summaryrefslogtreecommitdiff
path: root/lib
diff options
context:
space:
mode:
Diffstat (limited to 'lib')
-rw-r--r--lib/Target/AMDGPU/AMDILISelDAGToDAG.cpp47
-rw-r--r--lib/Target/AMDGPU/R600Defines.h19
-rw-r--r--lib/Target/AMDGPU/R600ISelLowering.cpp13
-rw-r--r--lib/Target/AMDGPU/R600Instructions.td311
4 files changed, 307 insertions, 83 deletions
diff --git a/lib/Target/AMDGPU/AMDILISelDAGToDAG.cpp b/lib/Target/AMDGPU/AMDILISelDAGToDAG.cpp
index 807113134d2..a9439b88003 100644
--- a/lib/Target/AMDGPU/AMDILISelDAGToDAG.cpp
+++ b/lib/Target/AMDGPU/AMDILISelDAGToDAG.cpp
@@ -70,6 +70,11 @@ private:
bool SelectADDR8BitOffset(SDValue Addr, SDValue& Base, SDValue& Offset);
bool SelectADDRReg(SDValue Addr, SDValue& Base, SDValue& Offset);
bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset);
+ bool SelectADDRFrame(SDValue Addr, SDValue &Base, SDValue &Offset);
+ bool SelectALUDst(SDValue N, SDValue &GPR, SDValue &WriteMask, SDValue &Omod,
+ SDValue &Rel, SDValue &Clamp);
+ bool SelectALUSrcOp2(SDValue N, SDValue &GPR, SDValue &Neg, SDValue &Rel,
+ SDValue &Abs);
// Include the pieces autogenerated from the target description.
#include "AMDGPUGenDAGISel.inc"
@@ -97,6 +102,8 @@ SDValue AMDGPUDAGToDAGISel::getSmallIPtrImm(unsigned int Imm) {
return CurDAG->getTargetConstant(Imm, MVT::i32);
}
+
+
bool AMDGPUDAGToDAGISel::SelectADDRParam(
SDValue Addr, SDValue& R1, SDValue& R2) {
@@ -393,3 +400,43 @@ bool AMDGPUDAGToDAGISel::SelectADDRReg(SDValue Addr, SDValue& Base,
return true;
}
+
+bool AMDGPUDAGToDAGISel::SelectADDRFrame(SDValue Addr, SDValue &Base,
+ SDValue &Offset) {
+ if (Addr.getOpcode() == ISD::FrameIndex) {
+ Base = CurDAG->getRegister(AMDGPU::ZERO, MVT::i32);
+ Offset = CurDAG->getConstant(0, MVT::i32);
+// unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister(RegIndex);
+ return true;
+ }
+
+ if (Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) {
+ if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr.getOperand(0))) {
+ Base = Addr.getOperand(1);
+ } else {
+ Base = Addr;
+ }
+
+ Offset = CurDAG->getConstant(0, MVT::i32);
+ return true;
+ }
+
+ return false;
+}
+
+bool AMDGPUDAGToDAGISel::SelectALUDst(SDValue N, SDValue &GPR,
+ SDValue &WriteMask, SDValue &Omod,
+ SDValue &Rel, SDValue &Clamp)
+{
+ return false;
+}
+
+bool AMDGPUDAGToDAGISel::SelectALUSrcOp2(SDValue N, SDValue &GPR, SDValue &Neg,
+ SDValue &Rel, SDValue &Abs)
+{
+ GPR = N;
+ Neg = CurDAG->getTargetConstant(0, MVT::i32);
+ Rel = CurDAG->getTargetConstant(0, MVT::i32);
+ Abs = CurDAG->getTargetConstant(0, MVT::i32);
+ return true;
+}
diff --git a/lib/Target/AMDGPU/R600Defines.h b/lib/Target/AMDGPU/R600Defines.h
