<div dir="ltr">Craig pointed this out (the similarity to InstCombine) & also points out InstCombine had the same problem (which I hadn't spotted). So I've fixed both - keeping them as separate libraries, but removing the dependence on Scalar.h<br><br>Thanks all!<br><br>- Dave<br><br><div class="gmail_quote"><div dir="ltr">On Tue, Apr 24, 2018 at 1:34 AM Aboud, Amjad <<a href="mailto:amjad.aboud@intel.com">amjad.aboud@intel.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div lang="EN-US" link="blue" vlink="purple">
<div class="m_8887644195325437420WordSection1">
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d">Hi Dave,<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d">Sorry for the late response.<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d">AggressiveInstCombine is an advanced InstCombine pass that will contain several pattern optimizations, Trunc pattern optimization was the first to be added.<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d">I thought that it makes sense to build similar solution as InstCombine. In fact, I simply imitated InstCombine solution.<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d">If you and others think that we better move this pass inside ScalarOpt library, then I have no objection.<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d">Craig, Sanjay do you have any opinion on this point?<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d">Thanks,<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d">Amjad<u></u><u></u></span></p>
<p class="MsoNormal"><a name="m_8887644195325437420__MailEndCompose"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1f497d"><u></u> <u></u></span></a></p>
<div style="border:none;border-left:solid blue 1.5pt;padding:0in 0in 0in 4.0pt">
<div>
<div style="border:none;border-top:solid #e1e1e1 1.0pt;padding:3.0pt 0in 0in 0in">
<p class="MsoNormal"><a name="m_8887644195325437420______replyseparator"></a><b><span style="font-size:11.0pt;font-family:"Calibri",sans-serif">From:</span></b><span style="font-size:11.0pt;font-family:"Calibri",sans-serif"> David Blaikie [mailto:<a href="mailto:dblaikie@gmail.com" target="_blank">dblaikie@gmail.com</a>]
<br>
<b>Sent:</b> Tuesday, April 24, 2018 01:45<br>
<b>To:</b> Aboud, Amjad <<a href="mailto:amjad.aboud@intel.com" target="_blank">amjad.aboud@intel.com</a>>; Craig Topper <<a href="mailto:craig.topper@gmail.com" target="_blank">craig.topper@gmail.com</a>>; Davide Italiano <<a href="mailto:davide@freebsd.org" target="_blank">davide@freebsd.org</a>><br>
<b>Cc:</b> <a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a><br>
<b>Subject:</b> Re: [llvm] r323321 - [InstCombine] Introducing Aggressive Instruction Combine pass (-aggressive-instcombine).<u></u><u></u></span></p>
</div>
</div></div></div></div><div lang="EN-US" link="blue" vlink="purple"><div class="m_8887644195325437420WordSection1"><div style="border:none;border-left:solid blue 1.5pt;padding:0in 0in 0in 4.0pt">
<p class="MsoNormal"><u></u> <u></u></p>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt">Ping - any ideas on this extra library? Can we remove it/fold it into ScalarOpt?<u></u><u></u></p>
<div>
<div>
<p class="MsoNormal">On Mon, Apr 2, 2018 at 9:11 AM David Blaikie <<a href="mailto:dblaikie@gmail.com" target="_blank">dblaikie@gmail.com</a>> wrote:<u></u><u></u></p>
</div>
<blockquote style="border:none;border-left:solid #cccccc 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<div>
<p class="MsoNormal">Hi Amjad, Craig,<br>
<br>
Just curious - is there a reason this ended up as not only a separate pass, but as an entirely separate library? Most passes are grouped into broad libraries like IPO/Scalar/etc. & it seems this one could be in Scalar, where it's entry point/construction is
declared anyway?<br>
<br>
I ask because this caught/caused some layering violations by inclnuding the Scalar.h header which is logically part of the Transforms/Scalar (ScalarOpt) library, which the AggressiveInstCombine library doesn't depend on.<br>
<br>
Could we move this pass into the ScalarOpt library?<br>
<br>
- Dave<u></u><u></u></p>
</div>
<div>
<p class="MsoNormal" style="margin-bottom:12.0pt"><u></u> <u></u></p>
<div>
<div>
<p class="MsoNormal">On Wed, Jan 24, 2018 at 4:44 AM Amjad Aboud via llvm-commits <<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a>> wrote:<u></u><u></u></p>
</div>
<blockquote style="border:none;border-left:solid #cccccc 1.0pt;padding:0in 0in 0in 6.0pt;margin-left:4.8pt;margin-right:0in">
<p class="MsoNormal">Author: aaboud<br>
Date: Wed Jan 24 04:42:42 2018<br>
New Revision: 323321<br>
<br>
URL: <a href="http://llvm.org/viewvc/llvm-project?rev=323321&view=rev" target="_blank">
http://llvm.org/viewvc/llvm-project?rev=323321&view=rev</a><br>
Log:<br>
[InstCombine] Introducing Aggressive Instruction Combine pass (-aggressive-instcombine).<br>
<br>
Combine expression patterns to form expressions with fewer, simple instructions.<br>
This pass does not modify the CFG.<br>
<br>
For example, this pass reduce width of expressions post-dominated by TruncInst<br>
into smaller width when applicable.<br>
<br>
It differs from instcombine pass in that it contains pattern optimization that<br>
requires higher complexity than the O(1), thus, it should run fewer times than<br>
instcombine pass.<br>
<br>
Differential Revision: <a href="https://reviews.llvm.org/D38313" target="_blank">
https://reviews.llvm.org/D38313</a><br>
<br>
Added:<br>
llvm/trunk/include/llvm/Transforms/AggressiveInstCombine/ (with props)<br>
llvm/trunk/include/llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h<br>
llvm/trunk/lib/Transforms/AggressiveInstCombine/ (with props)<br>
llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp<br>
llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombineInternal.h<br>
llvm/trunk/lib/Transforms/AggressiveInstCombine/CMakeLists.txt<br>
llvm/trunk/lib/Transforms/AggressiveInstCombine/LLVMBuild.txt<br>
llvm/trunk/lib/Transforms/AggressiveInstCombine/TruncInstCombine.cpp<br>
llvm/trunk/test/Transforms/AggressiveInstCombine/ (with props)<br>
llvm/trunk/test/Transforms/AggressiveInstCombine/trunc_multi_uses.ll<br>
Modified:<br>
llvm/trunk/docs/Passes.rst<br>
llvm/trunk/include/llvm/InitializePasses.h<br>
llvm/trunk/include/llvm/Transforms/Scalar.h<br>
llvm/trunk/lib/LTO/LLVMBuild.txt<br>
llvm/trunk/lib/Passes/LLVMBuild.txt<br>
llvm/trunk/lib/Passes/PassBuilder.cpp<br>
llvm/trunk/lib/Passes/PassRegistry.def<br>
llvm/trunk/lib/Transforms/CMakeLists.txt<br>
llvm/trunk/lib/Transforms/IPO/LLVMBuild.txt<br>
llvm/trunk/lib/Transforms/IPO/PassManagerBuilder.cpp<br>
llvm/trunk/lib/Transforms/LLVMBuild.txt<br>
llvm/trunk/lib/Transforms/Scalar/LLVMBuild.txt<br>
llvm/trunk/test/Other/new-pm-defaults.ll<br>
llvm/trunk/test/Other/new-pm-lto-defaults.ll<br>
llvm/trunk/test/Other/new-pm-thinlto-defaults.ll<br>
llvm/trunk/tools/opt/CMakeLists.txt<br>
llvm/trunk/tools/opt/opt.cpp<br>
<br>
Modified: llvm/trunk/docs/Passes.rst<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/docs/Passes.rst?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/docs/Passes.rst?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/docs/Passes.rst (original)<br>
+++ llvm/trunk/docs/Passes.rst Wed Jan 24 04:42:42 2018<br>
@@ -641,6 +641,21 @@ not library calls are simplified is cont<br>
:ref:`-functionattrs <passes-functionattrs>` pass and LLVM's knowledge of<br>
library calls on different targets.<br>
<br>
+.. _passes-aggressive-instcombine:<br>
+<br>
+``-aggressive-instcombine``: Combine expression patterns<br>
+--------------------------------------------------------<br>
+<br>
+Combine expression patterns to form expressions with fewer, simple instructions.<br>
+This pass does not modify the CFG.<br>
+<br>
+For example, this pass reduce width of expressions post-dominated by TruncInst<br>
+into smaller width when applicable.<br>
+<br>
+It differs from instcombine pass in that it contains pattern optimization that<br>
