let PrintMethod = "printInterpSlot";
}
+def SendMsgImm : Operand<i32> {
+ let PrintMethod = "printSendMsg";
+}
+
def isSI : Predicate<"Subtarget.getGeneration() "
">= AMDGPUSubtarget::SOUTHERN_ISLANDS">;
let mayLoad = 1 in {
-defm S_LOAD_DWORD : SMRD_Helper <0x00, "S_LOAD_DWORD", SReg_64, SReg_32>;
+// We are using the SGPR_32 and not the SReg_32 register class for 32-bit
+// SMRD instructions, because the SGPR_32 register class does not include M0
+// and writing to M0 from an SMRD instruction will hang the GPU.
+defm S_LOAD_DWORD : SMRD_Helper <0x00, "S_LOAD_DWORD", SReg_64, SGPR_32>;
defm S_LOAD_DWORDX2 : SMRD_Helper <0x01, "S_LOAD_DWORDX2", SReg_64, SReg_64>;
defm S_LOAD_DWORDX4 : SMRD_Helper <0x02, "S_LOAD_DWORDX4", SReg_64, SReg_128>;
defm S_LOAD_DWORDX8 : SMRD_Helper <0x03, "S_LOAD_DWORDX8", SReg_64, SReg_256>;
defm S_LOAD_DWORDX16 : SMRD_Helper <0x04, "S_LOAD_DWORDX16", SReg_64, SReg_512>;
defm S_BUFFER_LOAD_DWORD : SMRD_Helper <
- 0x08, "S_BUFFER_LOAD_DWORD", SReg_128, SReg_32
+ 0x08, "S_BUFFER_LOAD_DWORD", SReg_128, SGPR_32
>;
defm S_BUFFER_LOAD_DWORDX2 : SMRD_Helper <
def S_WAITCNT : SOPP <0x0000000c, (ins WAIT_FLAG:$simm16), "S_WAITCNT $simm16",
[]
>;
-} // End hasSideEffects
//def S_SETHALT : SOPP_ <0x0000000d, "S_SETHALT", []>;
//def S_SLEEP : SOPP_ <0x0000000e, "S_SLEEP", []>;
//def S_SETPRIO : SOPP_ <0x0000000f, "S_SETPRIO", []>;
-//def S_SENDMSG : SOPP_ <0x00000010, "S_SENDMSG", []>;
+
+let Uses = [EXEC] in {
+ def S_SENDMSG : SOPP <0x00000010, (ins SendMsgImm:$simm16, M0Reg:$m0), "S_SENDMSG $simm16",
+ [(int_SI_sendmsg imm:$simm16, M0Reg:$m0)]
+ > {
+ let DisableEncoding = "$m0";
+ }
+} // End Uses = [EXEC]
+
//def S_SENDMSGHALT : SOPP_ <0x00000011, "S_SENDMSGHALT", []>;
//def S_TRAP : SOPP_ <0x00000012, "S_TRAP", []>;
//def S_ICACHE_INV : SOPP_ <0x00000013, "S_ICACHE_INV", []>;
//def S_INCPERFLEVEL : SOPP_ <0x00000014, "S_INCPERFLEVEL", []>;
//def S_DECPERFLEVEL : SOPP_ <0x00000015, "S_DECPERFLEVEL", []>;
//def S_TTRACEDATA : SOPP_ <0x00000016, "S_TTRACEDATA", []>;
+} // End hasSideEffects
def V_CNDMASK_B32_e32 : VOP2 <0x00000000, (outs VReg_32:$dst),
(ins VSrc_32:$src0, VReg_32:$src1, VCCReg:$vcc),
$src2), sub1)
>;
-defm V_READLANE_B32 : VOP2_32 <0x00000001, "V_READLANE_B32", []>;
-defm V_WRITELANE_B32 : VOP2_32 <0x00000002, "V_WRITELANE_B32", []>;
+def V_READLANE_B32 : VOP2 <
+ 0x00000001,
+ (outs SReg_32:$vdst),
+ (ins VReg_32:$src0, SSrc_32:$vsrc1),
+ "V_READLANE_B32 $vdst, $src0, $vsrc1",
+ []
+>;
+
+def V_WRITELANE_B32 : VOP2 <
+ 0x00000002,
+ (outs VReg_32:$vdst),
+ (ins SReg_32:$src0, SSrc_32:$vsrc1),
+ "V_WRITELANE_B32 $vdst, $src0, $vsrc1",
+ []
+>;
let isCommutable = 1 in {
defm V_ADD_F32 : VOP2_32 <0x00000003, "V_ADD_F32",