index 8191c6a64a7..962f4888c5f 100644
--- a/lib/Target/AMDGPU/R600Defines.h
+++ b/lib/Target/AMDGPU/R600Defines.h
@@ -33,8 +33,9 @@ namespace R600_InstFlag {
FC = (1 << 3),
TRIG = (1 << 4),
OP3 = (1 << 5),
- VECTOR = (1 << 6)
+ VECTOR = (1 << 6),
//FlagOperand bits 7, 8
+ HAS_NATIVE_OPERANDS = (1 << 9)
};
}
@@ -42,4 +43,20 @@ namespace R600_InstFlag {
#define HW_REG_MASK 0x1ff
#define HW_CHAN_SHIFT 9
+namespace R600Op2OperandIndex {
+ enum ROI {
+ SRC0,
+ SRC0_NEG,
+ SRC0_REL,
+ IMM0,
+ SRC1,
+ SRC1_REL,
+ SRC1_NEG,
+ IMM1,
+ INDEX_MODE,
+ PRED_SEL,
+ LAST
+ };
+}
+
#endif // R600DEFINES_H_
diff --git a/lib/Target/AMDGPU/R600ISelLowering.cpp b/lib/Target/AMDGPU/R600ISelLowering.cpp
index da1de57777a..db208438739 100644
--- a/lib/Target/AMDGPU/R600ISelLowering.cpp
+++ b/lib/Target/AMDGPU/R600ISelLowering.cpp
@@ -112,7 +112,18 @@ MachineBasicBlock * R600TargetLowering::EmitInstrWithCustomInserter(
TII->addFlag(NewMI, 1, MO_FLAG_NEG);
break;
}
-
+#if 0
+ case AMDGPU::FrameLoad_i32:
+ case AMDGPU::FrameStore_i32:
+ {
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::MOV))
+ .addOperand(MI->getOperand(0))
+ .addOperand(MI->getOperand(1))
+ .addImm(0)
+ .addReg(AMDGPU::PRED_SEL_OFF);
+ break;
+ }
+#endif
case AMDGPU::R600_LOAD_CONST:
{
int64_t RegIndex = MI->getOperand(1).getImm();
diff --git a/lib/Target/AMDGPU/R600Instructions.td b/lib/Target/AMDGPU/R600Instructions.td
index 4ce256d1298..eba9e72a6c4 100644
--- a/lib/Target/AMDGPU/R600Instructions.td
+++ b/lib/Target/AMDGPU/R600Instructions.td
@@ -22,6 +22,7 @@ class InstR600 <bits<11> inst, dag outs, dag ins, string asm, list<dag> pattern,
bit Op3 = 0;
bit isVector = 0;
bits<2> FlagOperandIdx = 0;
+ bit HasNativeOperands = 0;
bits<11> op_code = inst;
//let Inst = inst;
@@ -39,6 +40,7 @@ class InstR600 <bits<11> inst, dag outs, dag ins, string asm, list<dag> pattern,
// instruction group
let TSFlags{6} = isVector;
let TSFlags{8-7} = FlagOperandIdx;
+ let TSFlags{9} = HasNativeOperands;
}
class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern> :
@@ -59,21 +61,163 @@ def MEMrr : Operand<iPTR> {
def InstFlag : OperandWithDefaultOps <i32, (ops (i32 0))>;
+def FRAMEri : Operand<iPTR> {
+ let MIOperandInfo = (ops R600_Reg32:$ptr, i32imm:$index);
+}
+
+class ALUDstBase <ValueType vt> : Operand <vt> {
+ let MIOperandInfo = (ops R600_Reg32:$gpr, InstFlag:$write_mask,
+ InstFlag:$omod, InstFlag:$rel, InstFlag:$clamp);
+}
+
+def ALUDst_I32 : ALUDstBase <i32>;
+def ALUDst_F32 : ALUDstBase <f32>;
+
+class ALUSrcOp2Base <ValueType vt, Operand immType> : Operand <vt> {
+ let MIOperandInfo = (ops R600_Reg32:$gpr, InstFlag:$neg, InstFlag:$rel,
