R600: Make sign_extend_inreg legal.

Don't know why I didn't just do this in the first place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206862 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 61 insertions, 91 deletions

diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index c1a607d25e..2990d0b91a 100644
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -1017,81 +1017,22 @@ SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
   MVT VT = Op.getSimpleValueType();
   MVT ScalarVT = VT.getScalarType();
 
-  unsigned SrcBits = ExtraVT.getScalarType().getSizeInBits();
-  unsigned DestBits = ScalarVT.getSizeInBits();
-  unsigned BitsDiff = DestBits - SrcBits;
-
-  if (!Subtarget->hasBFE())
-    return ExpandSIGN_EXTEND_INREG(Op, BitsDiff, DAG);
+  if (!VT.isVector())
+    return SDValue();
 
   SDValue Src = Op.getOperand(0);
-  if (VT.isVector()) {
-    SDLoc DL(Op);
-    // Need to scalarize this, and revisit each of the scalars later.
-    // TODO: Don't scalarize on Evergreen?
-    unsigned NElts = VT.getVectorNumElements();
-    SmallVector<SDValue, 8> Args;
-    DAG.ExtractVectorElements(Src, Args);
-
-    SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
-    for (unsigned I = 0; I < NElts; ++I)
-      Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
-
-    return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args.data(), Args.size());
-  }
-
-  if (SrcBits == 32) {
-    SDLoc DL(Op);
-
-    // If the source is 32-bits, this is really half of a 2-register pair, and
-    // we need to discard the unused half of the pair.
-    SDValue TruncSrc = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Src);
-    return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, TruncSrc);
-  }
-
-  unsigned NElts = VT.isVector() ? VT.getVectorNumElements() : 1;
-
-  // TODO: Match 64-bit BFE. SI has a 64-bit BFE, but it's scalar only so it
-  // might not be worth the effort, and will need to expand to shifts when
-  // fixing SGPR copies.
-  if (SrcBits < 32 && DestBits <= 32) {
-    SDLoc DL(Op);
-    MVT ExtVT = (NElts == 1) ? MVT::i32 : MVT::getVectorVT(MVT::i32, NElts);
-
-    if (DestBits != 32)
-      Src = DAG.getNode(ISD::ZERO_EXTEND, DL, ExtVT, Src);
-
-    // FIXME: This should use TargetConstant, but that hits assertions for
-    // Evergreen.
-    SDValue Ext = DAG.getNode(AMDGPUISD::BFE_I32, DL, ExtVT,
-                              Op.getOperand(0), // Operand
-                              DAG.getConstant(0, ExtVT), // Offset
-                              DAG.getConstant(SrcBits, ExtVT)); // Width
-
-    // Truncate to the original type if necessary.
-    if (ScalarVT == MVT::i32)
-      return Ext;
-    return DAG.getNode(ISD::TRUNCATE, DL, VT, Ext);
-  }
-
-  // For small types, extend to 32-bits first.
-  if (SrcBits < 32) {
-    SDLoc DL(Op);
-    MVT ExtVT = (NElts == 1) ? MVT::i32 : MVT::getVectorVT(MVT::i32, NElts);
+  SDLoc DL(Op);
 
-    SDValue TruncSrc = DAG.getNode(ISD::TRUNCATE, DL, ExtVT, Src);
-    SDValue Ext32 = DAG.getNode(AMDGPUISD::BFE_I32,
-                                DL,
-                                ExtVT,
-                                TruncSrc, // Operand
-                                DAG.getConstant(0, ExtVT), // Offset
-                                DAG.getConstant(SrcBits, ExtVT)); // Width
+  // TODO: Don't scalarize on Evergreen?
+  unsigned NElts = VT.getVectorNumElements();
+  SmallVector<SDValue, 8> Args;
+  DAG.ExtractVectorElements(Src, Args, 0, NElts);
 
-    return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Ext32);
-  }
+  SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
+  for (unsigned I = 0; I < NElts; ++I)
+    Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
 
-  // For everything else, use the standard bitshift expansion.
-  return ExpandSIGN_EXTEND_INREG(Op, BitsDiff, DAG);
+  return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args.data(), Args.size());
 }
 
 //===----------------------------------------------------------------------===//
diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index 6c2d3139a6..b7c35d71aa 100644
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -115,15 +115,15 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
   setOperationAction(ISD::SIGN_EXTEND, MVT::i64, Custom);
   setOperationAction(ISD::ZERO_EXTEND, MVT::i64, Custom);
 
-  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Custom);
+  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
 
-  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Custom);
+  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
 
-  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Custom);
+  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
 
@@ -165,9 +165,6 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
 
   setOperationAction(ISD::LOAD, MVT::i1, Custom);
 