def SI_KILL : InstSI <
(outs),
(ins VReg_32:$src),
- "SI_KIL $src",
+ "SI_KILL $src",
[(int_AMDGPU_kill f32:$src)]
>;
let Uses = [EXEC], Defs = [EXEC,VCC,M0] in {
-//defm SI_ : RegisterLoadStore <VReg_32, FRAMEri64, ADDRIndirect>;
+//defm SI_ : RegisterLoadStore <VReg_32, FRAMEri, ADDRIndirect>;
let UseNamedOperandTable = 1 in {
def SI_RegisterLoad : AMDGPUShaderInst <
(outs VReg_32:$dst, SReg_64:$temp),
- (ins FRAMEri64:$addr, i32imm:$chan),
+ (ins FRAMEri32:$addr, i32imm:$chan),
"", []
> {
let isRegisterLoad = 1;
class SIRegStore<dag outs> : AMDGPUShaderInst <
outs,
- (ins VReg_32:$val, FRAMEri64:$addr, i32imm:$chan),
+ (ins VReg_32:$val, FRAMEri32:$addr, i32imm:$chan),
"", []
> {
let isRegisterStore = 1;
/* int_SI_vs_load_input */
def : Pat<
(SIload_input i128:$tlst, IMM12bit:$attr_offset, i32:$buf_idx_vgpr),
- (BUFFER_LOAD_FORMAT_XYZW_IDXEN $tlst, $buf_idx_vgpr, imm:$attr_offset)
+ (BUFFER_LOAD_FORMAT_XYZW_IDXEN $tlst, $buf_idx_vgpr, imm:$attr_offset, 0, 0, 0, 0)
>;
/* int_SI_export */
def : BitConvert <v8i32, v32i8, SReg_256>;
def : BitConvert <v32i8, v8i32, SReg_256>;
def : BitConvert <v8i32, v32i8, VReg_256>;
+def : BitConvert <v8i32, v8f32, VReg_256>;
+def : BitConvert <v8f32, v8i32, VReg_256>;
def : BitConvert <v32i8, v8i32, VReg_256>;
+def : BitConvert <v16i32, v16f32, VReg_512>;
+def : BitConvert <v16f32, v16i32, VReg_512>;
+
/********** =================== **********/
/********** Src & Dst modifiers **********/
/********** =================== **********/
// 3. Offset in an 32Bit VGPR
def : Pat <
(SIload_constant i128:$sbase, i32:$voff),
- (BUFFER_LOAD_DWORD_OFFEN $sbase, $voff)
+ (BUFFER_LOAD_DWORD_OFFEN $sbase, $voff, 0, 0, 0, 0)
>;
// The multiplication scales from [0,1] to the unsigned integer range
defm : MUBUFStore_Pattern <BUFFER_STORE_DWORDX2, v2i32, global_store>;
defm : MUBUFStore_Pattern <BUFFER_STORE_DWORDX4, v4i32, global_store>;
+// BUFFER_LOAD_DWORD*, addr64=0
+multiclass MUBUF_Load_Dword <ValueType vt, MUBUF offset, MUBUF offen, MUBUF idxen,
+ MUBUF bothen> {
+
+ def : Pat <
+ (vt (int_SI_buffer_load_dword i128:$rsrc, i32:$vaddr, i32:$soffset,
+ imm:$offset, 0, 0, imm:$glc, imm:$slc,
+ imm:$tfe)),
+ (offset $rsrc, $vaddr, (as_i16imm $offset), $soffset, (as_i1imm $glc),
+ (as_i1imm $slc), (as_i1imm $tfe))
+ >;
+
+ def : Pat <
+ (vt (int_SI_buffer_load_dword i128:$rsrc, i32:$vaddr, i32:$soffset,
+ imm, 1, 0, imm:$glc, imm:$slc,
+ imm:$tfe)),
+ (offen $rsrc, $vaddr, $soffset, (as_i1imm $glc), (as_i1imm $slc),
+ (as_i1imm $tfe))
+ >;
+
+ def : Pat <
+ (vt (int_SI_buffer_load_dword i128:$rsrc, i32:$vaddr, i32:$soffset,
+ imm:$offset, 0, 1, imm:$glc, imm:$slc,
+ imm:$tfe)),
+ (idxen $rsrc, $vaddr, (as_i16imm $offset), $soffset, (as_i1imm $glc),