+ InstFlag:$abs);
+}
+def ALUSrcOp2_I32 : ALUSrcOp2Base <i32, i32imm>;
+def ALUSrcOp2_F32 : ALUSrcOp2Base <f32, f32imm>;
+
def ADDRParam : ComplexPattern<i32, 2, "SelectADDRParam", [], []>;
def ADDRDWord : ComplexPattern<i32, 1, "SelectADDRDWord", [], []>;
def ADDRVTX_READ : ComplexPattern<i32, 2, "SelectADDRVTX_READ", [], []>;
+def ADDRFrame : ComplexPattern<i32, 2, "SelectADDRFrame", [], []>;
+class ALUDstPat<ValueType vt> : ComplexPattern<vt, 5, "SelectALUDst", [], []>;
+class ALUSrcOp2Pat<ValueType vt> : ComplexPattern<vt, 4,
+ "SelectALUSrcOp2", [], []>;
-class R600_ALU {
- bits<7> DST_GPR = 0;
- bits<9> SRC0_SEL = 0;
- bits<1> SRC0_NEG = 0;
- bits<9> SRC1_SEL = 0;
- bits<1> SRC1_NEG = 0;
- bits<1> CLAMP = 0;
-
+class R600ALU_Word0 {
+ field bits<64> Inst;
+
+ bits<11> src0;
+ bits<1> src0_neg;
+ bits<1> src0_rel;
+ bits<11> src1;
+ bits<1> src1_rel;
+ bits<1> src1_neg;
+ bits<3> index_mode;
+ bits<2> pred_sel;
+ bits<1> last;
+
+ let Inst{8-0} = src0{8-0};
+ let Inst{9} = src0_rel;
+ let Inst{11-10} = src0{10-9};
+ let Inst{12} = src0_neg;
+ let Inst{21-13} = src1{8-0};
+ let Inst{22} = src1_rel;
+ let Inst{24-23} = src1{10-9};
+ let Inst{25} = src1_neg;
+ let Inst{28-26} = index_mode;
+ let Inst{30-29} = pred_sel;
+ let Inst{31} = last;
+}
+
+class R600ALU_Word1_OP2 <bits<11> alu_inst> {
+ field bits<64> Inst;
+
+ bits<1> src0_abs;
+ bits<1> src1_abs;
+ bits<1> update_exec_mask;
+ bits<1> update_pred;
+ bits<1> write_mask;
+ bits<2> omod;
+ bits<3> bank_swizzle;
+ bits<9> dst;
+ bits<1> dst_rel;
+ bits<1> clamp;
+
+ let Inst{32} = src0_abs;
+ let Inst{33} = src1_abs;
+ let Inst{34} = update_exec_mask;
+ let Inst{35} = update_pred;
+ let Inst{36} = write_mask;
+ let Inst{38-37} = omod;
+ let Inst{49-39} = alu_inst;
+ let Inst{52-50} = bank_swizzle;
+ let Inst{59-53} = dst{6-0};
+ let Inst{60} = dst_rel;
+ let Inst{62-61} = dst{8-7};
+ let Inst{63} = clamp;
}
+class R600ALU_OP2 <bits<11> inst> {
+ field bits<64> Inst;
+
+ bits<11> dst;
+ bits<11> src0;
+ bits<1> src0_neg;
+ bits<1> src0_rel;
+ bits<1> src0_abs;
+ bits<11> src1;
+ bits<1> src1_neg;
+ bits<1> src1_rel;
+ bits<1> src1_abs;
+ bits<2> pred_sel;
+
+ // For now, the following fields are manually encoded in R600MCCodeEmitter:
+ bits<1> update_exec_mask = 0;
+ bits<1> update_pred = 0;
+ bits<1> write_mask = 0;
+ bits<2> omod = 0;
+ bits<3> bank_swizzle = 0;
+ bits<1> dst_rel = 0;
+ bits<1> clamp = 0;
+ bits<3> index_mode = 0;
+ bits<1> last = 0;
+
+ let Inst{8-0} = src0{8-0};
+ let Inst{9} = src0_rel;
+ let Inst{11-10} = src0{10-9};
+ let Inst{12} = src0_neg;
+ let Inst{21-13} = src1{8-0};
+ let Inst{22} = src1_rel;
