R600/SI: Use bcnt instruction for ctpop

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@210567 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index cb4f8cde1bfe6142455da2ed4b89c58d6fe7741a..8ee453de3528aa0c66c6e3e88116ee3964d748aa 100644 (file)
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -220,12 +220,19 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
     setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
   }
 
+  for (MVT VT : { MVT::i32, MVT::i64 }) {
+    // TODO: Evergreen has BCNT_INT for CTPOP
+    setOperationAction(ISD::CTPOP, VT, Expand);
+    setOperationAction(ISD::CTTZ, VT, Expand);
+    setOperationAction(ISD::CTLZ, VT, Expand);
+  }
+
   static const MVT::SimpleValueType IntTypes[] = {
     MVT::v2i32, MVT::v4i32
   };
 
   for (MVT VT : IntTypes) {
-    //Expand the following operations for the current type by default
+    // Expand the following operations for the current type by default.
     setOperationAction(ISD::ADD,  VT, Expand);
     setOperationAction(ISD::AND,  VT, Expand);
     setOperationAction(ISD::FP_TO_SINT, VT, Expand);
@@ -244,6 +251,9 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
     setOperationAction(ISD::VSELECT, VT, Expand);
     setOperationAction(ISD::XOR,  VT, Expand);
     setOperationAction(ISD::BSWAP, VT, Expand);
+    setOperationAction(ISD::CTPOP, VT, Expand);
+    setOperationAction(ISD::CTTZ, VT, Expand);
+    setOperationAction(ISD::CTLZ, VT, Expand);
   }
 
   static const MVT::SimpleValueType FloatTypes[] = {

diff --git a/lib/Target/R600/AMDILISelLowering.cpp b/lib/Target/R600/AMDILISelLowering.cpp
index 7cea803f89871dbbff2a68aadaa67dbf868d92af..6001899ac8dd77cf284560ea2c5eca4aaa668f79 100644 (file)
--- a/lib/Target/R600/AMDILISelLowering.cpp
+++ b/lib/Target/R600/AMDILISelLowering.cpp
@@ -125,11 +125,6 @@ void AMDGPUTargetLowering::InitAMDILLowering() {
     setOperationAction(ISD::UMUL_LOHI, VT, Expand);
 
     setOperationAction(ISD::BSWAP, VT, Expand);
-
-    // GPU doesn't have any counting operators
-    setOperationAction(ISD::CTPOP, VT, Expand);
-    setOperationAction(ISD::CTTZ, VT, Expand);
-    setOperationAction(ISD::CTLZ, VT, Expand);
   }
 
   for (MVT VT : VectorTypes) {

diff --git a/lib/Target/R600/AMDILIntrinsics.td b/lib/Target/R600/AMDILIntrinsics.td
index 4a3e02e202bbb467a73f6707fe11c18b1b2f765e..d1eb6d6d6e7f38b728461e373025a61147ae8367 100644 (file)
--- a/lib/Target/R600/AMDILIntrinsics.td
+++ b/lib/Target/R600/AMDILIntrinsics.td
@@ -67,11 +67,8 @@ let TargetPrefix = "AMDIL", isTarget = 1 in {
 
 let TargetPrefix = "AMDIL", isTarget = 1 in {
   def int_AMDIL_abs : GCCBuiltin<"__amdil_abs">, UnaryIntInt;
-
   def int_AMDIL_bit_reverse_u32 : GCCBuiltin<"__amdil_ubit_reverse">,
           UnaryIntInt;
-  def int_AMDIL_bit_count_i32 : GCCBuiltin<"__amdil_count_bits">,
-          UnaryIntInt;
   def int_AMDIL_bit_find_first_lo : GCCBuiltin<"__amdil_ffb_lo">,
           UnaryIntInt;
   def int_AMDIL_bit_find_first_hi : GCCBuiltin<"__amdil_ffb_hi">,

diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index 608aad2d3ca0d1376184be0163b2c5f4ae6ec11d..48ce7d8b7ea0473836ac91c9bf1d1c56071acba4 100644 (file)
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -211,6 +211,8 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
     setOperationAction(ISD::FRINT, MVT::f64, Legal);
   }
 
+  setOperationAction(ISD::CTPOP, MVT::i32, Legal);
+
   setTargetDAGCombine(ISD::SELECT_CC);
   setTargetDAGCombine(ISD::SETCC);
 

diff --git a/lib/Target/R600/SIInstrInfo.cpp b/lib/Target/R600/SIInstrInfo.cpp
index dc96434919136a93721d1a4a38d6b3e25355bb75..fdebb2ffb4d541dcba36582566ce1379e0ae40c5 100644 (file)
--- a/lib/Target/R600/SIInstrInfo.cpp
+++ b/lib/Target/R600/SIInstrInfo.cpp
@@ -668,6 +668,7 @@ unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
   case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
   case AMDGPU::S_LOAD_DWORDX4_IMM:
   case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
+  case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e32;
   }
 }
 
