1 //===-- CaymanInstructions.td - CM Instruction defs -------*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // TableGen definitions for instructions which are available only on Cayman
13 //===----------------------------------------------------------------------===//
15 def isCayman : Predicate<"Subtarget->hasCaymanISA()">;
17 //===----------------------------------------------------------------------===//
18 // Cayman Instructions
19 //===----------------------------------------------------------------------===//
21 let Predicates = [isCayman] in {
23 def MULADD_INT24_cm : R600_3OP <0x08, "MULADD_INT24",
24 [(set i32:$dst, (AMDGPUmad_i24 i32:$src0, i32:$src1, i32:$src2))], VecALU
26 def MUL_INT24_cm : R600_2OP <0x5B, "MUL_INT24",
27 [(set i32:$dst, (AMDGPUmul_i24 i32:$src0, i32:$src1))], VecALU
30 def : IMad24Pat<MULADD_INT24_cm>;
34 def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
36 def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
37 def MULHI_INT_cm : MULHI_INT_Common<0x90>;
38 def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
39 def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
40 def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
41 def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
42 def LOG_IEEE_cm : LOG_IEEE_Common<0x83>;
43 def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
44 def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
45 def SIN_cm : SIN_Common<0x8D>;
46 def COS_cm : COS_Common<0x8E>;
49 def : RsqPat<RECIPSQRT_IEEE_cm, f32>;
51 def : POW_Common <LOG_IEEE_cm, EXP_IEEE_cm, MUL>;
53 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
55 // RECIP_UINT emulation for Cayman
56 // The multiplication scales from [0,1] to the unsigned integer range
58 (AMDGPUurecip i32:$src0),
59 (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)),
60 (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
63 def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> {
70 def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
72 class RAT_STORE_DWORD <RegisterClass rc, ValueType vt, bits<4> mask> :
73 CF_MEM_RAT_CACHELESS <0x14, 0, mask,
74 (ins rc:$rw_gpr, R600_TReg32_X:$index_gpr),
75 "STORE_DWORD $rw_gpr, $index_gpr",
76 [(global_store vt:$rw_gpr, i32:$index_gpr)]> {
77 let eop = 0; // This bit is not used on Cayman.
80 def RAT_STORE_DWORD32 : RAT_STORE_DWORD <R600_TReg32_X, i32, 0x1>;
81 def RAT_STORE_DWORD64 : RAT_STORE_DWORD <R600_Reg64, v2i32, 0x3>;
82 def RAT_STORE_DWORD128 : RAT_STORE_DWORD <R600_Reg128, v4i32, 0xf>;
84 def RAT_STORE_TYPED_cm: CF_MEM_RAT_STORE_TYPED<0> {
85 let eop = 0; // This bit is not used on Cayman.
88 class VTX_READ_cm <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
89 : VTX_WORD0_cm, VTX_READ<name, buffer_id, outs, pattern> {
94 let FETCH_WHOLE_QUAD = 0;
95 let BUFFER_ID = buffer_id;
97 // XXX: We can infer this field based on the SRC_GPR. This would allow us
98 // to store vertex addresses in any channel, not just X.
101 let STRUCTURED_READ = 0;
103 let COALESCED_READ = 0;
105 let Inst{31-0} = Word0;
108 class VTX_READ_8_cm <bits<8> buffer_id, list<dag> pattern>
109 : VTX_READ_cm <"VTX_READ_8 $dst_gpr, $src_gpr", buffer_id,
110 (outs R600_TReg32_X:$dst_gpr), pattern> {
113 let DST_SEL_Y = 7; // Masked
114 let DST_SEL_Z = 7; // Masked
115 let DST_SEL_W = 7; // Masked
116 let DATA_FORMAT = 1; // FMT_8
119 class VTX_READ_16_cm <bits<8> buffer_id, list<dag> pattern>
120 : VTX_READ_cm <"VTX_READ_16 $dst_gpr, $src_gpr", buffer_id,
121 (outs R600_TReg32_X:$dst_gpr), pattern> {
123 let DST_SEL_Y = 7; // Masked
124 let DST_SEL_Z = 7; // Masked
125 let DST_SEL_W = 7; // Masked
126 let DATA_FORMAT = 5; // FMT_16
130 class VTX_READ_32_cm <bits<8> buffer_id, list<dag> pattern>
131 : VTX_READ_cm <"VTX_READ_32 $dst_gpr, $src_gpr", buffer_id,
132 (outs R600_TReg32_X:$dst_gpr), pattern> {
135 let DST_SEL_Y = 7; // Masked
136 let DST_SEL_Z = 7; // Masked
137 let DST_SEL_W = 7; // Masked
138 let DATA_FORMAT = 0xD; // COLOR_32
140 // This is not really necessary, but there were some GPU hangs that appeared
141 // to be caused by ALU instructions in the next instruction group that wrote
142 // to the $src_gpr registers of the VTX_READ.