+requires higher complexity than the O(1), thus, it should run fewer times than<br>
+instcombine pass.<br>
+<br>
``-internalize``: Internalize Global Symbols<br>
--------------------------------------------<br>
<br>
<br>
Modified: llvm/trunk/include/llvm/InitializePasses.h<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/InitializePasses.h?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/InitializePasses.h?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/include/llvm/InitializePasses.h (original)<br>
+++ llvm/trunk/include/llvm/InitializePasses.h Wed Jan 24 04:42:42 2018<br>
@@ -64,6 +64,7 @@ void initializeADCELegacyPassPass(PassRe<br>
void initializeAddDiscriminatorsLegacyPassPass(PassRegistry&);<br>
void initializeAddressSanitizerModulePass(PassRegistry&);<br>
void initializeAddressSanitizerPass(PassRegistry&);<br>
+void initializeAggressiveInstCombinerLegacyPassPass(PassRegistry&);<br>
void initializeAliasSetPrinterPass(PassRegistry&);<br>
void initializeAlignmentFromAssumptionsPass(PassRegistry&);<br>
void initializeAlwaysInlinerLegacyPassPass(PassRegistry&);<br>
<br>
Propchange: llvm/trunk/include/llvm/Transforms/AggressiveInstCombine/<br>
------------------------------------------------------------------------------<br>
bugtraq:number = true<br>
<br>
Added: llvm/trunk/include/llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h?rev=323321&view=auto" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h?rev=323321&view=auto</a><br>
==============================================================================<br>
--- llvm/trunk/include/llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h (added)<br>
+++ llvm/trunk/include/llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h Wed Jan 24 04:42:42 2018<br>
@@ -0,0 +1,34 @@<br>
+//===- AggressiveInstCombine.h - AggressiveInstCombine pass -----*- C++ -*-===//<br>
+//<br>
+// The LLVM Compiler Infrastructure<br>
+//<br>
+// This file is distributed under the University of Illinois Open Source<br>
+// License. See LICENSE.TXT for details.<br>
+//<br>
+//===----------------------------------------------------------------------===//<br>
+/// \file<br>
+///<br>
+/// This file provides the primary interface to the aggressive instcombine pass.<br>
+/// This pass is suitable for use in the new pass manager. For a pass that works<br>
+/// with the legacy pass manager, please look for<br>
+/// \c createAggressiveInstCombinerPass() in Scalar.h.<br>
+///<br>
+//===----------------------------------------------------------------------===//<br>
+<br>
+#ifndef LLVM_TRANSFORMS_AGGRESSIVE_INSTCOMBINE_INSTCOMBINE_H<br>
+#define LLVM_TRANSFORMS_AGGRESSIVE_INSTCOMBINE_INSTCOMBINE_H<br>
+<br>
+#include "llvm/IR/Function.h"<br>
+#include "llvm/IR/PassManager.h"<br>
+<br>
+namespace llvm {<br>
+<br>
+class AggressiveInstCombinePass<br>
+ : public PassInfoMixin<AggressiveInstCombinePass> {<br>
+<br>
+public:<br>
+ PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);<br>
+};<br>
+}<br>
+<br>
+#endif<br>
<br>
Modified: llvm/trunk/include/llvm/Transforms/Scalar.h<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Transforms/Scalar.h?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Transforms/Scalar.h?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/include/llvm/Transforms/Scalar.h (original)<br>
+++ llvm/trunk/include/llvm/Transforms/Scalar.h Wed Jan 24 04:42:42 2018<br>
@@ -142,6 +142,13 @@ FunctionPass *createInstructionCombining<br>
<br>
//===----------------------------------------------------------------------===//<br>
//<br>
+// AggressiveInstCombiner - Combine expression patterns to form expressions with<br>
+// fewer, simple instructions. This pass does not modify the CFG.<br>
+//<br>
+FunctionPass *createAggressiveInstCombinerPass();<br>
+<br>
+//===----------------------------------------------------------------------===//<br>
+//<br>
// LICM - This pass is a loop invariant code motion and memory promotion pass.<br>
//<br>
Pass *createLICMPass();<br>
<br>
Modified: llvm/trunk/lib/LTO/LLVMBuild.txt<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/LTO/LLVMBuild.txt?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/LTO/LLVMBuild.txt?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/LTO/LLVMBuild.txt (original)<br>
+++ llvm/trunk/lib/LTO/LLVMBuild.txt Wed Jan 24 04:42:42 2018<br>
@@ -20,6 +20,7 @@ type = Library<br>
name = LTO<br>
parent = Libraries<br>
required_libraries =<br>
+ AggressiveInstCombine<br>
Analysis<br>
BitReader<br>
BitWriter<br>
<br>
Modified: llvm/trunk/lib/Passes/LLVMBuild.txt<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Passes/LLVMBuild.txt?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Passes/LLVMBuild.txt?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Passes/LLVMBuild.txt (original)<br>
+++ llvm/trunk/lib/Passes/LLVMBuild.txt Wed Jan 24 04:42:42 2018<br>
@@ -19,4 +19,4 @@<br>
type = Library<br>
name = Passes<br>
parent = Libraries<br>
-required_libraries = Analysis CodeGen Core IPO InstCombine Scalar Support Target TransformUtils Vectorize Instrumentation<br>
+required_libraries = AggressiveInstCombine Analysis CodeGen Core IPO InstCombine Scalar Support Target TransformUtils Vectorize Instrumentation<br>
<br>
Modified: llvm/trunk/lib/Passes/PassBuilder.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Passes/PassBuilder.cpp?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Passes/PassBuilder.cpp?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Passes/PassBuilder.cpp (original)<br>
+++ llvm/trunk/lib/Passes/PassBuilder.cpp Wed Jan 24 04:42:42 2018<br>
@@ -59,6 +59,7 @@<br>
#include "llvm/Support/Debug.h"<br>
#include "llvm/Support/Regex.h"<br>
#include "llvm/Target/TargetMachine.h"<br>
+#include "llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h"<br>
#include "llvm/Transforms/GCOVProfiler.h"<br>
#include "llvm/Transforms/IPO/AlwaysInliner.h"<br>
#include "llvm/Transforms/IPO/ArgumentPromotion.h"<br>
@@ -363,6 +364,8 @@ PassBuilder::buildFunctionSimplification<br>
FPM.addPass(JumpThreadingPass());<br>
FPM.addPass(CorrelatedValuePropagationPass());<br>
FPM.addPass(SimplifyCFGPass());<br>
+ if (Level == O3)<br>
+ FPM.addPass(AggressiveInstCombinePass());<br>
FPM.addPass(InstCombinePass());<br>
<br>
if (!isOptimizingForSize(Level))<br>
@@ -1010,6 +1013,8 @@ ModulePassManager PassBuilder::buildLTOD<br>
// function pointers. When this happens, we often have to resolve varargs<br>
// calls, etc, so let instcombine do this.<br>
FunctionPassManager PeepholeFPM(DebugLogging);<br>
+ if (Level == O3)<br>
+ PeepholeFPM.addPass(AggressiveInstCombinePass());<br>
PeepholeFPM.addPass(InstCombinePass());<br>
invokePeepholeEPCallbacks(PeepholeFPM, Level);<br>
<br>
<br>
Modified: llvm/trunk/lib/Passes/PassRegistry.def<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Passes/PassRegistry.def?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Passes/PassRegistry.def?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Passes/PassRegistry.def (original)<br>
+++ llvm/trunk/lib/Passes/PassRegistry.def Wed Jan 24 04:42:42 2018<br>
@@ -139,6 +139,7 @@ FUNCTION_ALIAS_ANALYSIS("type-based-aa",<br>
FUNCTION_PASS("aa-eval", AAEvaluator())<br>
FUNCTION_PASS("adce", ADCEPass())<br>
FUNCTION_PASS("add-discriminators", AddDiscriminatorsPass())<br>
+FUNCTION_PASS("aggressive-instcombine", AggressiveInstCombinePass())<br>
FUNCTION_PASS("alignment-from-assumptions", AlignmentFromAssumptionsPass())<br>
FUNCTION_PASS("bdce", BDCEPass())<br>
FUNCTION_PASS("bounds-checking", BoundsCheckingPass())<br>
<br>
Propchange: llvm/trunk/lib/Transforms/AggressiveInstCombine/<br>
------------------------------------------------------------------------------<br>
bugtraq:number = true<br>
<br>
Added: llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp?rev=323321&view=auto" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp?rev=323321&view=auto</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp (added)<br>