@@ -1218,6 +1219,10 @@ void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
       // 3 to not hit an assertion later in MCInstLower.
       Inst->addOperand(MachineOperand::CreateImm(0));
       Inst->addOperand(MachineOperand::CreateImm(0));
+    } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
+      // The VALU version adds the second operand to the result, so insert an
+      // extra 0 operand.
+      Inst->addOperand(MachineOperand::CreateImm(0));
     }
 
     addDescImplicitUseDef(NewDesc, Inst);

diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 43ac47dad292f8284975385cb5e9f95ede735a99..6525b49f3a5155415c7f07f46439a4d4e2410c1e 100644 (file)
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -107,7 +107,9 @@ def S_BREV_B64 : SOP1_64 <0x0000000c, "S_BREV_B64", []>;
 
 ////def S_BCNT0_I32_B32 : SOP1_BCNT0 <0x0000000d, "S_BCNT0_I32_B32", []>;
 ////def S_BCNT0_I32_B64 : SOP1_BCNT0 <0x0000000e, "S_BCNT0_I32_B64", []>;
-////def S_BCNT1_I32_B32 : SOP1_BCNT1 <0x0000000f, "S_BCNT1_I32_B32", []>;
+def S_BCNT1_I32_B32 : SOP1_32 <0x0000000f, "S_BCNT1_I32_B32",
+  [(set i32:$dst, (ctpop i32:$src0))]
+>;
 ////def S_BCNT1_I32_B64 : SOP1_BCNT1 <0x00000010, "S_BCNT1_I32_B64", []>;
 ////def S_FF0_I32_B32 : SOP1_FF0 <0x00000011, "S_FF0_I32_B32", []>;
 ////def S_FF0_I32_B64 : SOP1_FF0 <0x00000012, "S_FF0_I32_B64", []>;
@@ -1217,7 +1219,7 @@ defm V_BFM_B32 : VOP2_32 <0x0000001e, "V_BFM_B32",
 defm V_MAC_F32 : VOP2_32 <0x0000001f, "V_MAC_F32", []>;
 defm V_MADMK_F32 : VOP2_32 <0x00000020, "V_MADMK_F32", []>;
 defm V_MADAK_F32 : VOP2_32 <0x00000021, "V_MADAK_F32", []>;
-//defm V_BCNT_U32_B32 : VOP2_32 <0x00000022, "V_BCNT_U32_B32", []>;
+defm V_BCNT_U32_B32 : VOP2_32 <0x00000022, "V_BCNT_U32_B32", []>;
 defm V_MBCNT_LO_U32_B32 : VOP2_32 <0x00000023, "V_MBCNT_LO_U32_B32", []>;
 defm V_MBCNT_HI_U32_B32 : VOP2_32 <0x00000024, "V_MBCNT_HI_U32_B32", []>;
 
@@ -2508,6 +2510,11 @@ def : Pat <
   (S_ADD_I32 $src0, $src1)
 >;
 
+def : Pat <
+  (i32 (add (i32 (ctpop i32:$popcnt)), i32:$val)),
+  (V_BCNT_U32_B32_e32 $popcnt, $val)
+>;
+
 //============================================================================//
 // Miscellaneous Optimization Patterns
 //============================================================================//

