[PATCH] D11304: [NVPTX] run LSR before straight-line optimizations

[Jingyue Wu]

jingyue created this revision.
jingyue added reviewers: jholewinski, eliben.
jingyue added a subscriber: llvm-commits.
Herald added a subscriber: jholewinski.

Straight-line optimizations can simplify the loop body and make LSR's
cost analysis more precise. This significantly improves several Eigen3
CUDA benchmarks.

With this change, EigenContractionKernel runs up to 40% faster
(https://bitbucket.org/eigen/eigen/src/753ceee5f206ff7dde9f6a41a5a420749fc9406f/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h?at=default#cl-502).
EigenConvolutionKernel2D runs up to 10% faster
(https://bitbucket.org/eigen/eigen/src/753ceee5f206ff7dde9f6a41a5a420749fc9406f/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h?at=default#cl-605).

I have some difficulties writing small tests that benefit from this
reordering due to a seemingly issue with LSR (being discussed at
http://lists.cs.uiuc.edu/pipermail/llvmdev/2015-July/088244.html).

http://reviews.llvm.org/D11304

Files:
  lib/Target/NVPTX/NVPTXTargetMachine.cpp

Index: lib/Target/NVPTX/NVPTXTargetMachine.cpp
===================================================================
--- lib/Target/NVPTX/NVPTXTargetMachine.cpp
+++ lib/Target/NVPTX/NVPTXTargetMachine.cpp
@@ -167,9 +167,10 @@
   disablePass(&TailDuplicateID);
 
   addPass(createNVPTXImageOptimizerPass());
-  TargetPassConfig::addIRPasses();
   addPass(createNVPTXAssignValidGlobalNamesPass());
   addPass(createGenericToNVVMPass());
+
+  // === Propagate special address spaces ===
   addPass(createNVPTXLowerKernelArgsPass(&getNVPTXTargetMachine()));
   // NVPTXLowerKernelArgs emits alloca for byval parameters which can often
   // be eliminated by SROA.
@@ -180,6 +181,8 @@
   // them unused. We could remove dead code in an ad-hoc manner, but that
   // requires manual work and might be error-prone.
   addPass(createDeadCodeEliminationPass());
+
+  // === Straight-line scalar optimizations ===
   addPass(createSeparateConstOffsetFromGEPPass());
   addPass(createSpeculativeExecutionPass());
   // ReassociateGEPs exposes more opportunites for SLSR. See
@@ -197,6 +200,22 @@
   // NaryReassociate on GEPs creates redundant common expressions, so run
   // EarlyCSE after it.
   addPass(createEarlyCSEPass());
+
+  // === LSR and other generic IR passes ===
+  TargetPassConfig::addIRPasses();
+  // EarlyCSE is not strong enough to clean up what LSR produces. For example,
+  // GVN can combine
+  //
+  //   %0 = add %a, %b
+  //   %1 = add %b, %a
+  //
+  // and
+  //
+  //   %0 = shl nsw %a, 2
+  //   %1 = shl %a, 2
+  //
+  // but EarlyCSE can do neither of them.
+  addPass(llvm::createGVNPass());
 }
 
 bool NVPTXPassConfig::addInstSelector() {


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D11304.30015.patch
Type: text/x-patch
Size: 1661 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150717/6d026647/attachment.bin>