+++ llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombine.cpp Wed Jan 24 04:42:42 2018<br>
@@ -0,0 +1,110 @@<br>
+//===- AggressiveInstCombine.cpp ------------------------------------------===//<br>
+//<br>
+// The LLVM Compiler Infrastructure<br>
+//<br>
+// This file is distributed under the University of Illinois Open Source<br>
+// License. See LICENSE.TXT for details.<br>
+//<br>
+//===----------------------------------------------------------------------===//<br>
+//<br>
+// This file implements the aggressive expression pattern combiner classes.<br>
+// Currently, it handles expression patterns for:<br>
+// * Truncate instruction<br>
+//<br>
+//===----------------------------------------------------------------------===//<br>
+<br>
+#include "llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h"<br>
+#include "AggressiveInstCombineInternal.h"<br>
+#include "llvm/Analysis/AliasAnalysis.h"<br>
+#include "llvm/Analysis/BasicAliasAnalysis.h"<br>
+#include "llvm/Analysis/GlobalsModRef.h"<br>
+#include "llvm/Analysis/TargetLibraryInfo.h"<br>
+#include "llvm/IR/DataLayout.h"<br>
+#include "llvm/Pass.h"<br>
+#include "llvm/Transforms/Scalar.h"<br>
+using namespace llvm;<br>
+<br>
+#define DEBUG_TYPE "aggressive-instcombine"<br>
+<br>
+namespace {<br>
+/// Contains expression pattern combiner logic.<br>
+/// This class provides both the logic to combine expression patterns and<br>
+/// combine them. It differs from InstCombiner class in that each pattern<br>
+/// combiner runs only once as opposed to InstCombine's multi-iteration,<br>
+/// which allows pattern combiner to have higher complexity than the O(1)<br>
+/// required by the instruction combiner.<br>
+class AggressiveInstCombinerLegacyPass : public FunctionPass {<br>
+public:<br>
+ static char ID; // Pass identification, replacement for typeid<br>
+<br>
+ AggressiveInstCombinerLegacyPass() : FunctionPass(ID) {<br>
+ initializeAggressiveInstCombinerLegacyPassPass(<br>
+ *PassRegistry::getPassRegistry());<br>
+ }<br>
+<br>
+ void getAnalysisUsage(AnalysisUsage &AU) const override;<br>
+<br>
+ /// Run all expression pattern optimizations on the given /p F function.<br>
+ ///<br>
+ /// \param F function to optimize.<br>
+ /// \returns true if the IR is changed.<br>
+ bool runOnFunction(Function &F) override;<br>
+};<br>
+} // namespace<br>
+<br>
+void AggressiveInstCombinerLegacyPass::getAnalysisUsage(<br>
+ AnalysisUsage &AU) const {<br>
+ AU.setPreservesCFG();<br>
+ AU.addRequired<TargetLibraryInfoWrapperPass>();<br>
+ AU.addPreserved<AAResultsWrapperPass>();<br>
+ AU.addPreserved<BasicAAWrapperPass>();<br>
+ AU.addPreserved<GlobalsAAWrapperPass>();<br>
+}<br>
+<br>
+bool AggressiveInstCombinerLegacyPass::runOnFunction(Function &F) {<br>
+ auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();<br>
+ auto &DL = F.getParent()->getDataLayout();<br>
+<br>
+ bool MadeIRChange = false;<br>
+<br>
+ // Handle TruncInst patterns<br>
+ TruncInstCombine TIC(TLI, DL);<br>
+ MadeIRChange |= TIC.run(F);<br>
+<br>
+ // TODO: add more patterns to handle...<br>
+<br>
+ return MadeIRChange;<br>
+}<br>
+<br>
+PreservedAnalyses AggressiveInstCombinePass::run(Function &F,<br>
+ FunctionAnalysisManager &AM) {<br>
+ auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);<br>
+ auto &DL = F.getParent()->getDataLayout();<br>
+ bool MadeIRChange = false;<br>
+<br>
+ // Handle TruncInst patterns<br>
+ TruncInstCombine TIC(TLI, DL);<br>
+ MadeIRChange |= TIC.run(F);<br>
+ if (!MadeIRChange)<br>
+ // No changes, all analyses are preserved.<br>
+ return PreservedAnalyses::all();<br>
+<br>
+ // Mark all the analyses that instcombine updates as preserved.<br>
+ PreservedAnalyses PA;<br>
+ PA.preserveSet<CFGAnalyses>();<br>
+ PA.preserve<AAManager>();<br>
+ PA.preserve<GlobalsAA>();<br>
+ return PA;<br>
+}<br>
+<br>
+char AggressiveInstCombinerLegacyPass::ID = 0;<br>
+INITIALIZE_PASS_BEGIN(AggressiveInstCombinerLegacyPass,<br>
+ "aggressive-instcombine",<br>
+ "Combine pattern based expressions", false, false)<br>
+INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)<br>
+INITIALIZE_PASS_END(AggressiveInstCombinerLegacyPass, "aggressive-instcombine",<br>
+ "Combine pattern based expressions", false, false)<br>
+<br>
+FunctionPass *llvm::createAggressiveInstCombinerPass() {<br>
+ return new AggressiveInstCombinerLegacyPass();<br>
+}<br>
<br>
Added: llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombineInternal.h<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombineInternal.h?rev=323321&view=auto" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombineInternal.h?rev=323321&view=auto</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombineInternal.h (added)<br>
+++ llvm/trunk/lib/Transforms/AggressiveInstCombine/AggressiveInstCombineInternal.h Wed Jan 24 04:42:42 2018<br>
@@ -0,0 +1,119 @@<br>
+//===- AggressiveInstCombineInternal.h --------------------------*- C++ -*-===//<br>
+//<br>
+// The LLVM Compiler Infrastructure<br>
+//<br>
+// This file is distributed under the University of Illinois Open Source<br>
+// License. See LICENSE.TXT for details.<br>
+//<br>
+//===----------------------------------------------------------------------===//<br>
+//<br>
+// This file implements the instruction pattern combiner classes.<br>
+// Currently, it handles pattern expressions for:<br>
+// * Truncate instruction<br>
+//<br>
+//===----------------------------------------------------------------------===//<br>
+<br>
+#include "llvm/ADT/MapVector.h"<br>
+#include "llvm/ADT/STLExtras.h"<br>
+#include "llvm/Analysis/AliasAnalysis.h"<br>
+#include "llvm/Analysis/BasicAliasAnalysis.h"<br>
+#include "llvm/Analysis/ConstantFolding.h"<br>
+#include "llvm/Analysis/GlobalsModRef.h"<br>
+#include "llvm/Analysis/TargetLibraryInfo.h"<br>
+#include "llvm/IR/DataLayout.h"<br>
+#include "llvm/Pass.h"<br>
+#include "llvm/Transforms/Scalar.h"<br>
+using namespace llvm;<br>
+<br>
+//===----------------------------------------------------------------------===//<br>
+// TruncInstCombine - looks for expression dags dominated by trunc instructions<br>
+// and for each eligible dag, it will create a reduced bit-width expression and<br>
+// replace the old expression with this new one and remove the old one.<br>
+// Eligible expression dag is such that:<br>
+// 1. Contains only supported instructions.<br>
+// 2. Supported leaves: ZExtInst, SExtInst, TruncInst and Constant value.<br>
+// 3. Can be evaluated into type with reduced legal bit-width (or Trunc type).<br>
+// 4. All instructions in the dag must not have users outside the dag.<br>
+// Only exception is for {ZExt, SExt}Inst with operand type equal to the<br>
+// new reduced type chosen in (3).<br>
+//<br>
+// The motivation for this optimization is that evaluating and expression using<br>
+// smaller bit-width is preferable, especially for vectorization where we can<br>
+// fit more values in one vectorized instruction. In addition, this optimization<br>
+// may decrease the number of cast instructions, but will not increase it.<br>
+//===----------------------------------------------------------------------===//<br>
+<br>
+namespace llvm {<br>
+ class DataLayout;<br>
+ class TargetLibraryInfo;<br>
+<br>
+class TruncInstCombine {<br>
+ TargetLibraryInfo &TLI;<br>
+ const DataLayout &DL;<br>
+<br>
+ /// List of all TruncInst instructions to be processed.<br>
+ SmallVector<TruncInst *, 4> Worklist;<br>
+<br>
+ /// Current processed TruncInst instruction.<br>
+ TruncInst *CurrentTruncInst;<br>
+<br>
+ /// Information per each instruction in the expression dag.<br>
+ struct Info {<br>
+ /// Number of LSBs that are needed to generate a valid expression.<br>
+ unsigned ValidBitWidth = 0;<br>
+ /// Minimum number of LSBs needed to generate the ValidBitWidth.<br>
+ unsigned MinBitWidth = 0;<br>
+ /// The reduced value generated to replace the old instruction.<br>
+ Value *NewValue = nullptr;<br>
+ };<br>
+ /// An ordered map representing expression dag post-dominated by current<br>
