From 6b738b8719dc71499b557ce7c8c8066e7d8b81e7 Mon Sep 17 00:00:00 2001
From: Matthew Fluet <matthew.fluet@gmail.com>
Date: Thu, 8 Aug 2019 15:08:31 -0400
Subject: [PATCH] Add `-codegen-fuse-op-and-chk {false|true}` compile-time flag

It appears that GCC (and, to a lesser extent) Clang/LLVM do not always
successfully fuse adjacent `Word<N>_<op>` and
`Word{S,U}<N>_<op>CheckP` primitives.  The performance results
reported at https://github.com/MLton/mlton/pull/273 and
https://github.com/MLton/mlton/pull/292 suggest that this does not
always have significant impact, but a close look at the `md5`
benchmark shows that the native codegen significantly outperforms the
C codegen with gcc-9 due to redundant arithmetic computations (one for
`Word{S,U}<N>_<op>CheckP` and another for `Word<N>_<op>`).

This flag will be used to enable explicit fusing of adjacent
`Word<N>_<op>` and `Word{S,U}<N>_<op>CheckP` primitives in the
codegens.
---
 mlton/control/control-flags.sig | 3 +++
 mlton/control/control-flags.sml | 5 ++++-
 mlton/main/main.fun             | 2 ++
 3 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/mlton/control/control-flags.sig b/mlton/control/control-flags.sig
index 45ad6b429d..6f25b31343 100644
--- a/mlton/control/control-flags.sig
+++ b/mlton/control/control-flags.sig
@@ -77,6 +77,9 @@ signature CONTROL_FLAGS =
       (* whether or not to use comments in codegen *)
       val codegenComments: int ref
 
+      (* whether or not to fuse `op` and `opCheckP` primitives in codegen *)
+      val codegenFuseOpAndChk: bool ref
+
       val contifyIntoMain: bool ref
 
       (* Generate an executable with debugging info. *)
diff --git a/mlton/control/control-flags.sml b/mlton/control/control-flags.sml
index 85b635c49b..df383ce86e 100644
--- a/mlton/control/control-flags.sml
+++ b/mlton/control/control-flags.sml
@@ -190,7 +190,10 @@ val codegen = control {name = "codegen",
 val codegenComments = control {name = "codegen comments",
                                default = 0,
                                toString = Int.toString}
-   
+
+val codegenFuseOpAndChk = control {name = "fuse `op` and `opCheckP` primitives in codegen",
+                                   default = false,
+                                   toString = Bool.toString}
 
 val contifyIntoMain = control {name = "contifyIntoMain",
                                default = false,
diff --git a/mlton/main/main.fun b/mlton/main/main.fun
index f7b6e2ec62..b9bc10a16c 100644
--- a/mlton/main/main.fun
+++ b/mlton/main/main.fun
@@ -332,6 +332,8 @@ fun makeOptions {usage} =
                                      | NONE => usage (concat ["invalid -codegen flag: ", s]))))),
        (Expert, "codegen-comments", " <n>", "level of comments  (0)",
         intRef codegenComments),
+       (Expert, "codegen-fuse-op-and-chk", " {false|true}", "fuse `op` and `opCheckP` primitives in codegen",
+        boolRef codegenFuseOpAndChk),
        (Normal, "const", " '<name> <value>'", "set compile-time constant",
         SpaceString (fn s =>
                      case String.tokens (s, Char.isSpace) of