From 30f9dcd4f5e151601f2c75a0a4ae4419abf5aa37 Mon Sep 17 00:00:00 2001
From: Matt Pharr <matt.pharr@intel.com>
Date: Tue, 13 Sep 2011 15:36:34 -0700
Subject: [PATCH] Unroll loops by default, add --opt=disable-loop-unroll to
 disable.

Issue #78.
---
 examples/aobench_instrumented/Makefile |  2 +-
 ispc.cpp                               |  1 +
 ispc.h                                 |  4 ++++
 main.cpp                               | 31 +++++++++++++++++++-------
 opt.cpp                                |  6 +++--
 5 files changed, 33 insertions(+), 11 deletions(-)

diff --git a/examples/aobench_instrumented/Makefile b/examples/aobench_instrumented/Makefile
index cedff406..20a5a9b9 100644
--- a/examples/aobench_instrumented/Makefile
+++ b/examples/aobench_instrumented/Makefile
@@ -2,7 +2,7 @@
 CXX=g++ -m64
 CXXFLAGS=-Iobjs/ -g3 -Wall
 ISPC=ispc
-ISPCFLAGS=-O2 --fast-math --instrument --arch=x86-64
+ISPCFLAGS=-O2 --instrument --arch=x86-64
 
 default: ao
 
diff --git a/ispc.cpp b/ispc.cpp
index cf9e307f..8430f791 100644
--- a/ispc.cpp
+++ b/ispc.cpp
@@ -256,6 +256,7 @@ Opt::Opt() {
     level = 1;
     fastMath = false;
     fastMaskedVload = false;
+    unrollLoops = true;
     disableBlendedMaskedStores = false;
     disableCoherentControlFlow = false;
     disableUniformControlFlow = false;
diff --git a/ispc.h b/ispc.h
index 8536732d..4b9bec65 100644
--- a/ispc.h
+++ b/ispc.h
@@ -244,6 +244,10 @@ struct Opt {
         arrays, so is unsafe in general.) */
     bool fastMaskedVload;
 