+ /// processed TruncInst. It maps each instruction in the dag to its Info<br>
+ /// structure. The map is ordered such that each instruction appears before<br>
+ /// all other instructions in the dag that uses it.<br>
+ MapVector<Instruction *, Info> InstInfoMap;<br>
+<br>
+public:<br>
+ TruncInstCombine(TargetLibraryInfo &TLI, const DataLayout &DL)<br>
+ : TLI(TLI), DL(DL), CurrentTruncInst(nullptr) {}<br>
+<br>
+ /// Perform TruncInst pattern optimization on given function.<br>
+ bool run(Function &F);<br>
+<br>
+private:<br>
+ /// Build expression dag dominated by the /p CurrentTruncInst and append it to<br>
+ /// the InstInfoMap container.<br>
+ ///<br>
+ /// \return true only if succeed to generate an eligible sub expression dag.<br>
+ bool buildTruncExpressionDag();<br>
+<br>
+ /// Calculate the minimal allowed bit-width of the chain ending with the<br>
+ /// currently visited truncate's operand.<br>
+ ///<br>
+ /// \return minimum number of bits to which the chain ending with the<br>
+ /// truncate's operand can be shrunk to.<br>
+ unsigned getMinBitWidth();<br>
+<br>
+ /// Build an expression dag dominated by the current processed TruncInst and<br>
+ /// Check if it is eligible to be reduced to a smaller type.<br>
+ ///<br>
+ /// \return the scalar version of the new type to be used for the reduced<br>
+ /// expression dag, or nullptr if the expression dag is not eligible<br>
+ /// to be reduced.<br>
+ Type *getBestTruncatedType();<br>
+<br>
+ /// Given a \p V value and a \p SclTy scalar type return the generated reduced<br>
+ /// value of \p V based on the type \p SclTy.<br>
+ ///<br>
+ /// \param V value to be reduced.<br>
+ /// \param SclTy scalar version of new type to reduce to.<br>
+ /// \return the new reduced value.<br>
+ Value *getReducedOperand(Value *V, Type *SclTy);<br>
+<br>
+ /// Create a new expression dag using the reduced /p SclTy type and replace<br>
+ /// the old expression dag with it. Also erase all instructions in the old<br>
+ /// dag, except those that are still needed outside the dag.<br>
+ ///<br>
+ /// \param SclTy scalar version of new type to reduce expression dag into.<br>
+ void ReduceExpressionDag(Type *SclTy);<br>
+};<br>
+} // end namespace llvm.<br>
<br>
Added: llvm/trunk/lib/Transforms/AggressiveInstCombine/CMakeLists.txt<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/AggressiveInstCombine/CMakeLists.txt?rev=323321&view=auto" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/AggressiveInstCombine/CMakeLists.txt?rev=323321&view=auto</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/AggressiveInstCombine/CMakeLists.txt (added)<br>
+++ llvm/trunk/lib/Transforms/AggressiveInstCombine/CMakeLists.txt Wed Jan 24 04:42:42 2018<br>
@@ -0,0 +1,11 @@<br>
+add_llvm_library(LLVMAggressiveInstCombine<br>
+ AggressiveInstCombine.cpp<br>
+ TruncInstCombine.cpp<br>
+<br>
+ ADDITIONAL_HEADER_DIRS<br>
+ ${LLVM_MAIN_INCLUDE_DIR}/llvm/Transforms<br>
+ ${LLVM_MAIN_INCLUDE_DIR}/llvm/Transforms/AggressiveInstCombine<br>
+<br>
+ DEPENDS<br>
+ intrinsics_gen<br>
+ )<br>
<br>
Added: llvm/trunk/lib/Transforms/AggressiveInstCombine/LLVMBuild.txt<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/AggressiveInstCombine/LLVMBuild.txt?rev=323321&view=auto" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/AggressiveInstCombine/LLVMBuild.txt?rev=323321&view=auto</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/AggressiveInstCombine/LLVMBuild.txt (added)<br>
+++ llvm/trunk/lib/Transforms/AggressiveInstCombine/LLVMBuild.txt Wed Jan 24 04:42:42 2018<br>
@@ -0,0 +1,22 @@<br>
+;===- ./lib/Transforms/AggressiveInstCombine/LLVMBuild.txt -----*- Conf -*--===;<br>
+;<br>
+; The LLVM Compiler Infrastructure<br>
+;<br>
+; This file is distributed under the University of Illinois Open Source<br>
+; License. See LICENSE.TXT for details.<br>
+;<br>
+;===------------------------------------------------------------------------===;<br>
+;<br>
+; This is an LLVMBuild description file for the components in this subdirectory.<br>
+;<br>
+; For more information on the LLVMBuild system, please see:<br>
+;<br>
+; <a href="http://llvm.org/docs/LLVMBuild.html" target="_blank">http://llvm.org/docs/LLVMBuild.html</a><br>
+;<br>
+;===------------------------------------------------------------------------===;<br>
+<br>
+[component_0]<br>
+type = Library<br>
+name = AggressiveInstCombine<br>
+parent = Transforms<br>
+required_libraries = Analysis Core Support TransformUtils<br>
<br>
Added: llvm/trunk/lib/Transforms/AggressiveInstCombine/TruncInstCombine.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/AggressiveInstCombine/TruncInstCombine.cpp?rev=323321&view=auto" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/AggressiveInstCombine/TruncInstCombine.cpp?rev=323321&view=auto</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/AggressiveInstCombine/TruncInstCombine.cpp (added)<br>
+++ llvm/trunk/lib/Transforms/AggressiveInstCombine/TruncInstCombine.cpp Wed Jan 24 04:42:42 2018<br>
@@ -0,0 +1,401 @@<br>
+//===- TruncInstCombine.cpp -----------------------------------------------===//<br>
+//<br>
+// The LLVM Compiler Infrastructure<br>
+//<br>
+// This file is distributed under the University of Illinois Open Source<br>
+// License. See LICENSE.TXT for details.<br>
+//<br>
+//===----------------------------------------------------------------------===//<br>
+//<br>
+// TruncInstCombine - looks for expression dags post-dominated by TruncInst and<br>
+// for each eligible dag, it will create a reduced bit-width expression, replace<br>
+// the old expression with this new one and remove the old expression.<br>
+// Eligible expression dag is such that:<br>
+// 1. Contains only supported instructions.<br>
+// 2. Supported leaves: ZExtInst, SExtInst, TruncInst and Constant value.<br>
+// 3. Can be evaluated into type with reduced legal bit-width.<br>
+// 4. All instructions in the dag must not have users outside the dag.<br>
+// The only exception is for {ZExt, SExt}Inst with operand type equal to<br>
+// the new reduced type evaluated in (3).<br>
+//<br>
+// The motivation for this optimization is that evaluating and expression using<br>
+// smaller bit-width is preferable, especially for vectorization where we can<br>
+// fit more values in one vectorized instruction. In addition, this optimization<br>
+// may decrease the number of cast instructions, but will not increase it.<br>
+//<br>
+//===----------------------------------------------------------------------===//<br>
+<br>
+#include "AggressiveInstCombineInternal.h"<br>
+#include "llvm/ADT/MapVector.h"<br>
+#include "llvm/ADT/STLExtras.h"<br>
+#include "llvm/Analysis/ConstantFolding.h"<br>
+#include "llvm/Analysis/TargetLibraryInfo.h"<br>
+#include "llvm/IR/DataLayout.h"<br>
+#include "llvm/IR/IRBuilder.h"<br>
+using namespace llvm;<br>
+<br>
+#define DEBUG_TYPE "aggressive-instcombine"<br>
+<br>
+/// Given an instruction and a container, it fills all the relevant operands of<br>
+/// that instruction, with respect to the Trunc expression dag optimizaton.<br>
+static void getRelevantOperands(Instruction *I, SmallVectorImpl<Value *> &Ops) {<br>
+ unsigned Opc = I->getOpcode();<br>
+ switch (Opc) {<br>
+ case Instruction::Trunc:<br>
+ case Instruction::ZExt:<br>
+ case Instruction::SExt:<br>
+ // These CastInst are considered leaves of the evaluated expression, thus,<br>
+ // their operands are not relevent.<br>
+ break;<br>
+ case Instruction::Add:<br>
+ case Instruction::Sub:<br>
+ case Instruction::Mul:<br>
+ case Instruction::And:<br>
+ case Instruction::Or:<br>
+ case Instruction::Xor:<br>
+ Ops.push_back(I->getOperand(0));<br>
+ Ops.push_back(I->getOperand(1));<br>
+ break;<br>
+ default:<br>
+ llvm_unreachable("Unreachable!");<br>
+ }<br>
+}<br>
+<br>
+bool TruncInstCombine::buildTruncExpressionDag() {<br>
+ SmallVector<Value *, 8> Worklist;<br>
+ SmallVector<Instruction *, 8> Stack;<br>
+ // Clear old expression dag.<br>
+ InstInfoMap.clear();<br>
+<br>
