R600/SI: Add a calling convention for compute shaders

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@183137 91177308-0d34-0410-b5e6-96231b3b80d8

10 files changed, 49 insertions, 19 deletions

diff --git a/lib/Target/R600/AMDGPUCallingConv.td b/lib/Target/R600/AMDGPUCallingConv.td
index 9c30515..e57b5cd 100644
--- a/lib/Target/R600/AMDGPUCallingConv.td
+++ b/lib/Target/R600/AMDGPUCallingConv.td
@@ -32,17 +32,21 @@ def CC_SI : CallingConv<[
     VGPR8, VGPR9, VGPR10, VGPR11, VGPR12, VGPR13, VGPR14, VGPR15,
     VGPR16, VGPR17, VGPR18, VGPR19, VGPR20, VGPR21, VGPR22, VGPR23,
     VGPR24, VGPR25, VGPR26, VGPR27, VGPR28, VGPR29, VGPR30, VGPR31
-  ]>>>,
+  ]>>>
+
+]>;
 
-  // This is the default for i64 values.
-  // XXX: We should change this once clang understands the CC_AMDGPU.
-  CCIfType<[i64], CCAssignToRegWithShadow<
-   [ SGPR0, SGPR2, SGPR4, SGPR6, SGPR8, SGPR10, SGPR12, SGPR14 ],
-   [ SGPR1, SGPR3, SGPR5, SGPR7, SGPR9, SGPR11, SGPR13, SGPR15 ]
-  >>
+// Calling convention for SI compute kernels
+def CC_SI_Kernel : CallingConv<[
+  CCIfType<[i64],      CCAssignToStack <8, 4>>,
+  CCIfType<[i32, f32], CCAssignToStack <4, 4>>,
+  CCIfType<[i16],      CCAssignToStack <2, 4>>,
+  CCIfType<[i8],       CCAssignToStack <1, 4>>
 ]>;
 
 def CC_AMDGPU : CallingConv<[
+  CCIf<"State.getMachineFunction().getInfo<SIMachineFunctionInfo>()->"#
+       "ShaderType == ShaderType::COMPUTE", CCDelegateTo<CC_SI_Kernel>>,
   CCIf<"State.getTarget().getSubtarget<AMDGPUSubtarget>().device()"#
        "->getGeneration() == AMDGPUDeviceInfo::HD7XXX", CCDelegateTo<CC_SI>>
 ]>;
diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index 4a6f4cc..02d6fab 100644
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -14,9 +14,11 @@
 //===----------------------------------------------------------------------===//
 
 #include "AMDGPUISelLowering.h"
+#include "AMDGPU.h"
 #include "AMDGPURegisterInfo.h"
 #include "AMDGPUSubtarget.h"
 #include "AMDILIntrinsicInfo.h"
+#include "SIMachineFunctionInfo.h"
 #include "llvm/CodeGen/CallingConvLower.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index 825812f..65d5479 100644
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -121,7 +121,7 @@ SDValue SITargetLowering::LowerFormalArguments(
     }
 
     // Second split vertices into their elements
-    if (Arg.VT.isVector()) {
+    if (Info->ShaderType != ShaderType::COMPUTE && Arg.VT.isVector()) {
       ISD::InputArg NewArg = Arg;
       NewArg.Flags.setSplit();
       NewArg.VT = Arg.VT.getVectorElementType();
@@ -153,6 +153,14 @@ SDValue SITargetLowering::LowerFormalArguments(
     CCInfo.AllocateReg(AMDGPU::VGPR1);
   }
 
+  unsigned ArgReg = 0;
+  // The pointer to the list of arguments is stored in SGPR0, SGPR1
+  if (Info->ShaderType == ShaderType::COMPUTE) {
+    CCInfo.AllocateReg(AMDGPU::SGPR0);
+    CCInfo.AllocateReg(AMDGPU::SGPR1);
+    ArgReg = MF.addLiveIn(AMDGPU::SGPR0_SGPR1, &AMDGPU::SReg_64RegClass);
+  }
+
   AnalyzeFormalArguments(CCInfo, Splits);
 
   for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
@@ -164,10 +172,26 @@ SDValue SITargetLowering::LowerFormalArguments(
     }
 
