-def DS_MAX_RTN_F64 : DS_1A1D_RET <0x73, "DS_MAX_F64", VReg_64>;
+def DS_ADD_RTN_U64 : DS_1A1D_RET <0x60, "DS_ADD_RTN_U64", VReg_64, "DS_ADD_U64">;
+def DS_SUB_RTN_U64 : DS_1A1D_RET <0x61, "DS_SUB_RTN_U64", VReg_64, "DS_SUB_U64">;
+def DS_RSUB_RTN_U64 : DS_1A1D_RET <0x62, "DS_RSUB_RTN_U64", VReg_64, "DS_RSUB_U64">;
+def DS_INC_RTN_U64 : DS_1A1D_RET <0x63, "DS_INC_RTN_U64", VReg_64, "DS_INC_U64">;
+def DS_DEC_RTN_U64 : DS_1A1D_RET <0x64, "DS_DEC_RTN_U64", VReg_64, "DS_DEC_U64">;
+def DS_MIN_RTN_I64 : DS_1A1D_RET <0x65, "DS_MIN_RTN_I64", VReg_64, "DS_MIN_I64">;
+def DS_MAX_RTN_I64 : DS_1A1D_RET <0x66, "DS_MAX_RTN_I64", VReg_64, "DS_MAX_I64">;
+def DS_MIN_RTN_U64 : DS_1A1D_RET <0x67, "DS_MIN_RTN_U64", VReg_64, "DS_MIN_U64">;
+def DS_MAX_RTN_U64 : DS_1A1D_RET <0x68, "DS_MAX_RTN_U64", VReg_64, "DS_MAX_U64">;
+def DS_AND_RTN_B64 : DS_1A1D_RET <0x69, "DS_AND_RTN_B64", VReg_64, "DS_AND_B64">;
+def DS_OR_RTN_B64 : DS_1A1D_RET <0x6a, "DS_OR_RTN_B64", VReg_64, "DS_OR_B64">;
+def DS_XOR_RTN_B64 : DS_1A1D_RET <0x6b, "DS_XOR_RTN_B64", VReg_64, "DS_XOR_B64">;
+def DS_MSKOR_RTN_B64 : DS_1A1D_RET <0x6c, "DS_MSKOR_RTN_B64", VReg_64, "DS_MSKOR_B64">;
+def DS_WRXCHG_RTN_B64 : DS_1A1D_RET <0x6d, "DS_WRXCHG_RTN_B64", VReg_64, "DS_WRXCHG_B64">;
+//def DS_WRXCHG2_RTN_B64 : DS_2A0D_RET <0x6e, "DS_WRXCHG2_RTN_B64", VReg_64, "DS_WRXCHG2_B64">;
+//def DS_WRXCHG2ST64_RTN_B64 : DS_2A0D_RET <0x6f, "DS_WRXCHG2_RTN_B64", VReg_64, "DS_WRXCHG2ST64_B64">;
+def DS_CMPST_RTN_B64 : DS_1A2D_RET <0x70, "DS_CMPST_RTN_B64", VReg_64, "DS_CMPST_B64">;
+def DS_CMPST_RTN_F64 : DS_1A2D_RET <0x71, "DS_CMPST_RTN_F64", VReg_64, "DS_CMPST_F64">;
+def DS_MIN_RTN_F64 : DS_1A1D_RET <0x72, "DS_MIN_F64", VReg_64, "DS_MIN_F64">;
+def DS_MAX_RTN_F64 : DS_1A1D_RET <0x73, "DS_MAX_F64", VReg_64, "DS_MAX_F64">;
//let SubtargetPredicate = isCI in {
// DS_CONDXCHG32_RTN_B64
(EXTRACT_SUBREG $value, sub1), $offset0, $offset1)
>;
-multiclass DSAtomicRetPat<DS inst, ValueType vt, PatFrag frag> {
- def : Pat <
- (frag (add i32:$ptr, (i32 IMM16bit:$offset)), vt:$value),
- (inst (i1 0), $ptr, $value, (as_i16imm $offset))
- >;
-
- def : Pat <
- (frag i32:$ptr, vt:$val),
- (inst 0, $ptr, $val, 0)
- >;
-}
+class DSAtomicRetPat<DS inst, ValueType vt, PatFrag frag> : Pat <
+ (frag (DS1Addr1Offset i32:$ptr, i32:$offset), vt:$value),
+ (inst (i1 0), $ptr, $value, (as_i16imm $offset))
+>;
// Special case of DSAtomicRetPat for add / sub 1 -> inc / dec
//
// We also load this -1 with s_mov_b32 / s_mov_b64 even though this
// needs to be a VGPR. The SGPR copy pass will fix this, and it's
// easier since there is no v_mov_b64.