+ let Inst{24-23} = src1{10-9};
+ let Inst{25} = src1_neg;
+ let Inst{28-26} = index_mode;
+ let Inst{30-29} = pred_sel;
+ let Inst{31} = last;
+
+ let Inst{32} = src0_abs;
+ let Inst{33} = src1_abs;
+ let Inst{34} = update_exec_mask;
+ let Inst{35} = update_pred;
+ let Inst{36} = write_mask;
+ let Inst{38-37} = omod;
+ let Inst{49-39} = inst;
+ let Inst{52-50} = bank_swizzle;
+ let Inst{59-53} = dst{6-0};
+ let Inst{60} = dst_rel;
+ let Inst{62-61} = dst{10-9};
+ let Inst{63} = clamp;
+// R600ALU_Word0,
+// R600ALU_Word1_OP2 <inst>;
+}
+
+/*
+class R600ALU_Word1_OP2_r600 : R600ALU_Word1_OP2 {
+
+ bits<1> fog_merge;
+ bits<10> alu_inst;
+
+ let Inst{37} = fog_merge;
+ let Inst{39-38} = omod;
+ let Inst{49-40} = alu_inst;
+}
+
+
+class R600ALU_Word1_OP2_r700 : R600ALU_Word1_OP2 {
+
+ bits<11> alu_inst;
+
+ let Inst{38-37} = omod;
+ let Inst{49-39} = alu_inst;
+}
+
+*/
+
def R600_Pred : PredicateOperand<i32, (ops R600_Predicate),
(ops PRED_SEL_OFF)>;
@@ -82,18 +226,33 @@ let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
class R600_1OP <bits<11> inst, string opName, list<dag> pattern,
InstrItinClass itin = AnyALU> :
- InstR600 <inst,
- (outs R600_Reg32:$dst),
- (ins R600_Reg32:$src, R600_Pred:$p, variable_ops),
- !strconcat(opName, " $dst, $src $p"),
- pattern,
- itin>{
- bits<7> dst;
- bits<9> src;
- let Inst{8-0} = src;
- let Inst{49-39} = inst;
- let Inst{59-53} = dst;
- }
+ InstR600 <0,
+ (outs R600_Reg32:$dst),
+ (ins R600_Reg32:$src0, InstFlag:$src0_neg, InstFlag:$src0_rel,
+ InstFlag:$src0_abs, R600_Pred:$pred_sel, InstFlag:$literal),
+ !strconcat(opName, " $dst, $src0 (neg$src0_neg rel$src0_rel "
+ "abs$src0_abs) $pred_sel, $literal"),
+ pattern,
+ itin>,
+ R600ALU_OP2<inst> {
+
+ let src1 = 0;
+ let src1_rel = 0;
+ let src1_neg = 0;
+ let src1_abs = 0;
+}
+
+class R600_1OP_F32 <bits<11> inst, string opName, SDPatternOperator node,
+ InstrItinClass itin = AnyALU> :
+ R600_1OP <inst, opName,
+ [(set R600_Reg32:$dst, (node ALUSrcOp2Pat<f32>:$src0))]
+>;
+
+class R600_1OP_I32 <bits<11> inst, string opName, SDPatternOperator node,
+ InstrItinClass itin = AnyALU> :
+ R600_1OP <inst, opName,
+ [(set R600_Reg32:$dst, (node ALUSrcOp2Pat<i32>:$src0))]
+>;
class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
InstrItinClass itin = AnyALU> :
@@ -383,30 +542,15 @@ def SNE : R600_2OP <
COND_NE))]
>;
-def FRACT : R600_1OP <
- 0x10, "FRACT",
- [(set R600_Reg32:$dst, (AMDGPUfract R600_Reg32:$src))]
->;
+def FRACT : R600_1OP_F32 <0x10, "FRACT", AMDGPUfract>;
-def TRUNC : R600_1OP <
- 0x11, "TRUNC",
- [(set R600_Reg32:$dst, (int_AMDGPU_trunc R600_Reg32:$src))]
->;
