Add support for converting llvm.fma to fma4 instructions.

llvm-svn: 162999

3 files changed, 76 insertions, 36 deletions

diff --git a/llvm/lib/Target/X86/Disassembler/X86Disassembler.cpp b/llvm/lib/Target/X86/Disassembler/X86Disassembler.cpp
index 5039887e1a2..a54d35dda2a 100644
--- a/llvm/lib/Target/X86/Disassembler/X86Disassembler.cpp
+++ b/llvm/lib/Target/X86/Disassembler/X86Disassembler.cpp
@@ -379,6 +379,8 @@ static void translateImmediate(MCInst &mcInst, uint64_t immediate,
   }
 
   switch (type) {
+  case TYPE_XMM32:
+  case TYPE_XMM64:
   case TYPE_XMM128:
     mcInst.addOperand(MCOperand::CreateReg(X86::XMM0 + (immediate >> 4)));
     return;
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index b2b30fe387c..f9184f693db 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -1052,7 +1052,7 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
     setOperationAction(ISD::VSELECT,           MVT::v8i32, Legal);
     setOperationAction(ISD::VSELECT,           MVT::v8f32, Legal);
 
-    if (Subtarget->hasFMA()) {
+    if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
       setOperationAction(ISD::FMA,             MVT::v8f32, Custom);
       setOperationAction(ISD::FMA,             MVT::v4f64, Custom);
       setOperationAction(ISD::FMA,             MVT::v4f32, Custom);
@@ -15606,7 +15606,8 @@ static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
     return SDValue();
 
   EVT ScalarVT = VT.getScalarType();
-  if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) || !Subtarget->hasFMA())
+  if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
+      (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
     return SDValue();
 
   SDValue A = N->getOperand(0);
@@ -15628,9 +15629,10 @@ static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
 
   unsigned Opcode;
   if (!NegMul)
-    Opcode = (!NegC)? X86ISD::FMADD : X86ISD::FMSUB;
+    Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
   else
-    Opcode = (!NegC)? X86ISD::FNMADD : X86ISD::FNMSUB;
+    Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
+
   return DAG.getNode(Opcode, dl, VT, A, B, C);
 }
 
diff --git a/llvm/lib/Target/X86/X86InstrFMA.td b/llvm/lib/Target/X86/X86InstrFMA.td
index 445bbaa4c17..f99440d973f 100644
--- a/llvm/lib/Target/X86/X86InstrFMA.td
+++ b/llvm/lib/Target/X86/X86InstrFMA.td
@@ -193,34 +193,57 @@ defm VFNMSUB : fma3s<0x9F, 0xAF, 0xBF, "vfnmsub", int_x86_fma_vfnmsub_ss,
 //===----------------------------------------------------------------------===//
 
