GBE: fix popcount bugs.

We need to pass correct popcount source type to backend.

v2: rebase to master.

Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>
Reviewed-by: "Luo, Xionghu" <xionghu.luo@intel.com>

4 files changed, 18 insertions, 10 deletions

diff --git a/backend/src/backend/gen_insn_selection.cpp b/backend/src/backend/gen_insn_selection.cpp
index 0e33f5d6..628fa2f1 100644
--- a/backend/src/backend/gen_insn_selection.cpp
+++ b/backend/src/backend/gen_insn_selection.cpp
@@ -2044,10 +2044,12 @@ namespace gbe
   /*! Unary instruction patterns */
   DECL_PATTERN(UnaryInstruction)
   {
-    static ir::Type getType(const ir::Opcode opcode, const ir::Type insnType) {
+    static ir::Type getType(const ir::Opcode opcode, const ir::Type insnType, bool isSrc = false) {
+      if (opcode == ir::OP_CBIT)
+        return isSrc ? insnType : ir::TYPE_U32;
       if (insnType == ir::TYPE_S64 || insnType == ir::TYPE_U64 || insnType == ir::TYPE_S8 || insnType == ir::TYPE_U8)
         return insnType;
-      if (opcode == ir::OP_FBH || opcode == ir::OP_FBL || opcode == ir::OP_CBIT || opcode == ir::OP_LZD)
+      if (opcode == ir::OP_FBH || opcode == ir::OP_FBL || opcode == ir::OP_LZD)
         return ir::TYPE_U32;
       if (insnType == ir::TYPE_S16 || insnType == ir::TYPE_U16)
         return insnType;
@@ -2059,8 +2061,8 @@ namespace gbe
     INLINE bool emitOne(Selection::Opaque &sel, const ir::UnaryInstruction &insn, bool &markChildren) const {
       const ir::Opcode opcode = insn.getOpcode();
       const ir::Type insnType = insn.getType();
-      const GenRegister dst = sel.selReg(insn.getDst(0), getType(opcode, insnType));
-      const GenRegister src = sel.selReg(insn.getSrc(0), getType(opcode, insnType));
+      const GenRegister dst = sel.selReg(insn.getDst(0), getType(opcode, insnType, false));
+      const GenRegister src = sel.selReg(insn.getSrc(0), getType(opcode, insnType, true));
       sel.push();
         if (sel.isScalarReg(insn.getDst(0)) == true) {
           sel.curr.execWidth = 1;
diff --git a/backend/src/ir/instruction.cpp b/backend/src/ir/instruction.cpp
index 4d1ae058..795ff079 100644
--- a/backend/src/ir/instruction.cpp
+++ b/backend/src/ir/instruction.cpp
@@ -814,7 +814,8 @@ namespace ir {
       const RegisterFamily family = getFamily(this->type);
       if (UNLIKELY(checkSpecialRegForWrite(dst[0], fn, whyNot) == false))
         return false;
-      if (UNLIKELY(checkRegisterData(family, dst[0], fn, whyNot) == false))
+      if (opcode != OP_CBIT &&
+          UNLIKELY(checkRegisterData(family, dst[0], fn, whyNot) == false))
         return false;
       for (uint32_t srcID = 0; srcID < srcNum; ++srcID)
         if (UNLIKELY(checkRegisterData(family, src[srcID], fn, whyNot) == false))
diff --git a/backend/src/libocl/tmpl/ocl_integer.tmpl.cl b/backend/src/libocl/tmpl/ocl_integer.tmpl.cl
index 28b9c208..6da0bab2 100644
--- a/backend/src/libocl/tmpl/ocl_integer.tmpl.cl
+++ b/backend/src/libocl/tmpl/ocl_integer.tmpl.cl
@@ -19,7 +19,12 @@
 
 PURE CONST uint __gen_ocl_fbh(uint);
 PURE CONST uint __gen_ocl_fbl(uint);
-PURE CONST uint __gen_ocl_cbit(uint);
+PURE CONST OVERLOADABLE uint __gen_ocl_cbit(uint);
+PURE CONST OVERLOADABLE uint __gen_ocl_cbit(int);
+PURE CONST OVERLOADABLE uint __gen_ocl_cbit(ushort);
+PURE CONST OVERLOADABLE uint __gen_ocl_cbit(short);
+PURE CONST OVERLOADABLE uint __gen_ocl_cbit(uchar);
+PURE CONST OVERLOADABLE uint __gen_ocl_cbit(char);
 
 OVERLOADABLE char clz(char x) {
   if (x < 0)
diff --git a/backend/src/llvm/llvm_gen_backend.cpp b/backend/src/llvm/llvm_gen_backend.cpp
index ef8ce3bc..86030b9a 100644
--- a/backend/src/llvm/llvm_gen_backend.cpp
+++ b/backend/src/llvm/llvm_gen_backend.cpp
@@ -625,7 +625,7 @@ namespace gbe
 #undef DECL_VISIT_FN
 
     // Emit unary instructions from gen native function
-    void emitUnaryCallInst(CallInst &I, CallSite &CS, ir::Opcode opcode);
+    void emitUnaryCallInst(CallInst &I, CallSite &CS, ir::Opcode opcode, ir::Type = ir::TYPE_FLOAT);
     // Emit unary instructions from gen native function
     void emitAtomicInst(CallInst &I, CallSite &CS, ir::AtomicOps opcode);
 
@@ -3044,7 +3044,7 @@ error:
     };
   }
 
-  void GenWriter::emitUnaryCallInst(CallInst &I, CallSite &CS, ir::Opcode opcode) {
+  void GenWriter::emitUnaryCallInst(CallInst &I, CallSite &CS, ir::Opcode opcode, ir::Type type) {
     CallSite::arg_iterator AI = CS.arg_begin();
 #if GBE_DEBUG
     CallSite::arg_iterator AE = CS.arg_end();
@@ -3052,7 +3052,7 @@ error:
     GBE_ASSERT(AI != AE);
     const ir::Register src = this->getRegister(*AI);
     const ir::Register dst = this->getRegister(&I);
-    ctx.ALU1(opcode, ir::TYPE_FLOAT, dst, src);
+    ctx.ALU1(opcode, type, dst, src);
   }
 
   void GenWriter::emitAtomicInst(CallInst &I, CallSite &CS, ir::AtomicOps opcode) {
@@ -3396,7 +3396,7 @@ error:
           }
           case GEN_OCL_FBH: this->emitUnaryCallInst(I,CS,ir::OP_FBH); break;
           case GEN_OCL_FBL: this->emitUnaryCallInst(I,CS,ir::OP_FBL); break;
-          case GEN_OCL_CBIT: this->emitUnaryCallInst(I,CS,ir::OP_CBIT); break;
+          case GEN_OCL_CBIT: this->emitUnaryCallInst(I,CS,ir::OP_CBIT, getUnsignedType(ctx, (*AI)->getType())); break;
           case GEN_OCL_ABS:
           {
             const ir::Register src = this->getRegister(*AI);