[X86] int_x86_avx2_permps and X86ISD::VPERMV should take an integer vector for its...
[oota-llvm.git] / lib / Target / X86 / X86InstrFragmentsSIMD.td
index d7c90c1780aa847a24cf8732d95b4dcf1ef49685..8d7d48bd61aafb18e979b56a9a5e91796472d616 100644 (file)
@@ -76,13 +76,17 @@ def X86cvtudq2pd: SDNode<"X86ISD::CVTUDQ2PD",
                  SDTypeProfile<1, 1, [SDTCisVT<0, v2f64>,
                                       SDTCisVT<1, v4i32>]>>;
 def X86pshufb  : SDNode<"X86ISD::PSHUFB",
-                 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
+                 SDTypeProfile<1, 2, [SDTCVecEltisVT<0, i8>, SDTCisSameAs<0,1>,
                                       SDTCisSameAs<0,2>]>>;
 def X86psadbw  : SDNode<"X86ISD::PSADBW",
-                 SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
-                                      SDTCisSameAs<0,2>]>>;
+                 SDTypeProfile<1, 2, [SDTCVecEltisVT<0, i64>,
+                                      SDTCVecEltisVT<1, i8>,
+                                      SDTCisSameSizeAs<0,1>,
+                                      SDTCisSameAs<1,2>]>>;
 def X86dbpsadbw : SDNode<"X86ISD::DBPSADBW",
-                  SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisVec<1>,
+                  SDTypeProfile<1, 3, [SDTCVecEltisVT<0, i16>,
+                                       SDTCVecEltisVT<1, i8>,
+                                       SDTCisSameSizeAs<0,1>,
                                        SDTCisSameAs<1,2>, SDTCisInt<3>]>>;
 def X86andnp   : SDNode<"X86ISD::ANDNP",
                  SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
@@ -91,9 +95,11 @@ def X86psign   : SDNode<"X86ISD::PSIGN",
                  SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
                                       SDTCisSameAs<0,2>]>>;
 def X86pextrb  : SDNode<"X86ISD::PEXTRB",
-                 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
+                 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisVT<1, v16i8>,
+                                      SDTCisPtrTy<2>]>>;
 def X86pextrw  : SDNode<"X86ISD::PEXTRW",
-                 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
+                 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisVT<1, v8i16>,
+                                      SDTCisPtrTy<2>]>>;
 def X86pinsrb  : SDNode<"X86ISD::PINSRB",
                  SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
                                       SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
@@ -217,24 +223,26 @@ def X86vsrai   : SDNode<"X86ISD::VSRAI", SDTIntShiftOp>;
 
 def X86vprot   : SDNode<"X86ISD::VPROT",
                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
-                                      SDTCisVec<2>]>>;
+                                             SDTCisSameAs<0,2>]>>;
 def X86vproti  : SDNode<"X86ISD::VPROTI",
                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
-                                      SDTCisVT<2, i8>]>>;
+                                             SDTCisVT<2, i8>]>>;
 
 def X86vpshl   : SDNode<"X86ISD::VPSHL",
                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
-                                      SDTCisVec<2>]>>;
+                                             SDTCisSameAs<0,2>]>>;
 def X86vpsha   : SDNode<"X86ISD::VPSHA",
                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
-                                      SDTCisVec<2>]>>;
+                                             SDTCisSameAs<0,2>]>>;
 
 def X86vpcom   : SDNode<"X86ISD::VPCOM",
                         SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
-                                      SDTCisVec<2>, SDTCisVT<3, i8>]>>;
+                                             SDTCisSameAs<0,2>,
+                                             SDTCisVT<3, i8>]>>;
 def X86vpcomu  : SDNode<"X86ISD::VPCOMU",
                         SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
-                                      SDTCisVec<2>, SDTCisVT<3, i8>]>>;
+                                             SDTCisSameAs<0,2>,
+                                             SDTCisVT<3, i8>]>>;
 
 def SDTX86CmpPTest : SDTypeProfile<1, 2, [SDTCisVT<0, i32>,
                                           SDTCisVec<1>,
@@ -260,11 +268,15 @@ def X86testnm  : SDNode<"X86ISD::TESTNM", SDTypeProfile<1, 2, [SDTCisVec<0>,
 def X86select  : SDNode<"X86ISD::SELECT"     , SDTSelect>;
 