+def TRUNC : R600_1OP_F32 <0x11, "TRUNC", int_AMDGPU_trunc>;
-def CEIL : R600_1OP <
- 0x12, "CEIL",
- [(set R600_Reg32:$dst, (fceil R600_Reg32:$src))]
->;
+def CEIL : R600_1OP_F32 <0x12, "CEIL", fceil>;
-def RNDNE : R600_1OP <
- 0x13, "RNDNE",
- [(set R600_Reg32:$dst, (frint R600_Reg32:$src))]
->;
+def RNDNE : R600_1OP_F32 <0x13, "RNDNE", frint>;
-def FLOOR : R600_1OP <
- 0x14, "FLOOR",
- [(set R600_Reg32:$dst, (ffloor R600_Reg32:$src))]
->;
+def FLOOR : R600_1OP_F32 <0x14, "FLOOR", ffloor>;
let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
@@ -482,10 +626,7 @@ def XOR_INT : R600_2OP <
[(set R600_Reg32:$dst, (xor R600_Reg32:$src0, R600_Reg32:$src1))]
>;
-def NOT_INT : R600_1OP <
- 0x33, "NOT_INT",
- [(set R600_Reg32:$dst, (not R600_Reg32:$src))]
->;
+def NOT_INT : R600_1OP_I32 <0x33, "NOT_INT", not>;
def ADD_INT : R600_2OP <
0x34, "ADD_INT",
@@ -739,39 +880,32 @@ multiclass CUBE_Common <bits<11> inst> {
}
}
} // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
-class EXP_IEEE_Common <bits<11> inst> : R600_1OP <
- inst, "EXP_IEEE",
- [(set R600_Reg32:$dst, (fexp2 R600_Reg32:$src))]
+class EXP_IEEE_Common <bits<11> inst> : R600_1OP_F32 <
+ inst, "EXP_IEEE", fexp2
>;
-class FLT_TO_INT_Common <bits<11> inst> : R600_1OP <
- inst, "FLT_TO_INT",
- [(set R600_Reg32:$dst, (fp_to_sint R600_Reg32:$src))]
+class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_F32 <
+ inst, "FLT_TO_INT", fp_to_sint
>;
-class INT_TO_FLT_Common <bits<11> inst> : R600_1OP <
- inst, "INT_TO_FLT",
- [(set R600_Reg32:$dst, (sint_to_fp R600_Reg32:$src))]
+class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_I32 <
+ inst, "INT_TO_FLT", sint_to_fp
>;
-class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP <
- inst, "FLT_TO_UINT",
- [(set R600_Reg32:$dst, (fp_to_uint R600_Reg32:$src))]
+class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_F32 <
+ inst, "FLT_TO_UINT", fp_to_uint
>;
-class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP <
- inst, "UINT_TO_FLT",
- [(set R600_Reg32:$dst, (uint_to_fp R600_Reg32:$src))]
+class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_I32 <
+ inst, "UINT_TO_FLT", uint_to_fp
>;
class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
- inst, "LOG_CLAMPED",
- []
+ inst, "LOG_CLAMPED", []
>;
-class LOG_IEEE_Common <bits<11> inst> : R600_1OP <
- inst, "LOG_IEEE",
- [(set R600_Reg32:$dst, (flog2 R600_Reg32:$src))]
+class LOG_IEEE_Common <bits<11> inst> : R600_1OP_F32 <
+ inst, "LOG_IEEE", flog2
>;
class LSHL_Common <bits<11> inst> : R600_2OP <
@@ -810,28 +944,23 @@ class MULLO_UINT_Common <bits<11> inst> : R600_2OP <
>;
class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