+ Worklist.push_back(CurrentTruncInst->getOperand(0));<br>
+<br>
+ while (!Worklist.empty()) {<br>
+ Value *Curr = Worklist.back();<br>
+<br>
+ if (isa<Constant>(Curr)) {<br>
+ Worklist.pop_back();<br>
+ continue;<br>
+ }<br>
+<br>
+ auto *I = dyn_cast<Instruction>(Curr);<br>
+ if (!I)<br>
+ return false;<br>
+<br>
+ if (!Stack.empty() && Stack.back() == I) {<br>
+ // Already handled all instruction operands, can remove it from both the<br>
+ // Worklist and the Stack, and add it to the instruction info map.<br>
+ Worklist.pop_back();<br>
+ Stack.pop_back();<br>
+ // Insert I to the Info map.<br>
+ InstInfoMap.insert(std::make_pair(I, Info()));<br>
+ continue;<br>
+ }<br>
+<br>
+ if (InstInfoMap.count(I)) {<br>
+ Worklist.pop_back();<br>
+ continue;<br>
+ }<br>
+<br>
+ // Add the instruction to the stack before start handling its operands.<br>
+ Stack.push_back(I);<br>
+<br>
+ unsigned Opc = I->getOpcode();<br>
+ switch (Opc) {<br>
+ case Instruction::Trunc:<br>
+ case Instruction::ZExt:<br>
+ case Instruction::SExt:<br>
+ // trunc(trunc(x)) -> trunc(x)<br>
+ // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest<br>
+ // trunc(ext(x)) -> trunc(x) if the source type is larger than the new<br>
+ // dest<br>
+ break;<br>
+ case Instruction::Add:<br>
+ case Instruction::Sub:<br>
+ case Instruction::Mul:<br>
+ case Instruction::And:<br>
+ case Instruction::Or:<br>
+ case Instruction::Xor: {<br>
+ SmallVector<Value *, 2> Operands;<br>
+ getRelevantOperands(I, Operands);<br>
+ for (Value *Operand : Operands)<br>
+ Worklist.push_back(Operand);<br>
+ break;<br>
+ }<br>
+ default:<br>
+ // TODO: Can handle more cases here:<br>
+ // 1. select, shufflevector, extractelement, insertelement<br>
+ // 2. udiv, urem<br>
+ // 3. shl, lshr, ashr<br>
+ // 4. phi node(and loop handling)<br>
+ // ...<br>
+ return false;<br>
+ }<br>
+ }<br>
+ return true;<br>
+}<br>
+<br>
+unsigned TruncInstCombine::getMinBitWidth() {<br>
+ SmallVector<Value *, 8> Worklist;<br>
+ SmallVector<Instruction *, 8> Stack;<br>
+<br>
+ Value *Src = CurrentTruncInst->getOperand(0);<br>
+ Type *DstTy = CurrentTruncInst->getType();<br>
+ unsigned TruncBitWidth = DstTy->getScalarSizeInBits();<br>
+ unsigned OrigBitWidth =<br>
+ CurrentTruncInst->getOperand(0)->getType()->getScalarSizeInBits();<br>
+<br>
+ if (isa<Constant>(Src))<br>
+ return TruncBitWidth;<br>
+<br>
+ Worklist.push_back(Src);<br>
+ InstInfoMap[cast<Instruction>(Src)].ValidBitWidth = TruncBitWidth;<br>
+<br>
+ while (!Worklist.empty()) {<br>
+ Value *Curr = Worklist.back();<br>
+<br>
+ if (isa<Constant>(Curr)) {<br>
+ Worklist.pop_back();<br>
+ continue;<br>
+ }<br>
+<br>
+ // Otherwise, it must be an instruction.<br>
+ auto *I = cast<Instruction>(Curr);<br>
+<br>
+ auto &Info = InstInfoMap[I];<br>
+<br>
+ SmallVector<Value *, 2> Operands;<br>
+ getRelevantOperands(I, Operands);<br>
+<br>
+ if (!Stack.empty() && Stack.back() == I) {<br>
+ // Already handled all instruction operands, can remove it from both, the<br>
+ // Worklist and the Stack, and update MinBitWidth.<br>
+ Worklist.pop_back();<br>
+ Stack.pop_back();<br>
+ for (auto *Operand : Operands)<br>
+ if (auto *IOp = dyn_cast<Instruction>(Operand))<br>
+ Info.MinBitWidth =<br>
+ std::max(Info.MinBitWidth, InstInfoMap[IOp].MinBitWidth);<br>
+ continue;<br>
+ }<br>
+<br>
+ // Add the instruction to the stack before start handling its operands.<br>
+ Stack.push_back(I);<br>
+ unsigned ValidBitWidth = Info.ValidBitWidth;<br>
+<br>
+ // Update minimum bit-width before handling its operands. This is required<br>
+ // when the instruction is part of a loop.<br>
+ Info.MinBitWidth = std::max(Info.MinBitWidth, Info.ValidBitWidth);<br>
+<br>
+ for (auto *Operand : Operands)<br>
+ if (auto *IOp = dyn_cast<Instruction>(Operand)) {<br>
+ // If we already calculated the minimum bit-width for this valid<br>
+ // bit-width, or for a smaller valid bit-width, then just keep the<br>
+ // answer we already calculated.<br>
+ unsigned IOpBitwidth = InstInfoMap.lookup(IOp).ValidBitWidth;<br>
+ if (IOpBitwidth >= ValidBitWidth)<br>
+ continue;<br>
+ InstInfoMap[IOp].ValidBitWidth = std::max(ValidBitWidth, IOpBitwidth);<br>
+ Worklist.push_back(IOp);<br>
+ }<br>
+ }<br>
+ unsigned MinBitWidth = InstInfoMap.lookup(cast<Instruction>(Src)).MinBitWidth;<br>
+ assert(MinBitWidth >= TruncBitWidth);<br>
+<br>
+ if (MinBitWidth > TruncBitWidth) {<br>
+ // In this case reducing expression with vector type might generate a new<br>
+ // vector type, which is not preferable as it might result in generating<br>
+ // sub-optimal code.<br>
+ if (DstTy->isVectorTy())<br>
+ return OrigBitWidth;<br>
+ // Use the smallest integer type in the range [MinBitWidth, OrigBitWidth).<br>
+ Type *Ty = DL.getSmallestLegalIntType(DstTy->getContext(), MinBitWidth);<br>
+ // Update minimum bit-width with the new destination type bit-width if<br>
+ // succeeded to find such, otherwise, with original bit-width.<br>
+ MinBitWidth = Ty ? Ty->getScalarSizeInBits() : OrigBitWidth;<br>
+ } else { // MinBitWidth == TruncBitWidth<br>
+ // In this case the expression can be evaluated with the trunc instruction<br>
+ // destination type, and trunc instruction can be omitted. However, we<br>
+ // should not perform the evaluation if the original type is a legal scalar<br>
+ // type and the target type is illegal.<br>
+ bool FromLegal = MinBitWidth == 1 || DL.isLegalInteger(OrigBitWidth);<br>
+ bool ToLegal = MinBitWidth == 1 || DL.isLegalInteger(MinBitWidth);<br>
+ if (!DstTy->isVectorTy() && FromLegal && !ToLegal)<br>
+ return OrigBitWidth;<br>
+ }<br>
+ return MinBitWidth;<br>
+}<br>
+<br>
+Type *TruncInstCombine::getBestTruncatedType() {<br>
+ if (!buildTruncExpressionDag())<br>
+ return nullptr;<br>
+<br>
+ // We don't want to duplicate instructions, which isn't profitable. Thus, we<br>
+ // can't shrink something that has multiple users, unless all users are<br>
+ // post-dominated by the trunc instruction, i.e., were visited during the<br>
+ // expression evaluation.<br>
+ unsigned DesiredBitWidth = 0;<br>
+ for (auto Itr : InstInfoMap) {<br>
+ Instruction *I = Itr.first;<br>
+ if (I->hasOneUse())<br>
+ continue;<br>
+ bool IsExtInst = (isa<ZExtInst>(I) || isa<SExtInst>(I));<br>
+ for (auto *U : I->users())<br>
+ if (auto *UI = dyn_cast<Instruction>(U))<br>
+ if (UI != CurrentTruncInst && !InstInfoMap.count(UI)) {<br>
+ if (!IsExtInst)<br>
+ return nullptr;<br>
+ // If this is an extension from the dest type, we can eliminate it,<br>
+ // even if it has multiple users. Thus, update the DesiredBitWidth and<br>
+ // validate all extension instructions agrees on same DesiredBitWidth.<br>
+ unsigned ExtInstBitWidth =<br>
+ I->getOperand(0)->getType()->getScalarSizeInBits();<br>
+ if (DesiredBitWidth && DesiredBitWidth != ExtInstBitWidth)<br>
+ return nullptr;<br>
+ DesiredBitWidth = ExtInstBitWidth;<br>
+ }<br>
+ }<br>
+<br>
+ unsigned OrigBitWidth =<br>
+ CurrentTruncInst->getOperand(0)->getType()->getScalarSizeInBits();<br>
+<br>
+ // Calculate minimum allowed bit-width allowed for shrinking the currently<br>
+ // visited truncate's operand.<br>
+ unsigned MinBitWidth = getMinBitWidth();<br>
+<br>
+ // Check that we can shrink to smaller bit-width than original one and that<br>
+ // it is similar to the DesiredBitWidth is such exists.<br>
+ if (MinBitWidth >= OrigBitWidth ||<br>
+ (DesiredBitWidth && DesiredBitWidth != MinBitWidth))<br>
+ return nullptr;<br>
+<br>
+ return IntegerType::get(CurrentTruncInst->getContext(), MinBitWidth);<br>
+}<br>
+<br>
+/// Given a reduced scalar type \p Ty and a \p V value, return a reduced type<br>
+/// for \p V, according to its type, if it vector type, return the vector<br>
+/// version of \p Ty, otherwise return \p Ty.<br>
+static Type *getReducedType(Value *V, Type *Ty) {<br>
+ assert(Ty && !Ty->isVectorTy() && "Expect Scalar Type");<br>
+ if (auto *VTy = dyn_cast<VectorType>(V->getType()))<br>
+ return VectorType::get(Ty, VTy->getNumElements());<br>
+ return Ty;<br>
+}<br>
+<br>