 
-multiclass fma4s<bits<8> opc, string OpcodeStr, Operand memop,
-                 ComplexPattern mem_cpat, Intrinsic Int> {
-  def rr : FMA4<opc, MRMSrcReg, (outs VR128:$dst),
-           (ins VR128:$src1, VR128:$src2, VR128:$src3),
+multiclass fma4s<bits<8> opc, string OpcodeStr, RegisterClass RC,
+                 X86MemOperand x86memop, ValueType OpVT, SDNode OpNode,
+                 PatFrag mem_frag> {
+  def rr : FMA4<opc, MRMSrcReg, (outs RC:$dst),
+           (ins RC:$src1, RC:$src2, RC:$src3),
            !strconcat(OpcodeStr,
            "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
-           [(set VR128:$dst,
-             (Int VR128:$src1, VR128:$src2, VR128:$src3))]>, VEX_W, MemOp4;
-  def rm : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
-           (ins VR128:$src1, VR128:$src2, memop:$src3),
+           [(set RC:$dst,
+             (OpVT (OpNode RC:$src1, RC:$src2, RC:$src3)))]>, VEX_W, MemOp4;
+  def rm : FMA4<opc, MRMSrcMem, (outs RC:$dst),
+           (ins RC:$src1, RC:$src2, x86memop:$src3),
            !strconcat(OpcodeStr,
            "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
-           [(set VR128:$dst,
-             (Int VR128:$src1, VR128:$src2, mem_cpat:$src3))]>, VEX_W, MemOp4;
-  def mr : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
-           (ins VR128:$src1, memop:$src2, VR128:$src3),
+           [(set RC:$dst, (OpNode RC:$src1, RC:$src2,
+                           (mem_frag addr:$src3)))]>, VEX_W, MemOp4;
+  def mr : FMA4<opc, MRMSrcMem, (outs RC:$dst),
+           (ins RC:$src1, x86memop:$src2, RC:$src3),
            !strconcat(OpcodeStr,
            "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
-           [(set VR128:$dst,
-             (Int VR128:$src1, mem_cpat:$src2, VR128:$src3))]>;
+           [(set RC:$dst,
+             (OpNode RC:$src1, (mem_frag addr:$src2), RC:$src3))]>;
 // For disassembler
 let isCodeGenOnly = 1 in
-  def rr_REV : FMA4<opc, MRMSrcReg, (outs VR128:$dst),
-               (ins VR128:$src1, VR128:$src2, VR128:$src3),
+  def rr_REV : FMA4<opc, MRMSrcReg, (outs RC:$dst),
+               (ins RC:$src1, RC:$src2, RC:$src3),
                !strconcat(OpcodeStr,
                "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), []>;
 }
 
+multiclass fma4s_int<bits<8> opc, string OpcodeStr, Operand memop,
+                     ComplexPattern mem_cpat, Intrinsic Int> {
+  def rr_Int : FMA4<opc, MRMSrcReg, (outs VR128:$dst),
+               (ins VR128:$src1, VR128:$src2, VR128:$src3),
+               !strconcat(OpcodeStr,
+               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
+               [(set VR128:$dst,
+                 (Int VR128:$src1, VR128:$src2, VR128:$src3))]>, VEX_W, MemOp4;
+  def rm_Int : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
+               (ins VR128:$src1, VR128:$src2, memop:$src3),
+               !strconcat(OpcodeStr,
+               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
+               [(set VR128:$dst, (Int VR128:$src1, VR128:$src2,
+                                  mem_cpat:$src3))]>, VEX_W, MemOp4;
+  def mr_Int : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
+               (ins VR128:$src1, memop:$src2, VR128:$src3),
+               !strconcat(OpcodeStr,
+               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
+               [(set VR128:$dst,
+                 (Int VR128:$src1, mem_cpat:$src2, VR128:$src3))]>;
+}
+
 multiclass fma4p<bits<8> opc, string OpcodeStr, SDNode OpNode,
                  ValueType OpVT128, ValueType OpVT256,
                  PatFrag ld_frag128, PatFrag ld_frag256> {
@@ -277,34 +300,47 @@ let isCodeGenOnly = 1 in {
 
