let PrintMethod = "printInterpSlot";
}
-def isSI : Predicate<"Subtarget.device()"
- "->getGeneration() == AMDGPUDeviceInfo::HD7XXX">;
+def isSI : Predicate<"Subtarget.getGeneration() "
+ "== AMDGPUSubtarget::SOUTHERN_ISLANDS">;
let Predicates = [isSI] in {
} // End isCompare = 1
+def DS_WRITE_B32 : DS_Store_Helper <0x0000000d, "DS_WRITE_B32", VReg_32>;
+def DS_READ_B32 : DS_Load_Helper <0x00000036, "DS_READ_B32", VReg_32>;
+
//def BUFFER_LOAD_FORMAT_X : MUBUF_ <0x00000000, "BUFFER_LOAD_FORMAT_X", []>;
//def BUFFER_LOAD_FORMAT_XY : MUBUF_ <0x00000001, "BUFFER_LOAD_FORMAT_XY", []>;
//def BUFFER_LOAD_FORMAT_XYZ : MUBUF_ <0x00000002, "BUFFER_LOAD_FORMAT_XYZ", []>;
-def BUFFER_LOAD_FORMAT_XYZW : MUBUF_Load_Helper <0x00000003, "BUFFER_LOAD_FORMAT_XYZW", VReg_128>;
+defm BUFFER_LOAD_FORMAT_XYZW : MUBUF_Load_Helper <0x00000003, "BUFFER_LOAD_FORMAT_XYZW", VReg_128>;
//def BUFFER_STORE_FORMAT_X : MUBUF_ <0x00000004, "BUFFER_STORE_FORMAT_X", []>;
//def BUFFER_STORE_FORMAT_XY : MUBUF_ <0x00000005, "BUFFER_STORE_FORMAT_XY", []>;
//def BUFFER_STORE_FORMAT_XYZ : MUBUF_ <0x00000006, "BUFFER_STORE_FORMAT_XYZ", []>;
//def BUFFER_STORE_FORMAT_XYZW : MUBUF_ <0x00000007, "BUFFER_STORE_FORMAT_XYZW", []>;
-//def BUFFER_LOAD_UBYTE : MUBUF_ <0x00000008, "BUFFER_LOAD_UBYTE", []>;
+defm BUFFER_LOAD_UBYTE : MUBUF_Load_Helper <0x00000008, "BUFFER_LOAD_UBYTE", VReg_32>;
//def BUFFER_LOAD_SBYTE : MUBUF_ <0x00000009, "BUFFER_LOAD_SBYTE", []>;
//def BUFFER_LOAD_USHORT : MUBUF_ <0x0000000a, "BUFFER_LOAD_USHORT", []>;
//def BUFFER_LOAD_SSHORT : MUBUF_ <0x0000000b, "BUFFER_LOAD_SSHORT", []>;
-def BUFFER_LOAD_DWORD : MUBUF_Load_Helper <0x0000000c, "BUFFER_LOAD_DWORD", VReg_32>;
-def BUFFER_LOAD_DWORDX2 : MUBUF_Load_Helper <0x0000000d, "BUFFER_LOAD_DWORDX2", VReg_64>;
-def BUFFER_LOAD_DWORDX4 : MUBUF_Load_Helper <0x0000000e, "BUFFER_LOAD_DWORDX4", VReg_128>;
+defm BUFFER_LOAD_DWORD : MUBUF_Load_Helper <0x0000000c, "BUFFER_LOAD_DWORD", VReg_32>;
+defm BUFFER_LOAD_DWORDX2 : MUBUF_Load_Helper <0x0000000d, "BUFFER_LOAD_DWORDX2", VReg_64>;
+defm BUFFER_LOAD_DWORDX4 : MUBUF_Load_Helper <0x0000000e, "BUFFER_LOAD_DWORDX4", VReg_128>;
//def BUFFER_STORE_BYTE : MUBUF_ <0x00000018, "BUFFER_STORE_BYTE", []>;
//def BUFFER_STORE_SHORT : MUBUF_ <0x0000001a, "BUFFER_STORE_SHORT", []>;
def BUFFER_STORE_DWORDX2 : MUBUF_Store_Helper <
0x0000001d, "BUFFER_STORE_DWORDX2", VReg_64, i64
>;
-//def BUFFER_STORE_DWORDX4 : MUBUF_DWORDX4 <0x0000001e, "BUFFER_STORE_DWORDX4", []>;
+
+def BUFFER_STORE_DWORDX4 : MUBUF_Store_Helper <
+ 0x0000001e, "BUFFER_STORE_DWORDX4", VReg_128, v4i32
