From 3d200255d5b93344c1ab0a5ba4b47a52cfa5893e Mon Sep 17 00:00:00 2001
From: Justin Holewinski <jholewinski@nvidia.com>
Date: Thu, 29 Nov 2012 14:26:24 +0000
Subject: [PATCH] Allow targets to prefer TypeSplitVector over
 TypePromoteInteger when computing the legalization method for vectors

For some targets, it is desirable to prefer scalarizing <N x i1> instead of promoting to a larger legal type, such as <N x i32>.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@168882 91177308-0d34-0410-b5e6-96231b3b80d8
---
 include/llvm/Target/TargetLowering.h        |  5 +++++
 lib/CodeGen/SelectionDAG/TargetLowering.cpp |  2 +-
 lib/Target/NVPTX/NVPTXISelLowering.cpp      |  3 +++
 lib/Target/NVPTX/NVPTXISelLowering.h        |  4 ++++
 test/CodeGen/NVPTX/vector-compare.ll        | 19 +++++++++++++++++++
 5 files changed, 32 insertions(+), 1 deletion(-)
 create mode 100644 test/CodeGen/NVPTX/vector-compare.ll

diff --git a/include/llvm/Target/TargetLowering.h b/include/llvm/Target/TargetLowering.h
index 580a30fcd2d..aaa5f51f355 100644
--- a/include/llvm/Target/TargetLowering.h
+++ b/include/llvm/Target/TargetLowering.h
@@ -159,6 +159,11 @@ public:
 
   virtual bool isSelectSupported(SelectSupportKind kind) const { return true; }
 
+  /// shouldSplitVectorElementType - Return true if a vector of the given type
+  /// should be split (TypeSplitVector) instead of promoted
+  /// (TypePromoteInteger) during type legalization.
+  virtual bool shouldSplitVectorElementType(EVT VT) const { return false; }
+
   /// isIntDivCheap() - Return true if integer divide is usually cheaper than
   /// a sequence of several shifts, adds, and multiplies for this target.
   bool isIntDivCheap() const { return IntDivIsCheap; }
diff --git a/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 794935dad51..9ddcf7a9c6b 100644
--- a/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -825,7 +825,7 @@ void TargetLowering::computeRegisterProperties() {
     // that wider vector type.
     EVT EltVT = VT.getVectorElementType();
     unsigned NElts = VT.getVectorNumElements();
-    if (NElts != 1) {
+    if (NElts != 1 && !shouldSplitVectorElementType(EltVT)) {
       bool IsLegalWiderType = false;
       // First try to promote the elements of integer vectors. If no legal
       // promotion was found, fallback to the widen-vector method.
diff --git a/lib/Target/NVPTX/NVPTXISelLowering.cpp b/lib/Target/NVPTX/NVPTXISelLowering.cpp
index f1a99d77be9..a6f7e25718c 100644
--- a/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -271,6 +271,9 @@ const char *NVPTXTargetLowering::getTargetNodeName(unsigned Opcode) const {
   }
 }
 
+bool NVPTXTargetLowering::shouldSplitVectorElementType(EVT VT) const {
+  return VT == MVT::i1;
+}
 
 SDValue
 NVPTXTargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
diff --git a/lib/Target/NVPTX/NVPTXISelLowering.h b/lib/Target/NVPTX/NVPTXISelLowering.h
index 94a177ceb00..0a1833a7c9d 100644
--- a/lib/Target/NVPTX/NVPTXISelLowering.h
+++ b/lib/Target/NVPTX/NVPTXISelLowering.h
@@ -92,6 +92,8 @@ public:
   virtual unsigned getFunctionAlignment(const Function *F) const;
 
   virtual EVT getSetCCResultType(EVT VT) const {
+    if (VT.isVector())
+      return MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
     return MVT::i1;
   }
 
@@ -129,6 +131,8 @@ public:
     return MVT::i32;
   }
 
+  virtual bool shouldSplitVectorElementType(EVT VT) const;
+
 private:
   const NVPTXSubtarget &nvptxSubtarget;  // cache the subtarget here
 
diff --git a/test/CodeGen/NVPTX/vector-compare.ll b/test/CodeGen/NVPTX/vector-compare.ll
new file mode 100644
index 00000000000..21804999523
--- /dev/null
+++ b/test/CodeGen/NVPTX/vector-compare.ll
@@ -0,0 +1,19 @@
+; RUN: llc < %s -march=nvptx -mcpu=sm_20
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_20
+
+; This test makes sure that the result of vector compares are properly
+; scalarized.  If codegen fails, then the type legalizer incorrectly
+; tried to promote <2 x i1> to <2 x i8> and instruction selection failed.
+
+define void @foo(<2 x i32>* %a, <2 x i32>* %b, i32* %r1, i32* %r2) {
+  %aval = load <2 x i32>* %a
+  %bval = load <2 x i32>* %b
+  %res = icmp slt <2 x i32> %aval, %bval
+  %t1 = extractelement <2 x i1> %res, i32 0
+  %t2 = extractelement <2 x i1> %res, i32 1
+  %t1a = zext i1 %t1 to i32
+  %t2a = zext i1 %t2 to i32
+  store i32 %t1a, i32* %r1
+  store i32 %t2a, i32* %r2
+  ret void
+}