+Value *TruncInstCombine::getReducedOperand(Value *V, Type *SclTy) {<br>
+ Type *Ty = getReducedType(V, SclTy);<br>
+ if (auto *C = dyn_cast<Constant>(V)) {<br>
+ C = ConstantExpr::getIntegerCast(C, Ty, false);<br>
+ // If we got a constantexpr back, try to simplify it with DL info.<br>
+ if (Constant *FoldedC = ConstantFoldConstant(C, DL, &TLI))<br>
+ C = FoldedC;<br>
+ return C;<br>
+ }<br>
+<br>
+ auto *I = cast<Instruction>(V);<br>
+ Info Entry = InstInfoMap.lookup(I);<br>
+ assert(Entry.NewValue);<br>
+ return Entry.NewValue;<br>
+}<br>
+<br>
+void TruncInstCombine::ReduceExpressionDag(Type *SclTy) {<br>
+ for (auto &Itr : InstInfoMap) { // Forward<br>
+ Instruction *I = Itr.first;<br>
+<br>
+ assert(!InstInfoMap.lookup(I).NewValue && "Instruction has been evaluated");<br>
+<br>
+ IRBuilder<> Builder(I);<br>
+ Value *Res = nullptr;<br>
+ unsigned Opc = I->getOpcode();<br>
+ switch (Opc) {<br>
+ case Instruction::Trunc:<br>
+ case Instruction::ZExt:<br>
+ case Instruction::SExt: {<br>
+ Type *Ty = getReducedType(I, SclTy);<br>
+ // If the source type of the cast is the type we're trying for then we can<br>
+ // just return the source. There's no need to insert it because it is not<br>
+ // new.<br>
+ if (I->getOperand(0)->getType() == Ty) {<br>
+ InstInfoMap[I].NewValue = I->getOperand(0);<br>
+ continue;<br>
+ }<br>
+ // Otherwise, must be the same type of cast, so just reinsert a new one.<br>
+ // This also handles the case of zext(trunc(x)) -> zext(x).<br>
+ Res = Builder.CreateIntCast(I->getOperand(0), Ty,<br>
+ Opc == Instruction::SExt);<br>
+<br>
+ // Update Worklist entries with new value if needed.<br>
+ if (auto *NewCI = dyn_cast<TruncInst>(Res)) {<br>
+ auto Entry = find(Worklist, I);<br>
+ if (Entry != Worklist.end())<br>
+ *Entry = NewCI;<br>
+ }<br>
+ break;<br>
+ }<br>
+ case Instruction::Add:<br>
+ case Instruction::Sub:<br>
+ case Instruction::Mul:<br>
+ case Instruction::And:<br>
+ case Instruction::Or:<br>
+ case Instruction::Xor: {<br>
+ Value *LHS = getReducedOperand(I->getOperand(0), SclTy);<br>
+ Value *RHS = getReducedOperand(I->getOperand(1), SclTy);<br>
+ Res = Builder.CreateBinOp((Instruction::BinaryOps)Opc, LHS, RHS);<br>
+ break;<br>
+ }<br>
+ default:<br>
+ llvm_unreachable("Unhandled instruction");<br>
+ }<br>
+<br>
+ InstInfoMap[I].NewValue = Res;<br>
+ cast<Instruction>(Res)->takeName(I);<br>
+ }<br>
+<br>
+ Value *Res = getReducedOperand(CurrentTruncInst->getOperand(0), SclTy);<br>
+ Type *DstTy = CurrentTruncInst->getType();<br>
+ if (Res->getType() != DstTy) {<br>
+ IRBuilder<> Builder(CurrentTruncInst);<br>
+ Res = Builder.CreateIntCast(Res, DstTy, false);<br>
+ cast<Instruction>(Res)->takeName(CurrentTruncInst);<br>
+ }<br>
+ CurrentTruncInst->replaceAllUsesWith(Res);<br>
+<br>
+ // Erase old expression dag, which was replaced by the reduced expression dag.<br>
+ // We iterate backward, which means we visit the instruction before we visit<br>
+ // any of its operands, this way, when we get to the operand, we already<br>
+ // removed the instructions (from the expression dag) that uses it.<br>
+ CurrentTruncInst->eraseFromParent();<br>
+ for (auto I = InstInfoMap.rbegin(), E = InstInfoMap.rend(); I != E; ++I) {<br>
+ // We still need to check that the instruction has no users before we erase<br>
+ // it, because {SExt, ZExt}Inst Instruction might have other users that was<br>
+ // not reduced, in such case, we need to keep that instruction.<br>
+ if (!I->first->getNumUses())<br>
+ I->first->eraseFromParent();<br>
+ }<br>
+}<br>
+<br>
+bool TruncInstCombine::run(Function &F) {<br>
+ bool MadeIRChange = false;<br>
+<br>
+ // Collect all TruncInst in the function into the Worklist for evaluating.<br>
+ for (auto &BB : F)<br>
+ for (auto &I : BB)<br>
+ if (auto *CI = dyn_cast<TruncInst>(&I))<br>
+ Worklist.push_back(CI);<br>
+<br>
+ // Process all TruncInst in the Worklist, for each instruction:<br>
+ // 1. Check if it dominates an eligible expression dag to be reduced.<br>
+ // 2. Create a reduced expression dag and replace the old one with it.<br>
+ while (!Worklist.empty()) {<br>
+ CurrentTruncInst = Worklist.pop_back_val();<br>
+<br>
+ if (Type *NewDstSclTy = getBestTruncatedType()) {<br>
+ DEBUG(dbgs() << "ICE: TruncInstCombine reducing type of expression dag "<br>
+ "dominated by: "<br>
+ << CurrentTruncInst << '\n');<br>
+ ReduceExpressionDag(NewDstSclTy);<br>
+ MadeIRChange = true;<br>
+ }<br>
+ }<br>
+<br>
+ return MadeIRChange;<br>
+}<br>
<br>
Modified: llvm/trunk/lib/Transforms/CMakeLists.txt<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/CMakeLists.txt?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/CMakeLists.txt?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/CMakeLists.txt (original)<br>
+++ llvm/trunk/lib/Transforms/CMakeLists.txt Wed Jan 24 04:42:42 2018<br>
@@ -1,5 +1,6 @@<br>
add_subdirectory(Utils)<br>
add_subdirectory(Instrumentation)<br>
+add_subdirectory(AggressiveInstCombine)<br>
add_subdirectory(InstCombine)<br>
add_subdirectory(Scalar)<br>
add_subdirectory(IPO)<br>
<br>
Modified: llvm/trunk/lib/Transforms/IPO/LLVMBuild.txt<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/IPO/LLVMBuild.txt?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/IPO/LLVMBuild.txt?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/IPO/LLVMBuild.txt (original)<br>
+++ llvm/trunk/lib/Transforms/IPO/LLVMBuild.txt Wed Jan 24 04:42:42 2018<br>
@@ -20,4 +20,4 @@ type = Library<br>
name = IPO<br>
parent = Transforms<br>
library_name = ipo<br>
-required_libraries = Analysis BitReader BitWriter Core InstCombine IRReader Linker Object ProfileData Scalar Support TransformUtils Vectorize Instrumentation<br>
+required_libraries = AggressiveInstCombine Analysis BitReader BitWriter Core InstCombine IRReader Linker Object ProfileData Scalar Support TransformUtils Vectorize Instrumentation<br>
<br>
Modified: llvm/trunk/lib/Transforms/IPO/PassManagerBuilder.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/IPO/PassManagerBuilder.cpp?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/IPO/PassManagerBuilder.cpp?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/IPO/PassManagerBuilder.cpp (original)<br>
+++ llvm/trunk/lib/Transforms/IPO/PassManagerBuilder.cpp Wed Jan 24 04:42:42 2018<br>
@@ -318,6 +318,8 @@ void PassManagerBuilder::addFunctionSimp<br>
MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals<br>
MPM.add(createCFGSimplificationPass()); // Merge & remove BBs<br>
// Combine silly seq's<br>
+ if (OptLevel > 2)<br>
+ MPM.add(createAggressiveInstCombinerPass());<br>
addInstructionCombiningPass(MPM);<br>
if (SizeLevel == 0 && !DisableLibCallsShrinkWrap)<br>
MPM.add(createLibCallsShrinkWrapPass());<br>
@@ -765,6 +767,8 @@ void PassManagerBuilder::addLTOOptimizat<br>
// simplification opportunities, and both can propagate functions through<br>
// function pointers. When this happens, we often have to resolve varargs<br>
// calls, etc, so let instcombine do this.<br>
+ if (OptLevel > 2)<br>
+ PM.add(createAggressiveInstCombinerPass());<br>
addInstructionCombiningPass(PM);<br>
addExtensionsToPM(EP_Peephole, PM);<br>
<br>
<br>
Modified: llvm/trunk/lib/Transforms/LLVMBuild.txt<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/LLVMBuild.txt?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/LLVMBuild.txt?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/LLVMBuild.txt (original)<br>
+++ llvm/trunk/lib/Transforms/LLVMBuild.txt Wed Jan 24 04:42:42 2018<br>
@@ -16,7 +16,7 @@<br>
;===------------------------------------------------------------------------===;<u></u><u></u></p>
<p class="MsoNormal"><br>
[common]<br>
-subdirectories = Coroutines IPO InstCombine Instrumentation Scalar Utils Vectorize ObjCARC<br>
+subdirectories = AggressiveInstCombine Coroutines IPO InstCombine Instrumentation Scalar Utils Vectorize ObjCARC<br>
<br>
[component_0]<br>
type = Group<br>
<br>