-  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
-  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
-
   setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
   setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
   setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index a4ffd48287..a1d2cf43f7 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -74,6 +74,7 @@ def S_SEXT_I32_I8 : SOP1_32 <0x00000019, "S_SEXT_I32_I8",
 def S_SEXT_I32_I16 : SOP1_32 <0x0000001a, "S_SEXT_I32_I16",
   [(set i32:$dst, (sext_inreg i32:$src0, i16))]
 >;
+
 ////def S_BITSET0_B32 : SOP1_BITSET0 <0x0000001b, "S_BITSET0_B32", []>;
 ////def S_BITSET0_B64 : SOP1_BITSET0 <0x0000001c, "S_BITSET0_B64", []>;
 ////def S_BITSET1_B32 : SOP1_BITSET1 <0x0000001d, "S_BITSET1_B32", []>;
@@ -128,21 +129,6 @@ def S_CMPK_EQ_I32 : SOPK <
 >;
 */
 
-// Handle sext_inreg in i64
-def : Pat <
-  (i64 (sext_inreg i64:$src, i8)),
-  (INSERT_SUBREG (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
-    (S_SEXT_I32_I8 (EXTRACT_SUBREG i64:$src, sub0)), sub0),
-    (S_MOV_B32 -1), sub1)
->;
-
-def : Pat <
-  (i64 (sext_inreg i64:$src, i16)),
-  (INSERT_SUBREG (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
-    (S_SEXT_I32_I16 (EXTRACT_SUBREG i64:$src, sub0)), sub0),
-    (S_MOV_B32 -1), sub1)
->;
-
 let isCompare = 1 in {
 def S_CMPK_LG_I32 : SOPK_32 <0x00000004, "S_CMPK_LG_I32", []>;
 def S_CMPK_GT_I32 : SOPK_32 <0x00000005, "S_CMPK_GT_I32", []>;
@@ -2253,6 +2239,39 @@ def : Pat<
 >;
 
 //===----------------------------------------------------------------------===//
+// Conversion Patterns
+//===----------------------------------------------------------------------===//
+
+def : Pat<(i32 (sext_inreg i32:$src, i1)),
+  (S_BFE_I32 i32:$src, 65536)>; // 0 | 1 << 16
+
+// TODO: Match 64-bit BFE. SI has a 64-bit BFE, but it's scalar only so it
+// might not be worth the effort, and will need to expand to shifts when
+// fixing SGPR copies.
+
+// Handle sext_inreg in i64
+def : Pat <
+  (i64 (sext_inreg i64:$src, i1)),
+  (INSERT_SUBREG (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
+    (S_BFE_I32 (EXTRACT_SUBREG i64:$src, sub0), 65536), sub0), // 0 | 1 << 16
+    (S_MOV_B32 -1), sub1)
+>;
+
+def : Pat <
+  (i64 (sext_inreg i64:$src, i8)),
+  (INSERT_SUBREG (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
+    (S_SEXT_I32_I8 (EXTRACT_SUBREG i64:$src, sub0)), sub0),
+    (S_MOV_B32 -1), sub1)
+>;
+
+def : Pat <
+  (i64 (sext_inreg i64:$src, i16)),
+  (INSERT_SUBREG (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
+    (S_SEXT_I32_I16 (EXTRACT_SUBREG i64:$src, sub0)), sub0),
+    (S_MOV_B32 -1), sub1)
+>;
+
+//===----------------------------------------------------------------------===//
 // Miscellaneous Patterns
 //===----------------------------------------------------------------------===//
 
diff --git a/test/CodeGen/R600/sext-in-reg.ll b/test/CodeGen/R600/sext-in-reg.ll
index b722959aad..6ba7011b59 100644
--- a/test/CodeGen/R600/sext-in-reg.ll
+++ b/test/CodeGen/R600/sext-in-reg.ll
@@ -74,6 +74,19 @@ define void @sext_in_reg_i8_to_v1i32(<1 x i32> addrspace(1)* %out, <1 x i32> %a,
   ret void
 }
 
+; FUNC-LABEL: @sext_in_reg_i1_to_i64
+; SI: S_ADD_I32 [[VAL:s[0-9]+]],
+; SI: S_BFE_I32 s{{[0-9]+}}, s{{[0-9]+}}, 0x10000
+; SI: S_MOV_B32 {{s[0-9]+}}, -1
+; SI: BUFFER_STORE_DWORDX2
+define void @sext_in_reg_i1_to_i64(i64 addrspace(1)* %out, i64 %a, i64 %b) nounwind {
+  %c = add i64 %a, %b
+  %shl = shl i64 %c, 63
+  %ashr = ashr i64 %shl, 63
+  store i64 %ashr, i64 addrspace(1)* %out, align 8
+  ret void
+}
+
 ; FUNC-LABEL: @sext_in_reg_i8_to_i64
 ; SI: S_ADD_I32 [[VAL:s[0-9]+]],
 ; SI: S_SEXT_I32_I8 [[EXTRACT:s[0-9]+]], [[VAL]]