- inst, "RECIP_CLAMPED",
- []
+ inst, "RECIP_CLAMPED", []
>;
-class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
- inst, "RECIP_IEEE",
- [(set R600_Reg32:$dst, (int_AMDGPU_rcp R600_Reg32:$src))]
+class RECIP_IEEE_Common <bits<11> inst> : R600_1OP_F32 <
+ inst, "RECIP_IEEE", int_AMDGPU_rcp
>;
-class RECIP_UINT_Common <bits<11> inst> : R600_1OP <
- inst, "RECIP_INT $dst, $src",
- [(set R600_Reg32:$dst, (AMDGPUurecip R600_Reg32:$src))]
+class RECIP_UINT_Common <bits<11> inst> : R600_1OP_I32 <
+ inst, "RECIP_UINT", AMDGPUurecip
>;
-class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP <
- inst, "RECIPSQRT_CLAMPED",
- [(set R600_Reg32:$dst, (int_AMDGPU_rsq R600_Reg32:$src))]
+class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_F32 <
+ inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
>;
class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
- inst, "RECIPSQRT_IEEE",
- []
+ inst, "RECIPSQRT_IEEE", []
>;
class SIN_Common <bits<11> inst> : R600_1OP <
@@ -1037,11 +1166,11 @@ let Predicates = [isEGorCayman] in {
// XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
// which do not need to be truncated since the fp values are 0.0f or 1.0f.
// We should look into handling these cases separately.
- def : Pat<(fp_to_sint R600_Reg32:$src),
- (FLT_TO_INT_eg (TRUNC R600_Reg32:$src))>;
+ def : Pat<(fp_to_sint ALUSrcOp2Pat<f32>:$src0),
+ (FLT_TO_INT_eg (TRUNC R600_Reg32:$src0))>;
- def : Pat<(fp_to_uint R600_Reg32:$src),
- (FLT_TO_UINT_eg (TRUNC R600_Reg32:$src))>;
+ def : Pat<(fp_to_uint ALUSrcOp2Pat<f32>:$src0),
+ (FLT_TO_UINT_eg (TRUNC R600_Reg32:$src0))>;
def : Pat<(fsqrt R600_Reg32:$src),
(MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_eg R600_Reg32:$src))>;
@@ -1286,6 +1415,26 @@ def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
}
+//===----------------------------------------------------------------------===//
+// Frame loads and stores
+//===----------------------------------------------------------------------===//
+
+let usesCustomInserter = 1, isPseudo = 1, isCodeGenOnly =1 in {
+
+def FrameLoad_i32 : InstR600 <0x0,
+ (outs R600_Reg32:$dst), (ins FRAMEri:$addr),
+ "FrameLoad_i32 $dst, $addr",
+ [(set (i32 R600_Reg32:$dst), (load ADDRFrame:$addr))], NullALU
+>;
+
+def FrameStore_i32 : InstR600 <0x0,
+ (outs), (ins R600_Reg32:$val, FRAMEri:$addr),
+ "FrameStore_i32 $val, $addr",
+ [(store (i32 R600_Reg32:$val), ADDRFrame:$addr)], NullALU
+>;
+
+} // End usesCustomInserter = 1, isPseudo = 1, isCodeGenOnly = 1
+
let Predicates = [isCayman] in {
let isVector = 1 in {
generated by cgit v1.2.3 (git 2.39.1) at 2024-04-25 10:33:41 +0000