+>;
//def BUFFER_ATOMIC_SWAP : MUBUF_ <0x00000030, "BUFFER_ATOMIC_SWAP", []>;
//def BUFFER_ATOMIC_CMPSWAP : MUBUF_ <0x00000031, "BUFFER_ATOMIC_CMPSWAP", []>;
//def BUFFER_ATOMIC_ADD : MUBUF_ <0x00000032, "BUFFER_ATOMIC_ADD", []>;
//def IMAGE_SAMPLE_LZ : MIMG_NoPattern_ <"IMAGE_SAMPLE_LZ", 0x00000027>;
def IMAGE_SAMPLE_C : MIMG_Sampler_Helper <0x00000028, "IMAGE_SAMPLE_C">;
//def IMAGE_SAMPLE_C_CL : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_CL", 0x00000029>;
-//def IMAGE_SAMPLE_C_D : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_D", 0x0000002a>;
+def IMAGE_SAMPLE_C_D : MIMG_Sampler_Helper <0x0000002a, "IMAGE_SAMPLE_C_D">;
//def IMAGE_SAMPLE_C_D_CL : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_D_CL", 0x0000002b>;
def IMAGE_SAMPLE_C_L : MIMG_Sampler_Helper <0x0000002c, "IMAGE_SAMPLE_C_L">;
def IMAGE_SAMPLE_C_B : MIMG_Sampler_Helper <0x0000002d, "IMAGE_SAMPLE_C_B">;
} // End isBranch = 1
} // End isTerminator = 1
-//def S_BARRIER : SOPP_ <0x0000000a, "S_BARRIER", []>;
let hasSideEffects = 1 in {
+def S_BARRIER : SOPP <0x0000000a, (ins), "S_BARRIER",
+ [(int_AMDGPU_barrier_local)]
+> {
+ let SIMM16 = 0;
+ let isBarrier = 1;
+ let hasCtrlDep = 1;
+ let mayLoad = 1;
+ let mayStore = 1;
+}
+
def S_WAITCNT : SOPP <0x0000000c, (ins i32imm:$simm16), "S_WAITCNT $simm16",
[]
>;
defm V_MADMK_F32 : VOP2_32 <0x00000020, "V_MADMK_F32", []>;
defm V_MADAK_F32 : VOP2_32 <0x00000021, "V_MADAK_F32", []>;
//defm V_BCNT_U32_B32 : VOP2_32 <0x00000022, "V_BCNT_U32_B32", []>;
-//defm V_MBCNT_LO_U32_B32 : VOP2_32 <0x00000023, "V_MBCNT_LO_U32_B32", []>;
-//defm V_MBCNT_HI_U32_B32 : VOP2_32 <0x00000024, "V_MBCNT_HI_U32_B32", []>;
+defm V_MBCNT_LO_U32_B32 : VOP2_32 <0x00000023, "V_MBCNT_LO_U32_B32", []>;
+defm V_MBCNT_HI_U32_B32 : VOP2_32 <0x00000024, "V_MBCNT_HI_U32_B32", []>;
let isCommutable = 1, Defs = [VCC] in { // Carry-out goes to VCC
defm V_ADD_I32 : VOP2b_32 <0x00000025, "V_ADD_I32",
def V_FMA_F64 : VOP3_64 <0x0000014c, "V_FMA_F64", []>;
//def V_LERP_U8 : VOP3_U8 <0x0000014d, "V_LERP_U8", []>;
def V_ALIGNBIT_B32 : VOP3_32 <0x0000014e, "V_ALIGNBIT_B32", []>;
+def : ROTRPattern <V_ALIGNBIT_B32>;
+
def V_ALIGNBYTE_B32 : VOP3_32 <0x0000014f, "V_ALIGNBYTE_B32", []>;
def V_MULLIT_F32 : VOP3_32 <0x00000150, "V_MULLIT_F32", []>;
////def V_MIN3_F32 : VOP3_MIN3 <0x00000151, "V_MIN3_F32", []>;
} // Uses = [EXEC,VCC,M0], Defs = [EXEC,VCC,M0]
+// This psuedo instruction takes a pointer as input and outputs a resource
+// constant that can be used with the ADDR64 MUBUF instructions.