     CCValAssign &VA = ArgLocs[ArgIdx++];
+    EVT VT = VA.getLocVT();
+
+    if (VA.isMemLoc()) {
+      assert(ArgReg);
+      PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
+                                            AMDGPUAS::CONSTANT_ADDRESS);
+      EVT ArgVT = MVT::getIntegerVT(VT.getSizeInBits());
+      SDValue BasePtr =  DAG.getCopyFromReg(DAG.getRoot(), DL,
+                                            ArgReg, MVT::i64);
+      SDValue Ptr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
+                               DAG.getConstant(VA.getLocMemOffset(), MVT::i64));
+      SDValue Arg = DAG.getExtLoad(ISD::ZEXTLOAD, DL, VT, DAG.getRoot(), Ptr,
+                                MachinePointerInfo(UndefValue::get(PtrTy)),
+                                VA.getValVT(), false, false, ArgVT.getSizeInBits() >> 3);
+      InVals.push_back(Arg);
+      continue;
+    }
     assert(VA.isRegLoc() && "Parameter must be in a register!");
 
     unsigned Reg = VA.getLocReg();
-    MVT VT = VA.getLocVT();
 
     if (VT == MVT::i64) {
       // For now assume it is a pointer
diff --git a/test/CodeGen/R600/bfi_int.ll b/test/CodeGen/R600/bfi_int.ll
index 707abb3..a1bd09a 100644
--- a/test/CodeGen/R600/bfi_int.ll
+++ b/test/CodeGen/R600/bfi_int.ll
@@ -38,8 +38,8 @@ entry:
 ; R600-CHECK: @bfi_sha256_ma
 ; R600-CHECK: XOR_INT * [[DST:T[0-9]+\.[XYZW]]], {{T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ; R600-CHECK: BFI_INT * {{T[0-9]+\.[XYZW]}}, {{[[DST]]|PV\.[XYZW]}}, {{T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
-; SI-CHECK: V_XOR_B32_e32 [[DST:VGPR[0-9]+]], {{VGPR[0-9]+, VGPR[0-9]+}}
-; SI-CHECK: V_BFI_B32 {{VGPR[0-9]+}}, [[DST]], {{VGPR[0-9]+, VGPR[0-9]+}}
+; SI-CHECK: V_XOR_B32_e64 [[DST:VGPR[0-9]+]], {{[SV]GPR[0-9]+, [SV]GPR[0-9]+}}
+; SI-CHECK: V_BFI_B32 {{VGPR[0-9]+}}, [[DST]], {{[SV]GPR[0-9]+, [SV]GPR[0-9]+}}
 
 define void @bfi_sha256_ma(i32 addrspace(1)* %out, i32 %x, i32 %y, i32 %z) {
 entry:
diff --git a/test/CodeGen/R600/lshl.ll b/test/CodeGen/R600/lshl.ll
index fb698da..9e29b0d 100644
--- a/test/CodeGen/R600/lshl.ll
+++ b/test/CodeGen/R600/lshl.ll
@@ -1,6 +1,6 @@
 ;RUN: llc < %s -march=r600 -mcpu=verde | FileCheck %s
 
-;CHECK: V_LSHLREV_B32_e32 VGPR0, 1, VGPR0
+;CHECK: V_LSHL_B32_e64 VGPR{{[0-9]+}}, {{[SV]GPR[0-9]+}}, 1
 
 define void @test(i32 %p) {
    %i = mul i32 %p, 2
diff --git a/test/CodeGen/R600/lshr.ll b/test/CodeGen/R600/lshr.ll
index e0ed3ac..eab3fbf 100644
--- a/test/CodeGen/R600/lshr.ll
+++ b/test/CodeGen/R600/lshr.ll
@@ -1,6 +1,6 @@
 ;RUN: llc < %s -march=r600 -mcpu=verde | FileCheck %s
 