+    /** Indicates when loops should be unrolled (when doing so seems like
+        it will make sense. */
+    bool unrollLoops;
+
     /** On targets that don't have a masked store instruction but do have a
         blending instruction, by default, we simulate masked stores by
         loading the old value, blending, and storing the result.  This can
diff --git a/main.cpp b/main.cpp
index 4b74ffe6..c969f267 100644
--- a/main.cpp
+++ b/main.cpp
@@ -73,8 +73,6 @@ static void usage(int ret) {
     printf("    [--emit-asm]\t\t\tGenerate assembly language file as output\n");
     printf("    [--emit-llvm]\t\t\tEmit LLVM bitode file as output\n");
     printf("    [--emit-obj]\t\t\tGenerate object file file as output (default)\n");
-    printf("    [--fast-math]\t\t\tPerform non-IEEE-compliant optimizations of numeric expressions\n");
-    printf("    [--fast-masked-vload]\t\tFaster masked vector loads on SSE (may go past end of array)\n");
     printf("    [-g]\t\t\t\tGenerate debugging information\n");
     printf("    [--help]\t\t\t\tPrint help\n");
     printf("    [-h <name>/--header-outfile=<name>]\tOutput filename for header\n");
@@ -88,8 +86,11 @@ static void usage(int ret) {
     printf("    [--nocpp]\t\t\t\tDon't run the C preprocessor\n");
     printf("    [-o <name>/--outfile=<name>]\tOutput filename (may be \"-\" for standard output)\n");
     printf("    [-O0/-O1]\t\t\t\tSet optimization level (-O1 is default)\n");
-#if 0
     printf("    [--opt=<option>]\t\t\tSet optimization option\n");
+    printf("        disable-loop-unroll\t\tDisable loop unrolling.\n");
+    printf("        fast-masked-vload\t\tFaster masked vector loads on SSE (may go past end of array)\n");
+    printf("        fast-math\t\t\tPerform non-IEEE-compliant optimizations of numeric expressions\n");
+#if 0
     printf("        disable-blended-masked-stores\t\tScalarize masked stores on SSE (vs. using vblendps)\n");
     printf("        disable-coherent-control-flow\t\tDisable coherent control flow optimizations\n");
     printf("        disable-uniform-control-flow\t\tDisable uniform control flow optimizations\n");
@@ -198,10 +199,15 @@ int main(int Argc, char *Argv[]) {
             arch = argv[i] + 7;
         else if (!strncmp(argv[i], "--cpu=", 6))
             cpu = argv[i] + 6;
-        else if (!strcmp(argv[i], "--fast-math"))
-            g->opt.fastMath = true;
-        else if (!strcmp(argv[i], "--fast-masked-vload"))
-            g->opt.fastMaskedVload = true;
+        else if (!strcmp(argv[i], "--fast-math")) {
+            fprintf(stderr, "--fast-math option has been renamed to --opt=fast-math!\n");
+            usage(1);
+        }
+        else if (!strcmp(argv[i], "--fast-masked-vload")) {
+            fprintf(stderr, "--fast-masked-vload option has been renamed to "
+                    "--opt=fast-masked-vload!\n");
+            usage(1);
+        }
         else if (!strcmp(argv[i], "--debug"))
             g->debugPrint = true;
         else if (!strcmp(argv[i], "--instrument"))
@@ -238,7 +244,16 @@ int main(int Argc, char *Argv[]) {
         }
         else if (!strncmp(argv[i], "--opt=", 6)) {
             const char *opt = argv[i] + 6;
-            if (!strcmp(opt, "disable-blended-masked-stores"))
+            if (!strcmp(opt, "fast-math"))
+                g->opt.fastMath = true;
+            else if (!strcmp(opt, "fast-masked-vload"))
+                g->opt.fastMaskedVload = true;
+            else if (!strcmp(opt, "disable-loop-unroll"))
+                g->opt.unrollLoops = false;
+
+            // These are only used for performance tests of specific
+            // optimizations
+            else if (!strcmp(opt, "disable-blended-masked-stores"))
                 g->opt.disableBlendedMaskedStores = true;
             else if (!strcmp(opt, "disable-coherent-control-flow"))
                 g->opt.disableCoherentControlFlow = true;
diff --git a/opt.cpp b/opt.cpp
index 8dcf6a59..6fee84ba 100644
--- a/opt.cpp
+++ b/opt.cpp
@@ -294,7 +294,7 @@ Optimize(llvm::Module *module, int optLevel) {
         llvm::createStandardModulePasses(&optPM, 3, 
                                          false /* opt size */,
                                          true /* unit at a time */, 
-                                         false /* unroll loops */,
+                                         g->opt.unrollLoops,
                                          true /* simplify lib calls */,
                                          false /* may have exceptions */,
                                          llvm::createFunctionInliningPass());
@@ -309,7 +309,7 @@ Optimize(llvm::Module *module, int optLevel) {
         llvm::createStandardModulePasses(&optPM, 3, 
                                          false /* opt size */,
                                          true /* unit at a time */, 
-                                         false /* unroll loops */,
+                                         g->opt.unrollLoops,
                                          true /* simplify lib calls */,
                                          false /* may have exceptions */,
                                          llvm::createFunctionInliningPass());
@@ -318,6 +318,8 @@ Optimize(llvm::Module *module, int optLevel) {
         llvm::PassManagerBuilder builder;
         builder.OptLevel = 3;
         builder.Inliner = llvm::createFunctionInliningPass();
+        if (g->opt.unrollLoops == false)
+            builder.DisableUnrollLoops = true;
         builder.populateFunctionPassManager(funcPM);
         builder.populateModulePassManager(optPM);
         optPM.add(CreateIsCompileTimeConstantPass(true));