Move the primary fast-isel top-level comments to FastISel.cpp, where

they'll be a little more visible. Also, update and reword them a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56877 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/FastISel.cpp b/lib/CodeGen/SelectionDAG/FastISel.cpp
index dcb30ad18821b5002c0adecc9db25286357a5527..9f70bc998c905421112524920459d58300b81864 100644 (file)
--- a/lib/CodeGen/SelectionDAG/FastISel.cpp
+++ b/lib/CodeGen/SelectionDAG/FastISel.cpp
@@ -9,6 +9,34 @@
 //
 // This file contains the implementation of the FastISel class.
 //
+// "Fast" instruction selection is designed to emit very poor code quickly.
+// Also, it is not designed to be able to do much lowering, so most illegal
+// types (e.g. i64 on 32-bit targets) and operations (e.g. calls) are not
+// supported. It is also not intended to be able to do much optimization,
+// except in a few cases where doing optimizations reduces overall compile
+// time (e.g. folding constants into immediate fields, because it's cheap
+// and it reduces the number of instructions later phases have to examine).
+//
+// "Fast" instruction selection is able to fail gracefully and transfer
+// control to the SelectionDAG selector for operations that it doesn't
+// support. In many cases, this allows us to avoid duplicating a lot of
+// the complicated lowering logic that SelectionDAG currently has.
+//
+// The intended use for "fast" instruction selection is "-O0" mode
+// compilation, where the quality of the generated code is irrelevant when
+// weighed against the speed at which the code can be generated. Also,
+// at -O0, the LLVM optimizers are not running, and this makes the
+// compile time of codegen a much higher portion of the overall compile
+// time. Despite its limitations, "fast" instruction selection is able to
+// handle enough code on its own to provide noticeable overall speedups
+// in -O0 compiles.
+//
+// Basic operations are supported in a target-independent way, by reading
+// the same instruction descriptions that the SelectionDAG selector reads,
+// and identifying simple arithmetic operations that can be directly selected
+// from simple operators. More complicated operations currently require
+// target-specific code.
+//
 //===----------------------------------------------------------------------===//
 
 #include "llvm/Function.h"

diff --git a/utils/TableGen/FastISelEmitter.cpp b/utils/TableGen/FastISelEmitter.cpp
index 0aa9dfc49d4a3204efc6eb9cb083188840b21b57..5e3876369eef4e6587184c781742bc1f07671b61 100644 (file)
--- a/utils/TableGen/FastISelEmitter.cpp
+++ b/utils/TableGen/FastISelEmitter.cpp
@@ -7,32 +7,14 @@
 //
 //===----------------------------------------------------------------------===//
 //
-// This tablegen backend emits a "fast" instruction selector.
+// This tablegen backend emits code for use by the "fast" instruction
+// selection algorithm. See the comments at the top of
+// lib/CodeGen/SelectionDAG/FastISel.cpp for background.
 //
-// This instruction selection method is designed to emit very poor code
-// quickly. Also, it is not designed to do much lowering, so most illegal
-// types (e.g. i64 on 32-bit targets) and operations (e.g. calls) are not
-// supported and cannot easily be added. Blocks containing operations
-// that are not supported need to be handled by a more capable selector,
-// such as the SelectionDAG selector.
+// This file scans through the target's tablegen instruction-info files
+// and extracts instructions with obvious-looking patterns, and it emits
+// code to look up these instructions by type and operator.
 //
-// The intended use for "fast" instruction selection is "-O0" mode
-// compilation, where the quality of the generated code is irrelevant when
-// weighed against the speed at which the code can be generated.
-//
-// If compile time is so important, you might wonder why we don't just
-// skip codegen all-together, emit LLVM bytecode files, and execute them
-// with an interpreter. The answer is that it would complicate linking and
-// debugging, and also because that isn't how a compiler is expected to
-// work in some circles.
-//
-// If you need better generated code or more lowering than what this
-// instruction selector provides, use the SelectionDAG (DAGISel) instruction
-// selector instead. If you're looking here because SelectionDAG isn't fast
-// enough, consider looking into improving the SelectionDAG infastructure
-// instead. At the time of this writing there remain several major
-// opportunities for improvement.
-// 
 //===----------------------------------------------------------------------===//
 
 #include "FastISelEmitter.h"