diff --git a/test/CodeGen/R600/ctpop.ll b/test/CodeGen/R600/ctpop.ll
new file mode 100644 (file)
index 0000000..168b83c
--- /dev/null
+++ b/test/CodeGen/R600/ctpop.ll
@@ -0,0 +1,203 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
+
+declare i32 @llvm.ctpop.i32(i32) nounwind readnone
+declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32>) nounwind readnone
+declare <4 x i32> @llvm.ctpop.v4i32(<4 x i32>) nounwind readnone
+declare <8 x i32> @llvm.ctpop.v8i32(<8 x i32>) nounwind readnone
+declare <16 x i32> @llvm.ctpop.v16i32(<16 x i32>) nounwind readnone
+
+; FUNC-LABEL: @s_ctpop_i32:
+; SI: S_LOAD_DWORD [[SVAL:s[0-9]+]],
+; SI: S_BCNT1_I32_B32 [[SRESULT:s[0-9]+]], [[SVAL]]
+; SI: V_MOV_B32_e32 [[VRESULT:v[0-9]+]], [[SRESULT]]
+; SI: BUFFER_STORE_DWORD [[VRESULT]],
+; SI: S_ENDPGM
+define void @s_ctpop_i32(i32 addrspace(1)* noalias %out, i32 %val) nounwind {
+  %ctpop = call i32 @llvm.ctpop.i32(i32 %val) nounwind readnone
+  store i32 %ctpop, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; XXX - Why 0 in register?
+; FUNC-LABEL: @v_ctpop_i32:
+; SI: BUFFER_LOAD_DWORD [[VAL:v[0-9]+]],
+; SI: V_MOV_B32_e32 [[VZERO:v[0-9]+]], 0
+; SI: V_BCNT_U32_B32_e32 [[RESULT:v[0-9]+]], [[VAL]], [[VZERO]]
+; SI: BUFFER_STORE_DWORD [[RESULT]],
+; SI: S_ENDPGM
+define void @v_ctpop_i32(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
+  %val = load i32 addrspace(1)* %in, align 4
+  %ctpop = call i32 @llvm.ctpop.i32(i32 %val) nounwind readnone
+  store i32 %ctpop, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @v_ctpop_add_chain_i32
+; SI: BUFFER_LOAD_DWORD [[VAL0:v[0-9]+]],
+; SI: BUFFER_LOAD_DWORD [[VAL1:v[0-9]+]],
+; SI: V_MOV_B32_e32 [[VZERO:v[0-9]+]], 0
+; SI: V_BCNT_U32_B32_e32 [[MIDRESULT:v[0-9]+]], [[VAL1]], [[VZERO]]
+; SI-NOT: ADD
+; SI: V_BCNT_U32_B32_e64 [[RESULT:v[0-9]+]], [[VAL0]], [[MIDRESULT]]
+; SI: BUFFER_STORE_DWORD [[RESULT]],
+; SI: S_ENDPGM
+define void @v_ctpop_add_chain_i32(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in0, i32 addrspace(1)* noalias %in1) nounwind {
+  %val0 = load i32 addrspace(1)* %in0, align 4
+  %val1 = load i32 addrspace(1)* %in1, align 4
+  %ctpop0 = call i32 @llvm.ctpop.i32(i32 %val0) nounwind readnone
+  %ctpop1 = call i32 @llvm.ctpop.i32(i32 %val1) nounwind readnone
+  %add = add i32 %ctpop0, %ctpop1
+  store i32 %add, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @v_ctpop_v2i32:
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: S_ENDPGM
+define void @v_ctpop_v2i32(<2 x i32> addrspace(1)* noalias %out, <2 x i32> addrspace(1)* noalias %in) nounwind {
+  %val = load <2 x i32> addrspace(1)* %in, align 8
+  %ctpop = call <2 x i32> @llvm.ctpop.v2i32(<2 x i32> %val) nounwind readnone
+  store <2 x i32> %ctpop, <2 x i32> addrspace(1)* %out, align 8
+  ret void
+}
+
+; FUNC-LABEL: @v_ctpop_v4i32:
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: S_ENDPGM
+define void @v_ctpop_v4i32(<4 x i32> addrspace(1)* noalias %out, <4 x i32> addrspace(1)* noalias %in) nounwind {
+  %val = load <4 x i32> addrspace(1)* %in, align 16
+  %ctpop = call <4 x i32> @llvm.ctpop.v4i32(<4 x i32> %val) nounwind readnone
+  store <4 x i32> %ctpop, <4 x i32> addrspace(1)* %out, align 16
+  ret void
+}
+
+; FUNC-LABEL: @v_ctpop_v8i32:
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: S_ENDPGM
+define void @v_ctpop_v8i32(<8 x i32> addrspace(1)* noalias %out, <8 x i32> addrspace(1)* noalias %in) nounwind {
+  %val = load <8 x i32> addrspace(1)* %in, align 32
+  %ctpop = call <8 x i32> @llvm.ctpop.v8i32(<8 x i32> %val) nounwind readnone
+  store <8 x i32> %ctpop, <8 x i32> addrspace(1)* %out, align 32
+  ret void
+}
+
+; FUNC-LABEL: @v_ctpop_v16i32:
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: V_BCNT_U32_B32_e32