 let Predicates = [HasFMA4] in {
 
-defm VFMADDSS4    : fma4s<0x6A, "vfmaddss", ssmem, sse_load_f32,
-                          int_x86_fma_vfmadd_ss>;
-defm VFMADDSD4    : fma4s<0x6B, "vfmaddsd", sdmem, sse_load_f64,
-                          int_x86_fma_vfmadd_sd>;
+defm VFMADDSS4  : fma4s<0x6A, "vfmaddss", FR32, f32mem, f32, X86Fmadd, loadf32>,
+                  fma4s_int<0x6A, "vfmaddss", ssmem, sse_load_f32,
+                            int_x86_fma_vfmadd_ss>;
+defm VFMADDSD4  : fma4s<0x6B, "vfmaddsd", FR64, f64mem, f64, X86Fmadd, loadf64>,
+                  fma4s_int<0x6B, "vfmaddsd", sdmem, sse_load_f64,
+                            int_x86_fma_vfmadd_sd>;
+defm VFMSUBSS4  : fma4s<0x6E, "vfmsubss", FR32, f32mem, f32, X86Fmsub, loadf32>,
+                  fma4s_int<0x6E, "vfmsubss", ssmem, sse_load_f32,
+                            int_x86_fma_vfmsub_ss>;
+defm VFMSUBSD4  : fma4s<0x6F, "vfmsubsd", FR64, f64mem, f64, X86Fmsub, loadf64>,
+                  fma4s_int<0x6F, "vfmsubsd", sdmem, sse_load_f64,
+                            int_x86_fma_vfmsub_sd>;
+defm VFNMADDSS4 : fma4s<0x7A, "vfnmaddss", FR32, f32mem, f32,
+                        X86Fnmadd, loadf32>,
+                  fma4s_int<0x7A, "vfnmaddss", ssmem, sse_load_f32,
+                            int_x86_fma_vfnmadd_ss>;
+defm VFNMADDSD4 : fma4s<0x7B, "vfnmaddsd", FR64, f64mem, f64,
+                        X86Fnmadd, loadf64>,
+                  fma4s_int<0x7B, "vfnmaddsd", sdmem, sse_load_f64,
+                            int_x86_fma_vfnmadd_sd>;
+defm VFNMSUBSS4 : fma4s<0x7E, "vfnmsubss", FR32, f32mem, f32,
+                        X86Fnmsub, loadf32>,
+                  fma4s_int<0x7E, "vfnmsubss", ssmem, sse_load_f32,
+                            int_x86_fma_vfnmsub_ss>;
+defm VFNMSUBSD4 : fma4s<0x7F, "vfnmsubsd", FR64, f64mem, f64,
+                        X86Fnmsub, loadf64>,
+                  fma4s_int<0x7F, "vfnmsubsd", sdmem, sse_load_f64,
+                            int_x86_fma_vfnmsub_sd>;
+
 defm VFMADDPS4    : fma4p<0x68, "vfmaddps", X86Fmadd, v4f32, v8f32,
                           memopv4f32, memopv8f32>;
 defm VFMADDPD4    : fma4p<0x69, "vfmaddpd", X86Fmadd, v2f64, v4f64,
                           memopv2f64, memopv4f64>;
-defm VFMSUBSS4    : fma4s<0x6E, "vfmsubss", ssmem, sse_load_f32,
-                          int_x86_fma_vfmsub_ss>;
-defm VFMSUBSD4    : fma4s<0x6F, "vfmsubsd", sdmem, sse_load_f64,
-                          int_x86_fma_vfmsub_sd>;
 defm VFMSUBPS4    : fma4p<0x6C, "vfmsubps", X86Fmsub, v4f32, v8f32,
                           memopv4f32, memopv8f32>;
 defm VFMSUBPD4    : fma4p<0x6D, "vfmsubpd", X86Fmsub, v2f64, v4f64,
                           memopv2f64, memopv4f64>;
-defm VFNMADDSS4   : fma4s<0x7A, "vfnmaddss", ssmem, sse_load_f32,
-                          int_x86_fma_vfnmadd_ss>;
-defm VFNMADDSD4   : fma4s<0x7B, "vfnmaddsd", sdmem, sse_load_f64,
-                          int_x86_fma_vfnmadd_sd>;
 defm VFNMADDPS4   : fma4p<0x78, "vfnmaddps", X86Fnmadd, v4f32, v8f32,
                           memopv4f32, memopv8f32>;
 defm VFNMADDPD4   : fma4p<0x79, "vfnmaddpd", X86Fnmadd, v2f64, v4f64,
                           memopv2f64, memopv4f64>;
-defm VFNMSUBSS4   : fma4s<0x7E, "vfnmsubss", ssmem, sse_load_f32,
-                          int_x86_fma_vfnmsub_ss>;
-defm VFNMSUBSD4   : fma4s<0x7F, "vfnmsubsd", sdmem, sse_load_f64,
-                          int_x86_fma_vfnmsub_sd>;
 defm VFNMSUBPS4   : fma4p<0x7C, "vfnmsubps", X86Fnmsub, v4f32, v8f32,
                           memopv4f32, memopv8f32>;
 defm VFNMSUBPD4   : fma4p<0x7D, "vfnmsubpd", X86Fnmsub, v2f64, v4f64,