From 9735ccb7ead8598fded49a5ce2f0caf978afcccd Mon Sep 17 00:00:00 2001
From: Elena Demikhovsky <elena.demikhovsky@intel.com>
Date: Mon, 18 Aug 2014 11:59:06 +0000
Subject: [PATCH] AVX-512: Fixed a bug in emitting compare for MVT:i1 type.
 Added a test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215889 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86ISelDAGToDAG.cpp   | 26 ++++++++++++++++++++++----
 lib/Target/X86/X86ISelLowering.cpp   |  7 ++-----
 lib/Target/X86/X86InstrAVX512.td     |  4 ++++
 test/CodeGen/X86/avx512-cmp.ll       | 14 ++++++++++++++
 test/CodeGen/X86/avx512-trunc-ext.ll |  5 ++---
 5 files changed, 44 insertions(+), 12 deletions(-)

diff --git a/lib/Target/X86/X86ISelDAGToDAG.cpp b/lib/Target/X86/X86ISelDAGToDAG.cpp
index 1f5f44cf740..951c1a2ab33 100644
--- a/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -2544,12 +2544,30 @@ SDNode *X86DAGToDAGISel::Select(SDNode *Node) {
     SDValue N0 = Node->getOperand(0);
     SDValue N1 = Node->getOperand(1);
 
-    // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
-    // use a smaller encoding.
     if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
-        HasNoSignedComparisonUses(Node))
-      // Look past the truncate if CMP is the only use of it.
+        HasNoSignedComparisonUses(Node)) {
+      // Look for (X86cmp (truncate $op, i1), 0) and try to convert to a
+      // smaller encoding
+      if (Opcode == X86ISD::CMP && N0.getValueType() == MVT::i1 &&
+          X86::isZeroNode(N1)) {
+        SDValue Reg = N0.getOperand(0);
+        SDValue Imm = CurDAG->getTargetConstant(1, MVT::i8);
+
+        // Emit testb
+        if (Reg.getScalarValueSizeInBits() > 8)
+          Reg = CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Reg);
+        // Emit a testb.
+        SDNode *NewNode = CurDAG->getMachineNode(X86::TEST8ri, dl, MVT::i32,
+                                                Reg, Imm);
+        ReplaceUses(SDValue(Node, 0), SDValue(NewNode, 0));
+        return nullptr;
+      }
+
       N0 = N0.getOperand(0);
+    }
+    // Look for (X86cmp (and $op, $imm), 0) and see if we can convert it to
+    // use a smaller encoding.
+    // Look past the truncate if CMP is the only use of it.
     if ((N0.getNode()->getOpcode() == ISD::AND ||
          (N0.getResNo() == 0 && N0.getNode()->getOpcode() == X86ISD::AND)) &&
         N0.getNode()->hasOneUse() &&
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index b15588f1806..af8e1a18808 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -11917,12 +11917,9 @@ SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
   if (VT == MVT::i1) {
     assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
            "Invalid scalar TRUNCATE operation");
-    if (InVT == MVT::i32)
+    if (InVT.getSizeInBits() >= 32)
       return SDValue();
-    if (InVT.getSizeInBits() == 64)
-      In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::i32, In);
-    else if (InVT.getSizeInBits() < 32)
-      In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
+    In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
     return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
   }
   assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index 3678255e5f6..d5cfdb2ee9f 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -1198,6 +1198,10 @@ let Predicates = [HasBWI] in {
 }
 
 let Predicates = [HasAVX512] in {
+  def : Pat<(i1 (trunc (i64 GR64:$src))),
+            (COPY_TO_REGCLASS (KMOVWkr (AND32ri (EXTRACT_SUBREG $src, sub_32bit),
+                                        (i32 1))), VK1)>;
+
   def : Pat<(i1 (trunc (i32 GR32:$src))),
             (COPY_TO_REGCLASS (KMOVWkr (AND32ri $src, (i32 1))), VK1)>;
 
diff --git a/test/CodeGen/X86/avx512-cmp.ll b/test/CodeGen/X86/avx512-cmp.ll
index a44cb9d02f2..6e0d18558c5 100644
--- a/test/CodeGen/X86/avx512-cmp.ll
+++ b/test/CodeGen/X86/avx512-cmp.ll
@@ -85,3 +85,17 @@ define i32 @test8(i32 %a1, i32 %a2, i32 %a3) {
   %res = select i1 %tmp5, i32 1, i32 %a3
   ret i32 %res
 }
+
+; CHECK-LABEL: test9
+; CHECK: testb
+; CHECK-NOT: kmov
+; CHECK: ret
+define i32 @test9(i64 %a) {
+ %b = and i64 %a, 1
+ %cmp10.i = icmp eq i64 %b, 0
+ br i1 %cmp10.i, label %A, label %B
+A:
+ ret i32 6
+B:
+ ret i32 7
+}
diff --git a/test/CodeGen/X86/avx512-trunc-ext.ll b/test/CodeGen/X86/avx512-trunc-ext.ll
index 5e097be04cd..f1b639e110f 100644
--- a/test/CodeGen/X86/avx512-trunc-ext.ll
+++ b/test/CodeGen/X86/avx512-trunc-ext.ll
@@ -135,9 +135,8 @@ define <16 x i16> @trunc_v16i32_to_v16i16(<16 x i32> %x) {
 }
 
 ; CHECK-LABEL: trunc_i32_to_i1
-; CHECK: andl
-; CHECK: kmov
-; CHECK: kortest
+; CHECK: testb
+; CHECK: setne
 ; CKECK: orl
 ; CHECK: ret
 define i16 @trunc_i32_to_i1(i32 %a) {
-- 
2.34.1