[ROCM] Enable ROCM Backend #1.5: Address Remaining Comments from trit…

…on-lang#1312 (triton-lang#1434)

This PR address the remaing issues from triton-lang#1312. It does the following
*  LLVM String Join
* adds comment to GCNBuilder Class

---------

Co-authored-by: Rahul Batra <rahbatra@amd.com>

include/triton/Conversion/TritonGPUToLLVM/AsmFormat.h

@@ -4,6 +4,7 @@
 #include "mlir/IR/Value.h"
 #include "triton/Dialect/Triton/IR/Dialect.h"
 #include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/StringExtras.h"
 #include "llvm/ADT/StringRef.h"
 #include <memory>
 #include <string>
@@ -17,14 +18,7 @@ using llvm::StringRef;
 
 inline std::string strJoin(llvm::ArrayRef<std::string> strs,
                            llvm::StringRef delimiter) {
-  std::string osStr;
-  llvm::raw_string_ostream os(osStr);
-  for (size_t i = 0; !strs.empty() && i < strs.size() - 1; ++i)
-    os << strs[i] << delimiter;
-  if (!strs.empty())
-    os << strs.back();
-  os.flush();
-  return osStr;
+  return llvm::join(strs.begin(), strs.end(), delimiter);
 }
 
 } // namespace triton

include/triton/Conversion/TritonGPUToLLVM/GCNAsmFormat.h

@@ -19,6 +19,69 @@ class GCNInstr;
 class GCNInstrCommon;
 class GCNInstrExecution;
 
+// GCNBuilder helps to manage a GCN asm program consists of one or multiple
+// instructions.
+//
+// A helper for building an ASM program, the objective of GCNBuilder is to give
+// a thin encapsulation and make the ASM code for MLIR LLVM Dialect more clear.
+// Currently, several factors are introduced to reduce the need for mixing
+// string and C++ if-else code.
+//
+// Usage:
+// To create a multiplcation operation
+//
+//
+// GCNBuilder gcnBuilder;
+// unsigned bitwidth = elemTy.getIntOrFloatBitWidth();
+//
+// const std::string readConstraint = "v";
+// const std::string writeConstraint = "=v";
+// auto res = gcnBuilder.newOperand(writeConstraint);
+// auto lhs = gcnBuilder.newOperand(operands[0], readConstraint);
+// auto rhs = gcnBuilder.newOperand(operands[1], readConstraint);
+//
+// create inst
+// auto &mul_inst =
+// gcnBuilder.create<GCNInstr>("v_mul")->float_op_type(bitwidth);
+//
+// launch insts
+// mul_inst(res, lhs, rhs);
+//
+// return result
+// Value ret = gcnBuilder.launch(rewriter, loc, elemTy, false);
+// return ret;
+// To get the asm code:
+// builder.dump()
+//
+// To get all the mlir::Value used in the GCN code,
+//
+// builder.getAllMlirArgs() // get {pVal, iVal, jVal, kVal}
+//
+// To get the string containing all the constraints with "," separated,
+// builder.getConstraints() // get "=v,v,v"
+//
+// GCNBuilder can build a GCN asm with multiple instructions, sample code:
+//
+// GCNBuilder builder;
+// auto &rcp = gcnBuilder.create<GCNInstr>("v_rcp")->float_op_type(bitwidth);
+// auto &mul_inst =
+// gcnBuilder.create<GCNInstr>("v_mul")->float_op_type(bitwidth);
+//
+// rcp(...);
+// mul_inst(...);
+// This will get a GCN code with two instructions.
+//
+// Similar to a C function, a declared GCNInstr instance can be launched
+// multiple times with different operands, e.g.
+//
+//   auto &mul_inst =
+//   gcnBuilder.create<GCNInstr>("v_mul")->float_op_type(bitwidth); mul_inst(...
+//   some operands ...); mul_inst(... some different operands ...);
+//
+// Finally, we will get a GCN code with two mov instructions.
+//
+// There are several derived instruction type for typical instructions, for
+// example, the GCNIOInstr for ld and st instructions.
 struct GCNBuilder {
   struct Operand {
     std::string constraint;