144 // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
145 // %T2_X<def> = MOV %ZERO
146 //Adding this constraint prevents this from happening.
147 let Constraints = "$src_gpr.ptr = $dst_gpr";
150 class VTX_READ_64_cm <bits<8> buffer_id, list<dag> pattern>
151 : VTX_READ_cm <"VTX_READ_64 $dst_gpr, $src_gpr", buffer_id,
152 (outs R600_Reg64:$dst_gpr), pattern> {
158 let DATA_FORMAT = 0x1D; // COLOR_32_32
161 class VTX_READ_128_cm <bits<8> buffer_id, list<dag> pattern>
162 : VTX_READ_cm <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", buffer_id,
163 (outs R600_Reg128:$dst_gpr), pattern> {
169 let DATA_FORMAT = 0x22; // COLOR_32_32_32_32
171 // XXX: Need to force VTX_READ_128 instructions to write to the same register
172 // that holds its buffer address to avoid potential hangs. We can't use
173 // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst
174 // registers are different sizes.
177 //===----------------------------------------------------------------------===//
178 // VTX Read from parameter memory space
179 //===----------------------------------------------------------------------===//
180 def VTX_READ_PARAM_8_cm : VTX_READ_8_cm <0,
181 [(set i32:$dst_gpr, (load_param_exti8 ADDRVTX_READ:$src_gpr))]
184 def VTX_READ_PARAM_16_cm : VTX_READ_16_cm <0,
185 [(set i32:$dst_gpr, (load_param_exti16 ADDRVTX_READ:$src_gpr))]
188 def VTX_READ_PARAM_32_cm : VTX_READ_32_cm <0,
189 [(set i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
192 def VTX_READ_PARAM_64_cm : VTX_READ_64_cm <0,
193 [(set v2i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
196 def VTX_READ_PARAM_128_cm : VTX_READ_128_cm <0,
197 [(set v4i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))]
200 //===----------------------------------------------------------------------===//
201 // VTX Read from global memory space
202 //===----------------------------------------------------------------------===//
205 def VTX_READ_ID1_8_cm : VTX_READ_8_cm <1,
206 [(set i32:$dst_gpr, (vtx_id1_az_extloadi8 ADDRVTX_READ:$src_gpr))]
210 def VTX_READ_ID1_16_cm : VTX_READ_16_cm <1,
211 [(set i32:$dst_gpr, (vtx_id1_az_extloadi16 ADDRVTX_READ:$src_gpr))]
215 def VTX_READ_ID1_32_cm : VTX_READ_32_cm <1,
216 [(set i32:$dst_gpr, (vtx_id1_load ADDRVTX_READ:$src_gpr))]
220 def VTX_READ_ID1_64_cm : VTX_READ_64_cm <1,
221 [(set v2i32:$dst_gpr, (vtx_id1_load ADDRVTX_READ:$src_gpr))]
225 def VTX_READ_ID1_128_cm : VTX_READ_128_cm <1,
226 [(set v4i32:$dst_gpr, (vtx_id1_load ADDRVTX_READ:$src_gpr))]
230 def VTX_READ_ID2_8_cm : VTX_READ_8_cm <2,
231 [(set i32:$dst_gpr, (vtx_id2_az_extloadi8 ADDRVTX_READ:$src_gpr))]
235 def VTX_READ_ID2_16_cm : VTX_READ_16_cm <2,
236 [(set i32:$dst_gpr, (vtx_id2_az_extloadi16 ADDRVTX_READ:$src_gpr))]
240 def VTX_READ_ID2_32_cm : VTX_READ_32_cm <2,
241 [(set i32:$dst_gpr, (vtx_id2_load ADDRVTX_READ:$src_gpr))]
245 def VTX_READ_ID2_64_cm : VTX_READ_64_cm <2,
246 [(set v2i32:$dst_gpr, (vtx_id2_load ADDRVTX_READ:$src_gpr))]
250 def VTX_READ_ID2_128_cm : VTX_READ_128_cm <2,
251 [(set v4i32:$dst_gpr, (vtx_id2_load ADDRVTX_READ:$src_gpr))]