[(set i64:$dst, (sra i64:$src0, i32:$src1))]
>;
-} // End AddedComplexity = 1
def S_BFM_B32 : SOP2_32 <0x00000024, "S_BFM_B32", []>;
def S_BFM_B64 : SOP2_64 <0x00000025, "S_BFM_B64", []>;
-def S_MUL_I32 : SOP2_32 <0x00000026, "S_MUL_I32", []>;
+def S_MUL_I32 : SOP2_32 <0x00000026, "S_MUL_I32",
+ [(set i32:$dst, (mul i32:$src0, i32:$src1))]
+>;
+
+} // End AddedComplexity = 1
+
def S_BFE_U32 : SOP2_32 <0x00000027, "S_BFE_U32", []>;
def S_BFE_I32 : SOP2_32 <0x00000028, "S_BFE_I32", []>;
def S_BFE_U64 : SOP2_64 <0x00000029, "S_BFE_U64", []>;
// SOP2 Patterns
//===----------------------------------------------------------------------===//
-// V_ADD_I32_e32/S_ADD_I32 produces carry in VCC/SCC. For the vector
+// V_ADD_I32_e32/S_ADD_U32 produces carry in VCC/SCC. For the vector
// case, the sgpr-copies pass will fix this to use the vector version.
def : Pat <
(i32 (addc i32:$src0, i32:$src1)),
- (S_ADD_I32 $src0, $src1)
+ (S_ADD_U32 $src0, $src1)
>;
} // Predicates = [isSI, isCFDepth0]
(EXTRACT_SUBREG i64:$src1, sub1)), sub1)
>;
+def : BinOp64Pat <and, V_AND_B32_e32>;
def : BinOp64Pat <or, V_OR_B32_e32>;
def : BinOp64Pat <xor, V_XOR_B32_e32>;
(V_MUL_HI_I32 $src0, $src1)
>;
+def : Vop3ModPat<V_MAD_F32, VOP_F32_F32_F32_F32, AMDGPUmad>;
+
+
defm : BFIPatterns <V_BFI_B32, S_MOV_B32>;
def : ROTRPattern <V_ALIGNBIT_B32>;
def : DSReadPat <DS_READ_I16, i32, sextloadi16_local>;
def : DSReadPat <DS_READ_U16, i32, az_extloadi16_local>;
def : DSReadPat <DS_READ_B32, i32, local_load>;
-def : DSReadPat <DS_READ_B64, v2i32, local_load>;
+
+let AddedComplexity = 100 in {
+
+def : DSReadPat <DS_READ_B64, v2i32, local_load_aligned8bytes>;
+
+} // End AddedComplexity = 100
+
+def : Pat <
+ (v2i32 (local_load (DS64Bit4ByteAligned i32:$ptr, i8:$offset0,
+ i8:$offset1))),
+ (DS_READ2_B32 (i1 0), $ptr, $offset0, $offset1)
+>;
class DSWritePat <DS inst, ValueType vt, PatFrag frag> : Pat <
(frag vt:$value, (DS1Addr1Offset i32:$ptr, i32:$offset)),
def : DSWritePat <DS_WRITE_B8, i32, truncstorei8_local>;
def : DSWritePat <DS_WRITE_B16, i32, truncstorei16_local>;
def : DSWritePat <DS_WRITE_B32, i32, local_store>;
-def : DSWritePat <DS_WRITE_B64, v2i32, local_store>;
-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 $offset))
- >;
+let AddedComplexity = 100 in {
- def : Pat <
- (frag i32:$ptr, vt:$val),
- (inst 0, $ptr, $val, 0)
- >;
-}
+def : DSWritePat <DS_WRITE_B64, v2i32, local_store_aligned8bytes>;
+} // End AddedComplexity = 100
+
+def : Pat <
+ (local_store v2i32:$value, (DS64Bit4ByteAligned i32:$ptr, i8:$offset0,
+ i8:$offset1)),
+ (DS_WRITE2_B32 (i1 0), $ptr, (EXTRACT_SUBREG $value, sub0),
+ (EXTRACT_SUBREG $value, sub1), $offset0, $offset1)
+>;
+
+class DSAtomicRetPat<DS inst, ValueType vt, PatFrag frag> : Pat <
+ (frag (DS1Addr1Offset i32:$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>;
//===----------------------------------------------------------------------===//
projects / oota-llvm.git / blobdiff