summaryrefslogtreecommitdiffstats
path: root/llvm/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
diff options
context:
space:
mode:
authorMatt Arsenault <Matthew.Arsenault@amd.com>2014-09-15 15:41:53 +0000
committerMatt Arsenault <Matthew.Arsenault@amd.com>2014-09-15 15:41:53 +0000
commit3f98140c87728cec3247e2dbe0a9990e48afa984 (patch)
tree221a264a783c12b722ae3e15e68f0798a35ce101 /llvm/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
parent65f67e4dfe837d8aff8d9b30cecc7247466d6f52 (diff)
downloadbcm5719-llvm-3f98140c87728cec3247e2dbe0a9990e48afa984.tar.gz
bcm5719-llvm-3f98140c87728cec3247e2dbe0a9990e48afa984.zip
R600/SI: Add preliminary support for flat address space
llvm-svn: 217777
Diffstat (limited to 'llvm/lib/Target/R600/AMDGPUISelDAGToDAG.cpp')
-rw-r--r--llvm/lib/Target/R600/AMDGPUISelDAGToDAG.cpp76
1 files changed, 75 insertions, 1 deletions
diff --git a/llvm/lib/Target/R600/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
index ed86a18e5cb..4fa576f8055 100644
--- a/llvm/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
+++ b/llvm/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
@@ -65,6 +65,7 @@ private:
static bool checkPrivateAddress(const MachineMemOperand *Op);
static bool isGlobalStore(const StoreSDNode *N);
+ static bool isFlatStore(const StoreSDNode *N);
static bool isPrivateStore(const StoreSDNode *N);
static bool isLocalStore(const StoreSDNode *N);
static bool isRegionStore(const StoreSDNode *N);
@@ -72,6 +73,7 @@ private:
bool isCPLoad(const LoadSDNode *N) const;
bool isConstantLoad(const LoadSDNode *N, int cbID) const;
bool isGlobalLoad(const LoadSDNode *N) const;
+ bool isFlatLoad(const LoadSDNode *N) const;
bool isParamLoad(const LoadSDNode *N) const;
bool isPrivateLoad(const LoadSDNode *N) const;
bool isLocalLoad(const LoadSDNode *N) const;
@@ -104,6 +106,7 @@ private:
bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset,
SDValue &Offset, SDValue &GLC, SDValue &SLC,
SDValue &TFE) const;
+ SDNode *SelectAddrSpaceCast(SDNode *N);
bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
SDValue &Clamp, SDValue &Omod) const;
@@ -484,6 +487,8 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
case AMDGPUISD::DIV_SCALE: {
return SelectDIV_SCALE(N);
}
+ case ISD::ADDRSPACECAST:
+ return SelectAddrSpaceCast(N);
}
return SelectCode(N);
}
@@ -522,6 +527,10 @@ bool AMDGPUDAGToDAGISel::isLocalStore(const StoreSDNode *N) {
return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
}
+bool AMDGPUDAGToDAGISel::isFlatStore(const StoreSDNode *N) {
+ return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
+}
+
bool AMDGPUDAGToDAGISel::isRegionStore(const StoreSDNode *N) {
return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
}
@@ -553,6 +562,10 @@ bool AMDGPUDAGToDAGISel::isLocalLoad(const LoadSDNode *N) const {
return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
}
+bool AMDGPUDAGToDAGISel::isFlatLoad(const LoadSDNode *N) const {
+ return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
+}
+
bool AMDGPUDAGToDAGISel::isRegionLoad(const LoadSDNode *N) const {
return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
}
@@ -582,10 +595,11 @@ bool AMDGPUDAGToDAGISel::isPrivateLoad(const LoadSDNode *N) const {
const Value *MemVal = N->getMemOperand()->getValue();
if (!checkType(MemVal, AMDGPUAS::LOCAL_ADDRESS) &&
!checkType(MemVal, AMDGPUAS::GLOBAL_ADDRESS) &&
+ !checkType(MemVal, AMDGPUAS::FLAT_ADDRESS) &&
!checkType(MemVal, AMDGPUAS::REGION_ADDRESS) &&
!checkType(MemVal, AMDGPUAS::CONSTANT_ADDRESS) &&
!checkType(MemVal, AMDGPUAS::PARAM_D_ADDRESS) &&
- !checkType(MemVal, AMDGPUAS::PARAM_I_ADDRESS)){
+ !checkType(MemVal, AMDGPUAS::PARAM_I_ADDRESS)) {
return true;
}
return false;
@@ -1005,6 +1019,66 @@ bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
return false;
}
+// FIXME: This is incorrect and only enough to be able to compile.
+SDNode *AMDGPUDAGToDAGISel::SelectAddrSpaceCast(SDNode *N) {
+ AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(N);
+ SDLoc DL(N);
+
+ assert(Subtarget.hasFlatAddressSpace() &&
+ "addrspacecast only supported with flat address space!");
+
+ assert((ASC->getSrcAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS &&
+ ASC->getDestAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS) &&
+ "Cannot cast address space to / from constant address!");
+
+ assert((ASC->getSrcAddressSpace() == AMDGPUAS::FLAT_ADDRESS ||
+ ASC->getDestAddressSpace() == AMDGPUAS::FLAT_ADDRESS) &&
+ "Can only cast to / from flat address space!");
+
+ // The flat instructions read the address as the index of the VGPR holding the
+ // address, so casting should just be reinterpreting the base VGPR, so just
+ // insert trunc / bitcast / zext.
+
+ SDValue Src = ASC->getOperand(0);
+ EVT DestVT = ASC->getValueType(0);
+ EVT SrcVT = Src.getValueType();
+
+ unsigned SrcSize = SrcVT.getSizeInBits();
+ unsigned DestSize = DestVT.getSizeInBits();
+
+ if (SrcSize > DestSize) {
+ assert(SrcSize == 64 && DestSize == 32);
+ return CurDAG->getMachineNode(
+ TargetOpcode::EXTRACT_SUBREG,
+ DL,
+ DestVT,
+ Src,
+ CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32));
+ }
+
+
+ if (DestSize > SrcSize) {
+ assert(SrcSize == 32 && DestSize == 64);
+
+ SDValue RC = CurDAG->getTargetConstant(AMDGPU::VSrc_64RegClassID, MVT::i32);
+
+ const SDValue Ops[] = {
+ RC,
+ Src,
+ CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
+ SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
+ CurDAG->getConstant(0, MVT::i32)), 0),
+ CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32)
+ };
+
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
+ SDLoc(N), N->getValueType(0), Ops);
+ }
+
+ assert(SrcSize == 64 && DestSize == 64);
+ return CurDAG->getNode(ISD::BITCAST, DL, DestVT, Src).getNode();
+}
+
bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src,
SDValue &SrcMods) const {
OpenPOWER on IntegriCloud