+
+let usesCustomInserter = 1 in {
+
+def SI_ADDR64_RSRC : InstSI <
+ (outs SReg_128:$srsrc),
+ (ins SReg_64:$ptr),
+ "", []
+>;
+
+} // end usesCustomInserter
+
} // end IsCodeGenOnly, isPseudo
def : Pat<
/* int_SI_vs_load_input */
def : Pat<
- (int_SI_vs_load_input v16i8:$tlst, IMM12bit:$attr_offset,
- i32:$buf_idx_vgpr),
- (BUFFER_LOAD_FORMAT_XYZW imm:$attr_offset, 0, 1, 0, 0, 0,
- $buf_idx_vgpr, $tlst, 0, 0, 0)
+ (int_SI_vs_load_input v16i8:$tlst, IMM12bit:$attr_offset, i32:$buf_idx_vgpr),
+ (BUFFER_LOAD_FORMAT_XYZW_IDXEN $tlst, $buf_idx_vgpr, imm:$attr_offset)
>;
/* int_SI_export */
def : SampleArrayPattern <int_SI_sampleb, IMAGE_SAMPLE_B, addr_type>;
def : SampleShadowPattern <int_SI_sampleb, IMAGE_SAMPLE_C_B, addr_type>;
def : SampleShadowArrayPattern <int_SI_sampleb, IMAGE_SAMPLE_C_B, addr_type>;
+
+ def : SamplePattern <int_SI_sampled, IMAGE_SAMPLE_D, addr_type>;
+ def : SampleArrayPattern <int_SI_sampled, IMAGE_SAMPLE_D, addr_type>;
+ def : SampleShadowPattern <int_SI_sampled, IMAGE_SAMPLE_C_D, addr_type>;
+ def : SampleShadowArrayPattern <int_SI_sampled, IMAGE_SAMPLE_C_D, addr_type>;
}
defm : SamplePatterns<v2i32>;
// 3. Offset in an 32Bit VGPR
def : Pat <
(int_SI_load_const v16i8:$sbase, i32:$voff),
- (BUFFER_LOAD_DWORD 0, 1, 0, 0, 0, 0, $voff, $sbase, 0, 0, 0)
+ (BUFFER_LOAD_DWORD_OFFEN $sbase, $voff)
>;
// The multiplication scales from [0,1] to the unsigned integer range
(V_RCP_IFLAG_F32_e32 (V_CVT_F32_U32_e32 $src0))))
>;
+def : Pat <
+ (int_SI_tid),
+ (V_MBCNT_HI_U32_B32_e32 0xffffffff,
+ (V_MBCNT_LO_U32_B32_e64 0xffffffff, 0, 0, 0, 0, 0))
+>;
+
/********** ================== **********/
/********** VOP3 Patterns **********/
/********** ================== **********/
(V_MAD_F32 $src0, $src1, $src2)
>;
+/********** ======================= **********/
+/********** Load/Store Patterns **********/
+/********** ======================= **********/
+
+def : Pat <
+ (local_load i64:$src0),
+ (i32 (DS_READ_B32 0, (EXTRACT_SUBREG $src0, sub0),
+ (EXTRACT_SUBREG $src0, sub0), (EXTRACT_SUBREG $src0, sub0), 0, 0))
+>;
+
+def : Pat <
+ (local_store i32:$src1, i64:$src0),
+ (DS_WRITE_B32 0, (EXTRACT_SUBREG $src0, sub0), $src1, $src1, 0, 0)
+>;
+
/********** ================== **********/
/********** SMRD Patterns **********/
