AMDGPU: Use assert zext for workgroup sizes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@254328 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/SIISelLowering.cpp b/lib/Target/AMDGPU/SIISelLowering.cpp
index e2c644451b43c6ad083fb138a31a57b153d425d8..94fad32c007093529abd27db3ee5c1bc0cacd0bf 100644 (file)
--- a/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -1043,6 +1043,18 @@ SDValue SITargetLowering::copyToM0(SelectionDAG &DAG, SDValue Chain, SDLoc DL,
                                                       // a glue result.
 }
 
+SDValue SITargetLowering::lowerImplicitZextParam(SelectionDAG &DAG,
+                                                 SDValue Op,
+                                                 MVT VT,
+                                                 unsigned Offset) const {
+  SDLoc SL(Op);
+  SDValue Param = LowerParameter(DAG, MVT::i32, MVT::i32, SL,
+                                 DAG.getEntryNode(), Offset, false);
+  // The local size values will have the hi 16-bits as zero.
+  return DAG.getNode(ISD::AssertZext, SL, MVT::i32, Param,
+                     DAG.getValueType(VT));
+}
+
 SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
                                                   SelectionDAG &DAG) const {
   MachineFunction &MF = DAG.getMachineFunction();
@@ -1080,19 +1092,18 @@ SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
     return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
                           SI::KernelInputOffsets::GLOBAL_SIZE_Z, false);
   case Intrinsic::r600_read_local_size_x:
-    return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
-                          SI::KernelInputOffsets::LOCAL_SIZE_X, false);
+    return lowerImplicitZextParam(DAG, Op, MVT::i16,
+                                  SI::KernelInputOffsets::LOCAL_SIZE_X);
   case Intrinsic::r600_read_local_size_y:
-    return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
-                          SI::KernelInputOffsets::LOCAL_SIZE_Y, false);
+    return lowerImplicitZextParam(DAG, Op, MVT::i16,
+                                  SI::KernelInputOffsets::LOCAL_SIZE_Y);
   case Intrinsic::r600_read_local_size_z:
-    return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
-                          SI::KernelInputOffsets::LOCAL_SIZE_Z, false);
-
+    return lowerImplicitZextParam(DAG, Op, MVT::i16,
+                                  SI::KernelInputOffsets::LOCAL_SIZE_Z);
   case Intrinsic::AMDGPU_read_workdim:
-    return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
-                          getImplicitParameterOffset(MFI, GRID_DIM), false);
-
+    // Really only 2 bits.
+    return lowerImplicitZextParam(DAG, Op, MVT::i8,
+                                  getImplicitParameterOffset(MFI, GRID_DIM));
   case Intrinsic::r600_read_tgid_x:
     return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
       TRI->getPreloadedValue(MF, SIRegisterInfo::TGID_X), VT);

diff --git a/lib/Target/AMDGPU/SIISelLowering.h b/lib/Target/AMDGPU/SIISelLowering.h
index a8b8ad34ed9db1c989025cab5c9950c6486b6005..a358e3fc3c0d1b6c7a82f89c165ba6f9dbfb1432 100644 (file)
--- a/lib/Target/AMDGPU/SIISelLowering.h
+++ b/lib/Target/AMDGPU/SIISelLowering.h
@@ -28,6 +28,9 @@ class SITargetLowering : public AMDGPUTargetLowering {
   SDValue LowerGlobalAddress(AMDGPUMachineFunction *MFI, SDValue Op,
                              SelectionDAG &DAG) const override;
 
+  SDValue lowerImplicitZextParam(SelectionDAG &DAG, SDValue Op,
+                                 MVT VT, unsigned Offset) const;
+
   SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const;

diff --git a/test/CodeGen/AMDGPU/work-item-intrinsics.ll b/test/CodeGen/AMDGPU/work-item-intrinsics.ll
index 4328e964c1bf83f8984bc6ebe6efe8ea631dee02..ebe73429da9c0e33dc49be3aeb16454a34400661 100644 (file)
--- a/test/CodeGen/AMDGPU/work-item-intrinsics.ll
+++ b/test/CodeGen/AMDGPU/work-item-intrinsics.ll
@@ -213,6 +213,66 @@ entry:
   ret void
 }
 
+; FUNC-LABEL: {{^}}local_size_x_known_bits:
+; SI: s_load_dword [[VAL:s[0-9]+]], s[0:1], 0x6
+; VI: s_load_dword [[VAL:s[0-9]+]], s[0:1], 0x18
+; GCN-NOT: 0xffff
+; GCN: v_mov_b32_e32 [[VVAL:v[0-9]+]], [[VAL]]
+; GCN-NEXT: buffer_store_dword [[VVAL]]
+define void @local_size_x_known_bits(i32 addrspace(1)* %out) {
+entry:
+  %size = call i32 @llvm.r600.read.local.size.x() #0
+  %shl = shl i32 %size, 16
+  %shr = lshr i32 %shl, 16
+  store i32 %shr, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}local_size_y_known_bits:
+; SI: s_load_dword [[VAL:s[0-9]+]], s[0:1], 0x7
+; VI: s_load_dword [[VAL:s[0-9]+]], s[0:1], 0x1c
+; GCN-NOT: 0xffff
+; GCN: v_mov_b32_e32 [[VVAL:v[0-9]+]], [[VAL]]
+; GCN-NEXT: buffer_store_dword [[VVAL]]
+define void @local_size_y_known_bits(i32 addrspace(1)* %out) {
+entry:
+  %size = call i32 @llvm.r600.read.local.size.y() #0
+  %shl = shl i32 %size, 16
+  %shr = lshr i32 %shl, 16
+  store i32 %shr, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}local_size_z_known_bits:
+; SI: s_load_dword [[VAL:s[0-9]+]], s[0:1], 0x8
+; VI: s_load_dword [[VAL:s[0-9]+]], s[0:1], 0x20
+; GCN-NOT: 0xffff
+; GCN: v_mov_b32_e32 [[VVAL:v[0-9]+]], [[VAL]]
+; GCN-NEXT: buffer_store_dword [[VVAL]]
+define void @local_size_z_known_bits(i32 addrspace(1)* %out) {
+entry:
+  %size = call i32 @llvm.r600.read.local.size.z() #0
+  %shl = shl i32 %size, 16
+  %shr = lshr i32 %shl, 16
+  store i32 %shr, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}get_work_dim_known_bits:
+; SI: s_load_dword [[VAL:s[0-9]+]], s[0:1], 0xb
+; VI: s_load_dword [[VAL:s[0-9]+]], s[0:1], 0x2c
+; GCN-NOT: 0xff
+; GCN: v_mov_b32_e32 [[VVAL:v[0-9]+]], [[VAL]]
+; GCN: buffer_store_dword [[VVAL]]
+define void @get_work_dim_known_bits(i32 addrspace(1)* %out) {
+entry:
+  %dim = call i32 @llvm.AMDGPU.read.workdim() #0
+  %shl = shl i32 %dim, 24
+  %shr = lshr i32 %shl, 24
+  store i32 %shr, i32 addrspace(1)* %out
+  ret void
+}
+
 declare i32 @llvm.r600.read.ngroups.x() #0
 declare i32 @llvm.r600.read.ngroups.y() #0
 declare i32 @llvm.r600.read.ngroups.z() #0