Modified: llvm/trunk/lib/Transforms/Scalar/LLVMBuild.txt<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/LLVMBuild.txt?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/LLVMBuild.txt?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/Scalar/LLVMBuild.txt (original)<br>
+++ llvm/trunk/lib/Transforms/Scalar/LLVMBuild.txt Wed Jan 24 04:42:42 2018<br>
@@ -20,4 +20,4 @@ type = Library<br>
name = Scalar<br>
parent = Transforms<br>
library_name = ScalarOpts<br>
-required_libraries = Analysis Core InstCombine Support TransformUtils<br>
+required_libraries = AggressiveInstCombine Analysis Core InstCombine Support TransformUtils<br>
<br>
Modified: llvm/trunk/test/Other/new-pm-defaults.ll<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Other/new-pm-defaults.ll?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Other/new-pm-defaults.ll?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/test/Other/new-pm-defaults.ll (original)<br>
+++ llvm/trunk/test/Other/new-pm-defaults.ll Wed Jan 24 04:42:42 2018<br>
@@ -126,6 +126,7 @@<br>
; CHECK-O-NEXT: Running analysis: LazyValueAnalysis<br>
; CHECK-O-NEXT: Running pass: CorrelatedValuePropagationPass<br>
; CHECK-O-NEXT: Running pass: SimplifyCFGPass<br>
+; CHECK-O3-NEXT: AggressiveInstCombinePass<br>
; CHECK-O-NEXT: Running pass: InstCombinePass<br>
; CHECK-O1-NEXT: Running pass: LibCallsShrinkWrapPass<br>
; CHECK-O2-NEXT: Running pass: LibCallsShrinkWrapPass<br>
<br>
Modified: llvm/trunk/test/Other/new-pm-lto-defaults.ll<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Other/new-pm-lto-defaults.ll?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Other/new-pm-lto-defaults.ll?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/test/Other/new-pm-lto-defaults.ll (original)<br>
+++ llvm/trunk/test/Other/new-pm-lto-defaults.ll Wed Jan 24 04:42:42 2018<br>
@@ -10,7 +10,8 @@<br>
; RUN: | FileCheck %s --check-prefix=CHECK-O --check-prefix=CHECK-O2<br>
; RUN: opt -disable-verify -debug-pass-manager \<br>
; RUN: -passes='lto<O3>' -S %s 2>&1 \<br>
-; RUN: | FileCheck %s --check-prefix=CHECK-O --check-prefix=CHECK-O2<br>
+; RUN: | FileCheck %s --check-prefix=CHECK-O --check-prefix=CHECK-O2 \<br>
+; RUN: --check-prefix=CHECK-O3<br>
; RUN: opt -disable-verify -debug-pass-manager \<br>
; RUN: -passes='lto<Os>' -S %s 2>&1 \<br>
; RUN: | FileCheck %s --check-prefix=CHECK-O --check-prefix=CHECK-O2<br>
@@ -20,7 +21,7 @@<br>
; RUN: opt -disable-verify -debug-pass-manager \<br>
; RUN: -passes='lto<O3>' -S %s -passes-ep-peephole='no-op-function' 2>&1 \<br>
; RUN: | FileCheck %s --check-prefix=CHECK-O --check-prefix=CHECK-O2 \<br>
-; RUN: --check-prefix=CHECK-EP-Peephole<br>
+; RUN: --check-prefix=CHECK-O3 --check-prefix=CHECK-EP-Peephole<br>
<br>
; CHECK-O: Starting llvm::Module pass manager run.<br>
; CHECK-O-NEXT: Running pass: PassManager<{{.*}}Module<br>
@@ -60,6 +61,7 @@<br>
; CHECK-O2-NEXT: Running pass: DeadArgumentEliminationPass<br>
; CHECK-O2-NEXT: Running pass: ModuleToFunctionPassAdaptor<{{.*}}PassManager{{.*}}><br>
; CHECK-O2-NEXT: Starting llvm::Function pass manager run.<br>
+; CHECK-O3-NEXT: Running pass: AggressiveInstCombinePass<br>
; CHECK-O2-NEXT: Running pass: InstCombinePass<br>
; CHECK-EP-Peephole-NEXT: Running pass: NoOpFunctionPass<br>
; CHECK-O2-NEXT: Finished llvm::Function pass manager run.<br>
<br>
Modified: llvm/trunk/test/Other/new-pm-thinlto-defaults.ll<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Other/new-pm-thinlto-defaults.ll?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Other/new-pm-thinlto-defaults.ll?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/test/Other/new-pm-thinlto-defaults.ll (original)<br>
+++ llvm/trunk/test/Other/new-pm-thinlto-defaults.ll Wed Jan 24 04:42:42 2018<br>
@@ -111,6 +111,7 @@<br>
; CHECK-O-NEXT: Running analysis: LazyValueAnalysis<br>
; CHECK-O-NEXT: Running pass: CorrelatedValuePropagationPass<br>
; CHECK-O-NEXT: Running pass: SimplifyCFGPass<br>
+; CHECK-O3-NEXT: Running pass: AggressiveInstCombinePass<br>
; CHECK-O-NEXT: Running pass: InstCombinePass<br>
; CHECK-O1-NEXT: Running pass: LibCallsShrinkWrapPass<br>
; CHECK-O2-NEXT: Running pass: LibCallsShrinkWrapPass<br>
<br>
Propchange: llvm/trunk/test/Transforms/AggressiveInstCombine/<br>
------------------------------------------------------------------------------<br>
bugtraq:number = true<br>
<br>
Added: llvm/trunk/test/Transforms/AggressiveInstCombine/trunc_multi_uses.ll<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/AggressiveInstCombine/trunc_multi_uses.ll?rev=323321&view=auto" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/AggressiveInstCombine/trunc_multi_uses.ll?rev=323321&view=auto</a><br>
==============================================================================<br>
--- llvm/trunk/test/Transforms/AggressiveInstCombine/trunc_multi_uses.ll (added)<br>
+++ llvm/trunk/test/Transforms/AggressiveInstCombine/trunc_multi_uses.ll Wed Jan 24 04:42:42 2018<br>
@@ -0,0 +1,214 @@<br>
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py<br>
+; RUN: opt < %s -aggressive-instcombine -S | FileCheck %s<br>
+; RUN: opt < %s -passes=aggressive-instcombine -S | FileCheck %s<br>
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"<br>
+<br>
+; Aggressive Instcombine should be able to reduce width of these expressions.<br>
+<br>
+declare i32 @use32(i32)<br>
+declare i32 @use64(i64)<br>
+declare <2 x i32> @use32_vec(<2 x i32>)<br>
+declare <2 x i32> @use64_vec(<2 x i64>)<br>
+<br>
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;<br>
+;; These tests check cases where expression dag post-dominated by TruncInst<br>
+;; contains instruction, which has more than one usage.<br>
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;<br>
+<br>
+define void @multi_uses_add(i32 %X) {<br>
+; CHECK-LABEL: @multi_uses_add(<br>
+; CHECK-NEXT: [[A1:%.*]] = zext i32 [[X:%.*]] to i64<br>
+; CHECK-NEXT: [[B1:%.*]] = add i32 [[X]], 15<br>
+; CHECK-NEXT: [[C1:%.*]] = mul i32 [[B1]], [[B1]]<br>
+; CHECK-NEXT: [[TMP1:%.*]] = call i32 @use32(i32 [[C1]])<br>
+; CHECK-NEXT: [[TMP2:%.*]] = call i32 @use64(i64 [[A1]])<br>
+; CHECK-NEXT: ret void<br>
+;<br>
+ %A1 = zext i32 %X to i64<br>
+ %B1 = add i64 %A1, 15<br>
+ %C1 = mul i64 %B1, %B1<br>
+ %T1 = trunc i64 %C1 to i32<br>
+ call i32 @use32(i32 %T1)<br>
+ ; make sure zext have another use that is not post-dominated by the TruncInst.<br>
+ call i32 @use64(i64 %A1)<br>
+ ret void<br>
+}<br>
+<br>
+define void @multi_uses_or(i32 %X) {<br>
+; CHECK-LABEL: @multi_uses_or(<br>
+; CHECK-NEXT: [[A1:%.*]] = zext i32 [[X:%.*]] to i64<br>
+; CHECK-NEXT: [[B1:%.*]] = or i32 [[X]], 15<br>
+; CHECK-NEXT: [[C1:%.*]] = mul i32 [[B1]], [[B1]]<br>
+; CHECK-NEXT: [[TMP1:%.*]] = call i32 @use32(i32 [[C1]])<br>
+; CHECK-NEXT: [[TMP2:%.*]] = call i32 @use64(i64 [[A1]])<br>
+; CHECK-NEXT: ret void<br>
+;<br>
+ %A1 = zext i32 %X to i64<br>
+ %B1 = or i64 %A1, 15<br>
+ %C1 = mul i64 %B1, %B1<br>
+ %T1 = trunc i64 %C1 to i32<br>
+ call i32 @use32(i32 %T1)<br>
+ ; make sure zext have another use that is not post-dominated by the TruncInst.<br>
+ call i32 @use64(i64 %A1)<br>
+ ret void<br>
+}<br>
+<br>
+define void @multi_uses_xor(i32 %X) {<br>
+; CHECK-LABEL: @multi_uses_xor(<br>
+; CHECK-NEXT: [[A1:%.*]] = zext i32 [[X:%.*]] to i64<br>
+; CHECK-NEXT: [[B1:%.*]] = xor i32 [[X]], 15<br>
+; CHECK-NEXT: [[C1:%.*]] = mul i32 [[B1]], [[B1]]<br>
+; CHECK-NEXT: [[TMP1:%.*]] = call i32 @use32(i32 [[C1]])<br>
+; CHECK-NEXT: [[TMP2:%.*]] = call i32 @use64(i64 [[A1]])<br>
+; CHECK-NEXT: ret void<br>
+;<br>
+ %A1 = zext i32 %X to i64<br>
+ %B1 = xor i64 %A1, 15<br>
+ %C1 = mul i64 %B1, %B1<br>
+ %T1 = trunc i64 %C1 to i32<br>
+ call i32 @use32(i32 %T1)<br>
+ ; make sure zext have another use that is not post-dominated by the TruncInst.<br>
+ call i32 @use64(i64 %A1)<br>
+ ret void<br>
+}<br>
+<br>
+define void @multi_uses_and(i32 %X) {<br>
+; CHECK-LABEL: @multi_uses_and(<br>
+; CHECK-NEXT: [[A1:%.*]] = zext i32 [[X:%.*]] to i64<br>
+; CHECK-NEXT: [[B1:%.*]] = and i32 [[X]], 15<br>
+; CHECK-NEXT: [[C1:%.*]] = mul i32 [[B1]], [[B1]]<br>
+; CHECK-NEXT: [[TMP1:%.*]] = call i32 @use32(i32 [[C1]])<br>