+; SI: S_ENDPGM
+define void @v_ctpop_v16i32(<16 x i32> addrspace(1)* noalias %out, <16 x i32> addrspace(1)* noalias %in) nounwind {
+  %val = load <16 x i32> addrspace(1)* %in, align 32
+  %ctpop = call <16 x i32> @llvm.ctpop.v16i32(<16 x i32> %val) nounwind readnone
+  store <16 x i32> %ctpop, <16 x i32> addrspace(1)* %out, align 32
+  ret void
+}
+
+; FUNC-LABEL: @v_ctpop_i32_add_inline_constant:
+; SI: BUFFER_LOAD_DWORD [[VAL:v[0-9]+]],
+; SI: V_BCNT_U32_B32_e64 [[RESULT:v[0-9]+]], [[VAL]], 4
+; SI: BUFFER_STORE_DWORD [[RESULT]],
+; SI: S_ENDPGM
+define void @v_ctpop_i32_add_inline_constant(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
+  %val = load i32 addrspace(1)* %in, align 4
+  %ctpop = call i32 @llvm.ctpop.i32(i32 %val) nounwind readnone
+  %add = add i32 %ctpop, 4
+  store i32 %add, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @v_ctpop_i32_add_inline_constant_inv:
+; SI: BUFFER_LOAD_DWORD [[VAL:v[0-9]+]],
+; SI: V_BCNT_U32_B32_e64 [[RESULT:v[0-9]+]], [[VAL]], 4
+; SI: BUFFER_STORE_DWORD [[RESULT]],
+; SI: S_ENDPGM
+define void @v_ctpop_i32_add_inline_constant_inv(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
+  %val = load i32 addrspace(1)* %in, align 4
+  %ctpop = call i32 @llvm.ctpop.i32(i32 %val) nounwind readnone
+  %add = add i32 4, %ctpop
+  store i32 %add, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @v_ctpop_i32_add_literal:
+; SI: BUFFER_LOAD_DWORD [[VAL:v[0-9]+]],
+; SI: V_MOV_B32_e32 [[LIT:v[0-9]+]], 0x1869f
+; SI: V_BCNT_U32_B32_e32 [[RESULT:v[0-9]+]], [[VAL]], [[LIT]]
+; SI: BUFFER_STORE_DWORD [[RESULT]],
+; SI: S_ENDPGM
+define void @v_ctpop_i32_add_literal(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
+  %val = load i32 addrspace(1)* %in, align 4
+  %ctpop = call i32 @llvm.ctpop.i32(i32 %val) nounwind readnone
+  %add = add i32 %ctpop, 99999
+  store i32 %add, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @v_ctpop_i32_add_var:
+; SI-DAG: BUFFER_LOAD_DWORD [[VAL:v[0-9]+]],
+; SI-DAG: S_LOAD_DWORD [[VAR:s[0-9]+]],
+; SI: V_BCNT_U32_B32_e64 [[RESULT:v[0-9]+]], [[VAL]], [[VAR]]
+; SI: BUFFER_STORE_DWORD [[RESULT]],
+; SI: S_ENDPGM
+define void @v_ctpop_i32_add_var(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in, i32 %const) nounwind {
+  %val = load i32 addrspace(1)* %in, align 4
+  %ctpop = call i32 @llvm.ctpop.i32(i32 %val) nounwind readnone
+  %add = add i32 %ctpop, %const
+  store i32 %add, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @v_ctpop_i32_add_var_inv:
+; SI-DAG: BUFFER_LOAD_DWORD [[VAL:v[0-9]+]],
+; SI-DAG: S_LOAD_DWORD [[VAR:s[0-9]+]],
+; SI: V_BCNT_U32_B32_e64 [[RESULT:v[0-9]+]], [[VAL]], [[VAR]]
+; SI: BUFFER_STORE_DWORD [[RESULT]],
+; SI: S_ENDPGM
+define void @v_ctpop_i32_add_var_inv(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in, i32 %const) nounwind {
+  %val = load i32 addrspace(1)* %in, align 4
+  %ctpop = call i32 @llvm.ctpop.i32(i32 %val) nounwind readnone
+  %add = add i32 %const, %ctpop
+  store i32 %add, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @v_ctpop_i32_add_vvar_inv
+; SI-DAG: BUFFER_LOAD_DWORD [[VAL:v[0-9]+]], {{.*}} + 0x0
+; SI-DAG: BUFFER_LOAD_DWORD [[VAR:v[0-9]+]], {{.*}} + 0x10
+; SI: V_BCNT_U32_B32_e32 [[RESULT:v[0-9]+]], [[VAL]], [[VAR]]
+; SI: BUFFER_STORE_DWORD [[RESULT]],
+; SI: S_ENDPGM
+define void @v_ctpop_i32_add_vvar_inv(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in, i32 addrspace(1)* noalias %constptr) nounwind {
+  %val = load i32 addrspace(1)* %in, align 4
+  %ctpop = call i32 @llvm.ctpop.i32(i32 %val) nounwind readnone
+  %gep = getelementptr i32 addrspace(1)* %constptr, i32 4
+  %const = load i32 addrspace(1)* %gep, align 4
+  %add = add i32 %const, %ctpop
+  store i32 %add, i32 addrspace(1)* %out, align 4
+  ret void
+}