-;CHECK: V_LSHRREV_B32_e32 VGPR0, 1, VGPR0
+;CHECK: V_LSHR_B32_e64 {{VGPR[0-9]+}}, {{[SV]GPR[0-9]+}}, 1
 
 define void @test(i32 %p) {
    %i = udiv i32 %p, 2
diff --git a/test/CodeGen/R600/mulhu.ll b/test/CodeGen/R600/mulhu.ll
index bc17a59..eb379d1 100644
--- a/test/CodeGen/R600/mulhu.ll
+++ b/test/CodeGen/R600/mulhu.ll
@@ -1,7 +1,7 @@
 ;RUN: llc < %s -march=r600 -mcpu=verde | FileCheck %s
 
-;CHECK: V_MOV_B32_e32 VGPR1, -1431655765
-;CHECK-NEXT: V_MUL_HI_U32 VGPR0, VGPR0, VGPR1, 0, 0, 0, 0, 0
+;CHECK: V_MOV_B32_e32 VGPR{{[0-9]+}}, -1431655765
+;CHECK: V_MUL_HI_U32 VGPR0, {{[SV]GPR[0-9]+}}, {{VGPR[0-9]+}}
 ;CHECK-NEXT: V_LSHRREV_B32_e32 VGPR0, 1, VGPR0
 
 define void @test(i32 %p) {
diff --git a/test/CodeGen/R600/rotr.ll b/test/CodeGen/R600/rotr.ll
index efbdcbf..960d30d 100644
--- a/test/CodeGen/R600/rotr.ll
+++ b/test/CodeGen/R600/rotr.ll
@@ -22,8 +22,8 @@ entry:
 ; R600-CHECK: BIT_ALIGN_INT {{\** T[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW], PV.[XYZW]}}
 
 ; SI-CHECK: @rotl
-; SI-CHECK: V_SUB_I32_e32 [[DST:VGPR[0-9]+]], 32, {{VGPR[0-9]+}}
-; SI-CHECK: V_ALIGNBIT_B32 {{VGPR[0-9]+, VGPR[0-9]+, VGPR[0-9]+}}, [[DST]]
+; SI-CHECK: V_SUB_I32_e64 [[DST:VGPR[0-9]+]], 32, {{[SV]GPR[0-9]+}}
+; SI-CHECK: V_ALIGNBIT_B32 {{VGPR[0-9]+, [SV]GPR[0-9]+, VGPR[0-9]+}}, [[DST]]
 define void @rotl(i32 addrspace(1)* %in, i32 %x, i32 %y) {
 entry:
   %0 = shl i32 %x, %y
diff --git a/test/CodeGen/R600/seto.ll b/test/CodeGen/R600/seto.ll
index 4622203..19716f8 100644
--- a/test/CodeGen/R600/seto.ll
+++ b/test/CodeGen/R600/seto.ll
@@ -1,6 +1,6 @@
 ;RUN: llc < %s -march=r600 -mcpu=verde | FileCheck %s
 
-;CHECK: V_CMP_O_F32_e64 SGPR0_SGPR1, VGPR0, VGPR0, 0, 0, 0, 0
+;CHECK: V_CMP_O_F32_e64 SGPR0_SGPR1, {{[SV]GPR[0-9]+, [SV]GPR[0-9]+}}, 0, 0, 0, 0
 
 define void @main(float %p) {
 main_body:
diff --git a/test/CodeGen/R600/setuo.ll b/test/CodeGen/R600/setuo.ll
index 0bf5801..929dbb1 100644
--- a/test/CodeGen/R600/setuo.ll
+++ b/test/CodeGen/R600/setuo.ll
@@ -1,6 +1,6 @@
 ;RUN: llc < %s -march=r600 -mcpu=verde | FileCheck %s
 
-;CHECK: V_CMP_U_F32_e64 SGPR0_SGPR1, VGPR0, VGPR0, 0, 0, 0, 0
+;CHECK: V_CMP_U_F32_e64 SGPR0_SGPR1, {{[SV]GPR[0-9]+, [SV]GPR[0-9]+}}, 0, 0, 0, 0
 
 define void @main(float %p) {
 main_body: