From 33eac3ee5309ee8ac0919e141fe74c9ba76ad92f Mon Sep 17 00:00:00 2001
From: Justin Holewinski <jholewinski@nvidia.com>
Date: Mon, 26 Jan 2015 19:52:20 +0000
Subject: [PATCH] [NVPTX] Generate a more optimal sequence for select of i1

Instead of creating a pattern like "(p && a) || ((!p) && b)",
just expand the i8 operands to i32 and perform the selp on them.

Fixes PR22246

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@227123 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/NVPTX/NVPTXISelLowering.cpp | 21 +++++++++++++++++++++
 lib/Target/NVPTX/NVPTXISelLowering.h   |  2 ++
 lib/Target/NVPTX/NVPTXInstrInfo.td     |  5 -----
 test/CodeGen/NVPTX/bug22246.ll         | 14 ++++++++++++++
 4 files changed, 37 insertions(+), 5 deletions(-)
 create mode 100644 test/CodeGen/NVPTX/bug22246.ll

diff --git a/lib/Target/NVPTX/NVPTXISelLowering.cpp b/lib/Target/NVPTX/NVPTXISelLowering.cpp
index 093ba1a2b82..3a13dc05b67 100644
--- a/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -259,6 +259,9 @@ NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM)
   setOperationAction(ISD::CTPOP, MVT::i32, Legal);
   setOperationAction(ISD::CTPOP, MVT::i64, Legal);
 
+  // PTX does not directly support SELP of i1, so promote to i32 first
+  setOperationAction(ISD::SELECT, MVT::i1, Custom);
+
   // We have some custom DAG combine patterns for these nodes
   setTargetDAGCombine(ISD::ADD);
   setTargetDAGCombine(ISD::AND);
@@ -1803,11 +1806,29 @@ NVPTXTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::SRA_PARTS:
   case ISD::SRL_PARTS:
     return LowerShiftRightParts(Op, DAG);
+  case ISD::SELECT:
+    return LowerSelect(Op, DAG);
   default:
     llvm_unreachable("Custom lowering not defined for operation");
   }
 }
 
+SDValue NVPTXTargetLowering::LowerSelect(SDValue Op, SelectionDAG &DAG) const {
+  SDValue Op0 = Op->getOperand(0);
+  SDValue Op1 = Op->getOperand(1);
+  SDValue Op2 = Op->getOperand(2);
+  SDLoc DL(Op.getNode());
+
+  assert(Op.getValueType() == MVT::i1 && "Custom lowering enabled only for i1");
+
+  Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op1);
+  Op2 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op2);
+  SDValue Select = DAG.getNode(ISD::SELECT, DL, MVT::i32, Op0, Op1, Op2);
+  SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Select);
+
+  return Trunc;
+}
+
 SDValue NVPTXTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
   if (Op.getValueType() == MVT::i1)
     return LowerLOADi1(Op, DAG);
diff --git a/lib/Target/NVPTX/NVPTXISelLowering.h b/lib/Target/NVPTX/NVPTXISelLowering.h
index b3fea3f4a36..c14e120b973 100644
--- a/lib/Target/NVPTX/NVPTXISelLowering.h
+++ b/lib/Target/NVPTX/NVPTXISelLowering.h
@@ -529,6 +529,8 @@ private:
   SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
 
+  SDValue LowerSelect(SDValue Op, SelectionDAG &DAG) const;
+
   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
                           SelectionDAG &DAG) const override;
   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
diff --git a/lib/Target/NVPTX/NVPTXInstrInfo.td b/lib/Target/NVPTX/NVPTXInstrInfo.td
index 5b8a169fb96..1c63668f0dd 100644
--- a/lib/Target/NVPTX/NVPTXInstrInfo.td
+++ b/lib/Target/NVPTX/NVPTXInstrInfo.td
@@ -1356,11 +1356,6 @@ defm SELP_u64 : SELP<"u64", Int64Regs, i64imm>;
 defm SELP_f32 : SELP_PATTERN<"f32", Float32Regs, f32imm, fpimm>;
 defm SELP_f64 : SELP_PATTERN<"f64", Float64Regs, f64imm, fpimm>;
 
-// Special select for predicate operands
-def : Pat<(i1 (select Int1Regs:$p, Int1Regs:$a, Int1Regs:$b)),
-              (ORb1rr (ANDb1rr Int1Regs:$p, Int1Regs:$a),
-              (ANDb1rr (NOT1 Int1Regs:$p), Int1Regs:$b))>;
-
 //
 // Funnnel shift in clamp mode
 //
diff --git a/test/CodeGen/NVPTX/bug22246.ll b/test/CodeGen/NVPTX/bug22246.ll
new file mode 100644
index 00000000000..70e7e12336e
--- /dev/null
+++ b/test/CodeGen/NVPTX/bug22246.ll
@@ -0,0 +1,14 @@
+; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
+
+target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64-nvidia-cuda"
+
+; CHECK-LABEL: _Z3foobbbPb
+define void @_Z3foobbbPb(i1 zeroext %p1, i1 zeroext %p2, i1 zeroext %p3, i8* nocapture %output) {
+entry:
+; CHECK: selp.b32       %r{{[0-9]+}}, %r{{[0-9]+}}, %r{{[0-9]+}}, %p{{[0-9]+}}
+  %.sink.v = select i1 %p1, i1 %p2, i1 %p3
+  %frombool5 = zext i1 %.sink.v to i8
+  store i8 %frombool5, i8* %output, align 1
+  ret void
+}
-- 
2.34.1