-multiclass DSAtomicIncRetPat<DS inst, ValueType vt,
- Instruction LoadImm, PatFrag frag> {
- def : Pat <
- (frag (add i32:$ptr, (i32 IMM16bit:$offset)), (vt 1)),
- (inst (i1 0), $ptr, (LoadImm (vt -1)), (as_i16imm $offset))
- >;
-
- def : Pat <
- (frag i32:$ptr, (vt 1)),
- (inst 0, $ptr, (LoadImm (vt -1)), 0)
- >;
-}
+class DSAtomicIncRetPat<DS inst, ValueType vt,
+ Instruction LoadImm, PatFrag frag> : Pat <
+ (frag (DS1Addr1Offset i32:$ptr, i32:$offset), (vt 1)),
+ (inst (i1 0), $ptr, (LoadImm (vt -1)), (as_i16imm $offset))
+>;
-multiclass DSAtomicCmpXChg <DS inst, ValueType vt, PatFrag frag> {
- def : Pat <
- (frag (add i32:$ptr, (i32 IMM16bit:$offset)), vt:$cmp, vt:$swap),
- (inst (i1 0), $ptr, $cmp, $swap, (as_i16imm $offset))
- >;
- def : Pat <
- (frag i32:$ptr, vt:$cmp, vt:$swap),
- (inst 0, $ptr, $cmp, $swap, 0)
- >;
-}
+class DSAtomicCmpXChg <DS inst, ValueType vt, PatFrag frag> : Pat <
+ (frag (DS1Addr1Offset i32:$ptr, i32:$offset), vt:$cmp, vt:$swap),
+ (inst (i1 0), $ptr, $cmp, $swap, (as_i16imm $offset))
+>;
// 32-bit atomics.
-defm : DSAtomicIncRetPat<DS_INC_RTN_U32, i32,
- S_MOV_B32, atomic_load_add_local>;
-defm : DSAtomicIncRetPat<DS_DEC_RTN_U32, i32,
- S_MOV_B32, atomic_load_sub_local>;
-
-defm : DSAtomicRetPat<DS_WRXCHG_RTN_B32, i32, atomic_swap_local>;
-defm : DSAtomicRetPat<DS_ADD_RTN_U32, i32, atomic_load_add_local>;
-defm : DSAtomicRetPat<DS_SUB_RTN_U32, i32, atomic_load_sub_local>;
-defm : DSAtomicRetPat<DS_AND_RTN_B32, i32, atomic_load_and_local>;
-defm : DSAtomicRetPat<DS_OR_RTN_B32, i32, atomic_load_or_local>;
-defm : DSAtomicRetPat<DS_XOR_RTN_B32, i32, atomic_load_xor_local>;
-defm : DSAtomicRetPat<DS_MIN_RTN_I32, i32, atomic_load_min_local>;
-defm : DSAtomicRetPat<DS_MAX_RTN_I32, i32, atomic_load_max_local>;
-defm : DSAtomicRetPat<DS_MIN_RTN_U32, i32, atomic_load_umin_local>;
-defm : DSAtomicRetPat<DS_MAX_RTN_U32, i32, atomic_load_umax_local>;
-
-defm : DSAtomicCmpXChg<DS_CMPST_RTN_B32, i32, atomic_cmp_swap_32_local>;
+def : DSAtomicIncRetPat<DS_INC_RTN_U32, i32,
+ S_MOV_B32, atomic_load_add_local>;
+def : DSAtomicIncRetPat<DS_DEC_RTN_U32, i32,
+ S_MOV_B32, atomic_load_sub_local>;
+
+def : DSAtomicRetPat<DS_WRXCHG_RTN_B32, i32, atomic_swap_local>;
+def : DSAtomicRetPat<DS_ADD_RTN_U32, i32, atomic_load_add_local>;
+def : DSAtomicRetPat<DS_SUB_RTN_U32, i32, atomic_load_sub_local>;
+def : DSAtomicRetPat<DS_AND_RTN_B32, i32, atomic_load_and_local>;
+def : DSAtomicRetPat<DS_OR_RTN_B32, i32, atomic_load_or_local>;
+def : DSAtomicRetPat<DS_XOR_RTN_B32, i32, atomic_load_xor_local>;
+def : DSAtomicRetPat<DS_MIN_RTN_I32, i32, atomic_load_min_local>;
+def : DSAtomicRetPat<DS_MAX_RTN_I32, i32, atomic_load_max_local>;
+def : DSAtomicRetPat<DS_MIN_RTN_U32, i32, atomic_load_umin_local>;
+def : DSAtomicRetPat<DS_MAX_RTN_U32, i32, atomic_load_umax_local>;
+
+def : DSAtomicCmpXChg<DS_CMPST_RTN_B32, i32, atomic_cmp_swap_32_local>;
// 64-bit atomics.