+ (as_i1imm $slc), (as_i1imm $tfe))
+ >;
+
+ def : Pat <
+ (vt (int_SI_buffer_load_dword i128:$rsrc, v2i32:$vaddr, i32:$soffset,
+ imm, 1, 1, imm:$glc, imm:$slc,
+ imm:$tfe)),
+ (bothen $rsrc, $vaddr, $soffset, (as_i1imm $glc), (as_i1imm $slc),
+ (as_i1imm $tfe))
+ >;
+}
+
+defm : MUBUF_Load_Dword <i32, BUFFER_LOAD_DWORD_OFFSET, BUFFER_LOAD_DWORD_OFFEN,
+ BUFFER_LOAD_DWORD_IDXEN, BUFFER_LOAD_DWORD_BOTHEN>;
+defm : MUBUF_Load_Dword <v2i32, BUFFER_LOAD_DWORDX2_OFFSET, BUFFER_LOAD_DWORDX2_OFFEN,
+ BUFFER_LOAD_DWORDX2_IDXEN, BUFFER_LOAD_DWORDX2_BOTHEN>;
+defm : MUBUF_Load_Dword <v4i32, BUFFER_LOAD_DWORDX4_OFFSET, BUFFER_LOAD_DWORDX4_OFFEN,
+ BUFFER_LOAD_DWORDX4_IDXEN, BUFFER_LOAD_DWORDX4_BOTHEN>;
+
//===----------------------------------------------------------------------===//
// MTBUF Patterns
//===----------------------------------------------------------------------===//
/********** Indirect adressing **********/
/********** ====================== **********/
-multiclass SI_INDIRECT_Pattern <ValueType vt, SI_INDIRECT_DST IndDst> {
+multiclass SI_INDIRECT_Pattern <ValueType vt, ValueType eltvt, SI_INDIRECT_DST IndDst> {
// 1. Extract with offset
def : Pat<
// 3. Insert with offset
def : Pat<
- (vector_insert vt:$vec, f32:$val, (add i32:$idx, imm:$off)),
+ (vector_insert vt:$vec, eltvt:$val, (add i32:$idx, imm:$off)),
(IndDst (IMPLICIT_DEF), $vec, $idx, imm:$off, $val)
>;
// 4. Insert without offset
def : Pat<
- (vector_insert vt:$vec, f32:$val, i32:$idx),
+ (vector_insert vt:$vec, eltvt:$val, i32:$idx),
(IndDst (IMPLICIT_DEF), $vec, $idx, 0, $val)
>;
}
-defm : SI_INDIRECT_Pattern <v2f32, SI_INDIRECT_DST_V2>;
-defm : SI_INDIRECT_Pattern <v4f32, SI_INDIRECT_DST_V4>;
-defm : SI_INDIRECT_Pattern <v8f32, SI_INDIRECT_DST_V8>;
-defm : SI_INDIRECT_Pattern <v16f32, SI_INDIRECT_DST_V16>;
+defm : SI_INDIRECT_Pattern <v2f32, f32, SI_INDIRECT_DST_V2>;
+defm : SI_INDIRECT_Pattern <v4f32, f32, SI_INDIRECT_DST_V4>;
+defm : SI_INDIRECT_Pattern <v8f32, f32, SI_INDIRECT_DST_V8>;
+defm : SI_INDIRECT_Pattern <v16f32, f32, SI_INDIRECT_DST_V16>;
+
+defm : SI_INDIRECT_Pattern <v2i32, i32, SI_INDIRECT_DST_V2>;
+defm : SI_INDIRECT_Pattern <v4i32, i32, SI_INDIRECT_DST_V4>;
+defm : SI_INDIRECT_Pattern <v8i32, i32, SI_INDIRECT_DST_V8>;
+defm : SI_INDIRECT_Pattern <v16i32, i32, SI_INDIRECT_DST_V16>;
/********** =============== **********/
/********** Conditions **********/
(EXTRACT_SUBREG $a, sub0)
>;
+def : Pat <
+ (i1 (trunc i32:$a)),
+ (V_CMP_EQ_I32_e64 (V_AND_B32_e32 (i32 1), $a), 1)
+>;
+
// V_ADD_I32_e32/S_ADD_I32 produces carry in VCC/SCC. For the vector
// case, the sgpr-copies pass will fix this to use the vector version.
def : Pat <
projects / oota-llvm.git / blobdiff