[Beignet] [PATCH V2] Backend: Refine gen ir ALU1 inst getType

[Xiuli Pan]

From: Pan Xiuli <xiuli.pan at intel.com>

ALU1 getType function used to override some inst type, refine it by set
the right type for these ALU1 inst.
This will fix MOV for half and double type.
V2: Keep the optimization for MOV with float type.

Signed-off-by: Pan Xiuli <xiuli.pan at intel.com>
---
 backend/src/backend/gen_insn_selection.cpp | 13 ++++---------
 backend/src/llvm/llvm_gen_backend.cpp      |  9 +++++----
 2 files changed, 9 insertions(+), 13 deletions(-)

diff --git a/backend/src/backend/gen_insn_selection.cpp b/backend/src/backend/gen_insn_selection.cpp
index 3b21fb5..67f3b5f 100644
--- a/backend/src/backend/gen_insn_selection.cpp
+++ b/backend/src/backend/gen_insn_selection.cpp
@@ -2786,17 +2786,12 @@ extern bool OCL_DEBUGINFO; // first defined by calling BVAR in program.cpp
     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_LZD)
-        return ir::TYPE_U32;
-      if (opcode == ir::OP_SIMD_ANY || opcode == ir::OP_SIMD_ALL)
-        return ir::TYPE_S32;
-      if (insnType == ir::TYPE_S16 || insnType == ir::TYPE_U16)
-        return insnType;
       if (insnType == ir::TYPE_BOOL)
         return ir::TYPE_U16;
-      return ir::TYPE_FLOAT;
+      else if (insnType == ir::TYPE_U32 || insnType == ir::TYPE_S32)
+        return ir::TYPE_FLOAT;
+      else
+        return insnType;
     }
 
     INLINE bool emitOne(Selection::Opaque &sel, const ir::UnaryInstruction &insn, bool &markChildren) const {
diff --git a/backend/src/llvm/llvm_gen_backend.cpp b/backend/src/llvm/llvm_gen_backend.cpp
index 5135950..7005d60 100644
--- a/backend/src/llvm/llvm_gen_backend.cpp
+++ b/backend/src/llvm/llvm_gen_backend.cpp
@@ -4195,13 +4195,14 @@ namespace gbe
               ir::Register tmp1 = ctx.reg(getFamily(tmpType));
               ir::Register tmp2 = ctx.reg(getFamily(tmpType));
               ctx.CVT(tmpType, srcType, tmp0, src);
-              ctx.ALU1(ir::OP_LZD, tmpType, tmp1, tmp0);
+              ctx.ALU1(ir::OP_LZD, ir::TYPE_U32, tmp1, tmp0);
               ctx.SUB(tmpType, tmp2, tmp1, immReg);
               ctx.CVT(dstType, tmpType, dst, tmp2);
             }
             else
             {
-              ctx.ALU1(ir::OP_LZD, dstType, dst, src);
+              GBE_ASSERT(srcType == ir::TYPE_U32);
+              ctx.ALU1(ir::OP_LZD, srcType, dst, src);
             }
           }
           break;
@@ -4258,8 +4259,8 @@ namespace gbe
 #endif /* GBE_DEBUG */
 
         switch (genIntrinsicID) {
-          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_FBH: this->emitUnaryCallInst(I,CS,ir::OP_FBH, ir::TYPE_U32); break;
+          case GEN_OCL_FBL: this->emitUnaryCallInst(I,CS,ir::OP_FBL, ir::TYPE_U32); break;
           case GEN_OCL_CBIT: this->emitUnaryCallInst(I,CS,ir::OP_CBIT, getUnsignedType(ctx, (*AI)->getType())); break;
           case GEN_OCL_ABS:
           {
-- 
2.7.4