/********** ================== **********/
defm : SMRD_Pattern <S_LOAD_DWORD_IMM, S_LOAD_DWORD_SGPR, f32>;
defm : SMRD_Pattern <S_LOAD_DWORD_IMM, S_LOAD_DWORD_SGPR, i32>;
+defm : SMRD_Pattern <S_LOAD_DWORDX2_IMM, S_LOAD_DWORDX2_SGPR, i64>;
defm : SMRD_Pattern <S_LOAD_DWORDX4_IMM, S_LOAD_DWORDX4_SGPR, v16i8>;
defm : SMRD_Pattern <S_LOAD_DWORDX8_IMM, S_LOAD_DWORDX8_SGPR, v32i8>;
+//===----------------------------------------------------------------------===//
+// MUBUF Patterns
+//===----------------------------------------------------------------------===//
+
+multiclass MUBUFLoad_Pattern <MUBUF Instr_ADDR64, ValueType vt,
+ PatFrag global_ld, PatFrag constant_ld> {
+ def : Pat <
+ (vt (global_ld (add i64:$ptr, (i64 IMM12bit:$offset)))),
+ (Instr_ADDR64 (SI_ADDR64_RSRC (i64 0)), $ptr, (as_i16imm $offset))
+ >;
+
+ def : Pat <
+ (vt (global_ld i64:$ptr)),
+ (Instr_ADDR64 (SI_ADDR64_RSRC (i64 0)), $ptr, 0)
+ >;
+
+ def : Pat <
+ (vt (global_ld (add i64:$ptr, i64:$offset))),
+ (Instr_ADDR64 (SI_ADDR64_RSRC $ptr), $offset, 0)
+ >;
+
+ def : Pat <
+ (vt (constant_ld (add i64:$ptr, i64:$offset))),
+ (Instr_ADDR64 (SI_ADDR64_RSRC $ptr), $offset, 0)
+ >;
+}
+
+defm : MUBUFLoad_Pattern <BUFFER_LOAD_DWORD_ADDR64, i32,
+ global_load, constant_load>;
+defm : MUBUFLoad_Pattern <BUFFER_LOAD_UBYTE_ADDR64, i32,
+ zextloadi8_global, zextloadi8_constant>;
+defm : MUBUFLoad_Pattern <BUFFER_LOAD_DWORDX2_ADDR64, v2i32,
+ global_load, constant_load>;
+defm : MUBUFLoad_Pattern <BUFFER_LOAD_DWORDX4_ADDR64, v4i32,
+ global_load, constant_load>;
+
+multiclass MUBUFStore_Pattern <MUBUF Instr, ValueType vt> {
+
+ def : Pat <
+ (global_store vt:$value, i64:$ptr),
+ (Instr $value, (SI_ADDR64_RSRC (i64 0)), $ptr, 0)
+ >;
+
+ def : Pat <
+ (global_store vt:$value, (add i64:$ptr, i64:$offset)),
+ (Instr $value, (SI_ADDR64_RSRC $ptr), $offset, 0)
+ >;
+}
+
+defm : MUBUFStore_Pattern <BUFFER_STORE_DWORD, i32>;
+defm : MUBUFStore_Pattern <BUFFER_STORE_DWORDX2, i64>;
+defm : MUBUFStore_Pattern <BUFFER_STORE_DWORDX2, v2i32>;
+defm : MUBUFStore_Pattern <BUFFER_STORE_DWORDX4, v4i32>;
+
/********** ====================== **********/
/********** Indirect adressing **********/
/********** ====================== **********/
projects / oota-llvm.git / blobdiff