+; CHECK-NEXT: [[TMP2:%.*]] = call i32 @use64(i64 [[A1]])<br>
+; CHECK-NEXT: ret void<br>
+;<br>
+ %A1 = zext i32 %X to i64<br>
+ %B1 = and i64 %A1, 15<br>
+ %C1 = mul i64 %B1, %B1<br>
+ %T1 = trunc i64 %C1 to i32<br>
+ call i32 @use32(i32 %T1)<br>
+ ; make sure zext have another use that is not post-dominated by the TruncInst.<br>
+ call i32 @use64(i64 %A1)<br>
+ ret void<br>
+}<br>
+<br>
+define void @multi_uses_sub(i32 %X, i32 %Y) {<br>
+; CHECK-LABEL: @multi_uses_sub(<br>
+; CHECK-NEXT: [[A1:%.*]] = zext i32 [[X:%.*]] to i64<br>
+; CHECK-NEXT: [[A2:%.*]] = zext i32 [[Y:%.*]] to i64<br>
+; CHECK-NEXT: [[B1:%.*]] = sub i32 [[X]], [[Y]]<br>
+; CHECK-NEXT: [[C1:%.*]] = mul i32 [[B1]], [[B1]]<br>
+; CHECK-NEXT: [[TMP1:%.*]] = call i32 @use32(i32 [[C1]])<br>
+; CHECK-NEXT: [[TMP2:%.*]] = call i32 @use64(i64 [[A1]])<br>
+; CHECK-NEXT: [[TMP3:%.*]] = call i32 @use64(i64 [[A2]])<br>
+; CHECK-NEXT: ret void<br>
+;<br>
+ %A1 = zext i32 %X to i64<br>
+ %A2 = zext i32 %Y to i64<br>
+ %B1 = sub i64 %A1, %A2<br>
+ %C1 = mul i64 %B1, %B1<br>
+ %T1 = trunc i64 %C1 to i32<br>
+ call i32 @use32(i32 %T1)<br>
+ ; make sure zext have another use that is not post-dominated by the TruncInst.<br>
+ call i32 @use64(i64 %A1)<br>
+ call i32 @use64(i64 %A2)<br>
+ ret void<br>
+}<br>
+<br>
+define void @multi_use_vec_add(<2 x i32> %X) {<br>
+; CHECK-LABEL: @multi_use_vec_add(<br>
+; CHECK-NEXT: [[A1:%.*]] = zext <2 x i32> [[X:%.*]] to <2 x i64><br>
+; CHECK-NEXT: [[B1:%.*]] = add <2 x i32> [[X]], <i32 15, i32 15><br>
+; CHECK-NEXT: [[C1:%.*]] = mul <2 x i32> [[B1]], [[B1]]<br>
+; CHECK-NEXT: [[TMP1:%.*]] = call <2 x i32> @use32_vec(<2 x i32> [[C1]])<br>
+; CHECK-NEXT: [[TMP2:%.*]] = call <2 x i32> @use64_vec(<2 x i64> [[A1]])<br>
+; CHECK-NEXT: ret void<br>
+;<br>
+ %A1 = zext <2 x i32> %X to <2 x i64><br>
+ %B1 = add <2 x i64> %A1, <i64 15, i64 15><br>
+ %C1 = mul <2 x i64> %B1, %B1<br>
+ %T1 = trunc <2 x i64> %C1 to <2 x i32><br>
+ call <2 x i32> @use32_vec(<2 x i32> %T1)<br>
+ ; make sure zext have another use that is not post-dominated by the TruncInst.<br>
+ call <2 x i32> @use64_vec(<2 x i64> %A1)<br>
+ ret void<br>
+}<br>
+<br>
+define void @multi_use_vec_or(<2 x i32> %X) {<br>
+; CHECK-LABEL: @multi_use_vec_or(<br>
+; CHECK-NEXT: [[A1:%.*]] = zext <2 x i32> [[X:%.*]] to <2 x i64><br>
+; CHECK-NEXT: [[B1:%.*]] = or <2 x i32> [[X]], <i32 15, i32 15><br>
+; CHECK-NEXT: [[C1:%.*]] = mul <2 x i32> [[B1]], [[B1]]<br>
+; CHECK-NEXT: [[TMP1:%.*]] = call <2 x i32> @use32_vec(<2 x i32> [[C1]])<br>
+; CHECK-NEXT: [[TMP2:%.*]] = call <2 x i32> @use64_vec(<2 x i64> [[A1]])<br>
+; CHECK-NEXT: ret void<br>
+;<br>
+ %A1 = zext <2 x i32> %X to <2 x i64><br>
+ %B1 = or <2 x i64> %A1, <i64 15, i64 15><br>
+ %C1 = mul <2 x i64> %B1, %B1<br>
+ %T1 = trunc <2 x i64> %C1 to <2 x i32><br>
+ call <2 x i32> @use32_vec(<2 x i32> %T1)<br>
+ ; make sure zext have another use that is not post-dominated by the TruncInst.<br>
+ call <2 x i32> @use64_vec(<2 x i64> %A1)<br>
+ ret void<br>
+}<br>
+<br>
+define void @multi_use_vec_xor(<2 x i32> %X) {<br>
+; CHECK-LABEL: @multi_use_vec_xor(<br>
+; CHECK-NEXT: [[A1:%.*]] = zext <2 x i32> [[X:%.*]] to <2 x i64><br>
+; CHECK-NEXT: [[B1:%.*]] = xor <2 x i32> [[X]], <i32 15, i32 15><br>
+; CHECK-NEXT: [[C1:%.*]] = mul <2 x i32> [[B1]], [[B1]]<br>
+; CHECK-NEXT: [[TMP1:%.*]] = call <2 x i32> @use32_vec(<2 x i32> [[C1]])<br>
+; CHECK-NEXT: [[TMP2:%.*]] = call <2 x i32> @use64_vec(<2 x i64> [[A1]])<br>
+; CHECK-NEXT: ret void<br>
+;<br>
+ %A1 = zext <2 x i32> %X to <2 x i64><br>
+ %B1 = xor <2 x i64> %A1, <i64 15, i64 15><br>
+ %C1 = mul <2 x i64> %B1, %B1<br>
+ %T1 = trunc <2 x i64> %C1 to <2 x i32><br>
+ call <2 x i32> @use32_vec(<2 x i32> %T1)<br>
+ ; make sure zext have another use that is not post-dominated by the TruncInst.<br>
+ call <2 x i32> @use64_vec(<2 x i64> %A1)<br>
+ ret void<br>
+}<br>
+<br>
+define void @multi_use_vec_and(<2 x i32> %X) {<br>
+; CHECK-LABEL: @multi_use_vec_and(<br>
+; CHECK-NEXT: [[A1:%.*]] = zext <2 x i32> [[X:%.*]] to <2 x i64><br>
+; CHECK-NEXT: [[B1:%.*]] = and <2 x i32> [[X]], <i32 15, i32 15><br>
+; CHECK-NEXT: [[C1:%.*]] = mul <2 x i32> [[B1]], [[B1]]<br>
+; CHECK-NEXT: [[TMP1:%.*]] = call <2 x i32> @use32_vec(<2 x i32> [[C1]])<br>
+; CHECK-NEXT: [[TMP2:%.*]] = call <2 x i32> @use64_vec(<2 x i64> [[A1]])<br>
+; CHECK-NEXT: ret void<br>
+;<br>
+ %A1 = zext <2 x i32> %X to <2 x i64><br>
+ %B1 = and <2 x i64> %A1, <i64 15, i64 15><br>
+ %C1 = mul <2 x i64> %B1, %B1<br>
+ %T1 = trunc <2 x i64> %C1 to <2 x i32><br>
+ call <2 x i32> @use32_vec(<2 x i32> %T1)<br>
+ ; make sure zext have another use that is not post-dominated by the TruncInst.<br>
+ call <2 x i32> @use64_vec(<2 x i64> %A1)<br>
+ ret void<br>
+}<br>
+<br>
+define void @multi_use_vec_sub(<2 x i32> %X, <2 x i32> %Y) {<br>
+; CHECK-LABEL: @multi_use_vec_sub(<br>
+; CHECK-NEXT: [[A1:%.*]] = zext <2 x i32> [[X:%.*]] to <2 x i64><br>
+; CHECK-NEXT: [[A2:%.*]] = zext <2 x i32> [[Y:%.*]] to <2 x i64><br>
+; CHECK-NEXT: [[B1:%.*]] = sub <2 x i32> [[X]], [[Y]]<br>
+; CHECK-NEXT: [[C1:%.*]] = mul <2 x i32> [[B1]], [[B1]]<br>
+; CHECK-NEXT: [[TMP1:%.*]] = call <2 x i32> @use32_vec(<2 x i32> [[C1]])<br>
+; CHECK-NEXT: [[TMP2:%.*]] = call <2 x i32> @use64_vec(<2 x i64> [[A1]])<br>
+; CHECK-NEXT: [[TMP3:%.*]] = call <2 x i32> @use64_vec(<2 x i64> [[A2]])<br>
+; CHECK-NEXT: ret void<br>
+;<br>
+ %A1 = zext <2 x i32> %X to <2 x i64><br>
+ %A2 = zext <2 x i32> %Y to <2 x i64><br>
+ %B1 = sub <2 x i64> %A1, %A2<br>
+ %C1 = mul <2 x i64> %B1, %B1<br>
+ %T1 = trunc <2 x i64> %C1 to <2 x i32><br>
+ call <2 x i32> @use32_vec(<2 x i32> %T1)<br>
+ ; make sure zext have another use that is not post-dominated by the TruncInst.<br>
+ call <2 x i32> @use64_vec(<2 x i64> %A1)<br>
+ call <2 x i32> @use64_vec(<2 x i64> %A2)<br>
+ ret void<br>
+}<br>
<br>
Modified: llvm/trunk/tools/opt/CMakeLists.txt<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/tools/opt/CMakeLists.txt?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/tools/opt/CMakeLists.txt?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/tools/opt/CMakeLists.txt (original)<br>
+++ llvm/trunk/tools/opt/CMakeLists.txt Wed Jan 24 04:42:42 2018<br>
@@ -1,5 +1,6 @@<br>
set(LLVM_LINK_COMPONENTS<br>
${LLVM_TARGETS_TO_BUILD}<br>
+ AggressiveInstCombine<br>
Analysis<br>
BitWriter<br>
CodeGen<br>
<br>
Modified: llvm/trunk/tools/opt/opt.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/tools/opt/opt.cpp?rev=323321&r1=323320&r2=323321&view=diff" target="_blank">
http://llvm.org/viewvc/llvm-project/llvm/trunk/tools/opt/opt.cpp?rev=323321&r1=323320&r2=323321&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/tools/opt/opt.cpp (original)<br>
+++ llvm/trunk/tools/opt/opt.cpp Wed Jan 24 04:42:42 2018<br>
@@ -395,6 +395,7 @@ int main(int argc, char **argv) {<br>
initializeAnalysis(Registry);<br>
initializeTransformUtils(Registry);<br>
initializeInstCombine(Registry);<br>
+ initializeAggressiveInstCombinerLegacyPassPass(Registry);<br>
initializeInstrumentation(Registry);<br>
initializeTarget(Registry);<br>
// For codegen passes, only passes that do IR to IR transformation are<br>
<br>
<br>
_______________________________________________<br>
llvm-commits mailing list<br>
<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><u></u><u></u></p>
</blockquote>
</div>
</div>
</blockquote>
</div>
</div>
</div></div></div><div lang="EN-US" link="blue" vlink="purple"><div class="m_8887644195325437420WordSection1"><div style="border:none;border-left:solid blue 1.5pt;padding:0in 0in 0in 4.0pt"></div>
</div>
<p>---------------------------------------------------------------------<br>
Intel Israel (74) Limited</p>
<p>This e-mail and any attachments may contain confidential material for<br>
the sole use of the intended recipient(s). Any review or distribution<br>
by others is strictly prohibited. If you are not the intended<br>
recipient, please contact the sender and delete all copies.</p></div>
</blockquote></div></div>