-defm : DSAtomicIncRetPat<DS_INC_RTN_U64, i64,
- S_MOV_B64, atomic_load_add_local>;
-defm : DSAtomicIncRetPat<DS_DEC_RTN_U64, i64,
- S_MOV_B64, atomic_load_sub_local>;
-
-defm : DSAtomicRetPat<DS_WRXCHG_RTN_B64, i64, atomic_swap_local>;
-defm : DSAtomicRetPat<DS_ADD_RTN_U64, i64, atomic_load_add_local>;
-defm : DSAtomicRetPat<DS_SUB_RTN_U64, i64, atomic_load_sub_local>;
-defm : DSAtomicRetPat<DS_AND_RTN_B64, i64, atomic_load_and_local>;
-defm : DSAtomicRetPat<DS_OR_RTN_B64, i64, atomic_load_or_local>;
-defm : DSAtomicRetPat<DS_XOR_RTN_B64, i64, atomic_load_xor_local>;
-defm : DSAtomicRetPat<DS_MIN_RTN_I64, i64, atomic_load_min_local>;
-defm : DSAtomicRetPat<DS_MAX_RTN_I64, i64, atomic_load_max_local>;
-defm : DSAtomicRetPat<DS_MIN_RTN_U64, i64, atomic_load_umin_local>;
-defm : DSAtomicRetPat<DS_MAX_RTN_U64, i64, atomic_load_umax_local>;
-
-defm : DSAtomicCmpXChg<DS_CMPST_RTN_B64, i64, atomic_cmp_swap_64_local>;
+def : DSAtomicIncRetPat<DS_INC_RTN_U64, i64,
+ S_MOV_B64, atomic_load_add_local>;
+def : DSAtomicIncRetPat<DS_DEC_RTN_U64, i64,
+ S_MOV_B64, atomic_load_sub_local>;
+
+def : DSAtomicRetPat<DS_WRXCHG_RTN_B64, i64, atomic_swap_local>;
+def : DSAtomicRetPat<DS_ADD_RTN_U64, i64, atomic_load_add_local>;
+def : DSAtomicRetPat<DS_SUB_RTN_U64, i64, atomic_load_sub_local>;
+def : DSAtomicRetPat<DS_AND_RTN_B64, i64, atomic_load_and_local>;
+def : DSAtomicRetPat<DS_OR_RTN_B64, i64, atomic_load_or_local>;
+def : DSAtomicRetPat<DS_XOR_RTN_B64, i64, atomic_load_xor_local>;
+def : DSAtomicRetPat<DS_MIN_RTN_I64, i64, atomic_load_min_local>;
+def : DSAtomicRetPat<DS_MAX_RTN_I64, i64, atomic_load_max_local>;
+def : DSAtomicRetPat<DS_MIN_RTN_U64, i64, atomic_load_umin_local>;
+def : DSAtomicRetPat<DS_MAX_RTN_U64, i64, atomic_load_umax_local>;
+
+def : DSAtomicCmpXChg<DS_CMPST_RTN_B64, i64, atomic_cmp_swap_64_local>;
//===----------------------------------------------------------------------===//