 def X86pmuludq : SDNode<"X86ISD::PMULUDQ",
-                        SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
-                                      SDTCisSameAs<1,2>]>>;
+                        SDTypeProfile<1, 2, [SDTCVecEltisVT<0, i64>,
+                                             SDTCVecEltisVT<1, i32>,
+                                             SDTCisSameSizeAs<0,1>,
+                                             SDTCisSameAs<1,2>]>>;
 def X86pmuldq  : SDNode<"X86ISD::PMULDQ",
-                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
-                                       SDTCisSameAs<1,2>]>>;
+                        SDTypeProfile<1, 2, [SDTCVecEltisVT<0, i64>,
+                                             SDTCVecEltisVT<1, i32>,
+                                             SDTCisSameSizeAs<0,1>,
+                                             SDTCisSameAs<1,2>]>>;
 
 def X86extrqi : SDNode<"X86ISD::EXTRQI",
                   SDTypeProfile<1, 3, [SDTCisVT<0, v2i64>, SDTCisSameAs<0,1>,
@@ -284,25 +296,27 @@ def SDTShuff3Op : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
                                 SDTCisSameAs<0,2>, SDTCisSameAs<0,3>]>;
 
 def SDTShuff2OpM : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
-                                        SDTCisVec<2>]>;
+                                        SDTCisSameSizeAs<0,2>,
+                                        SDTCisSameNumEltsAs<0,2>]>;
 def SDTShuff2OpI : SDTypeProfile<1, 2, [SDTCisVec<0>,
-                                 SDTCisSameAs<0,1>, SDTCisInt<2>]>;
+                                 SDTCisSameAs<0,1>, SDTCisVT<2, i8>]>;
 def SDTShuff3OpI : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
-                                 SDTCisSameAs<0,2>, SDTCisInt<3>]>;
+                                 SDTCisSameAs<0,2>, SDTCisVT<3, i8>]>;
 def SDTFPBinOpImmRound: SDTypeProfile<1, 4, [SDTCisVec<0>, SDTCisSameAs<0,1>,
                              SDTCisSameAs<0,2>, SDTCisInt<3>, SDTCisInt<4>]>;
 def SDTFPUnaryOpImmRound: SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
                               SDTCisInt<2>, SDTCisInt<3>]>;
 
 def SDTVBroadcast  : SDTypeProfile<1, 1, [SDTCisVec<0>]>;
-def SDTVBroadcastm : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>]>;
+def SDTVBroadcastm : SDTypeProfile<1, 1, [SDTCisVec<0>,
+                                          SDTCisInt<0>, SDTCisInt<1>]>;
 
 def SDTBlend : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
                              SDTCisSameAs<1,2>, SDTCisVT<3, i8>]>;
 
 def SDTTernlog  : SDTypeProfile<1, 4, [SDTCisVec<0>, SDTCisSameAs<0,1>,
                                 SDTCisSameAs<0,2>, SDTCisSameAs<0,3>,
-                                SDTCisInt<4>]>;
+                                SDTCisVT<4, i8>]>;
 
 def SDTFPBinOpRound : SDTypeProfile<1, 3, [      // fadd_round, fmul_round, etc.
   SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>, SDTCisFP<0>, SDTCisInt<3>]>;
@@ -315,11 +329,11 @@ def SDTFma : SDTypeProfile<1, 3, [SDTCisSameAs<0,1>,
 def SDTFmaRound : SDTypeProfile<1, 4, [SDTCisSameAs<0,1>,
                            SDTCisSameAs<1,2>, SDTCisSameAs<1,3>, SDTCisInt<4>]>;
 def STDFp1SrcRm : SDTypeProfile<1, 2, [SDTCisSameAs<0,1>,
-                           SDTCisVec<0>, SDTCisInt<2>]>;
+                           SDTCisVec<0>, SDTCisVT<2, i32>]>;
 def STDFp2SrcRm : SDTypeProfile<1, 3, [SDTCisSameAs<0,1>,
-                           SDTCisVec<0>, SDTCisInt<3>]>;
+                           SDTCisVec<0>, SDTCisVT<3, i32>]>;
 def STDFp3SrcRm : SDTypeProfile<1, 4, [SDTCisSameAs<0,1>,
-                           SDTCisVec<0>, SDTCisInt<3>, SDTCisInt<4>]>;
+                           SDTCisVec<0>, SDTCisVT<3, i32>, SDTCisVT<4, i32>]>;
 
 def X86PAlignr : SDNode<"X86ISD::PALIGNR", SDTShuff3OpI>;
 def X86VAlign  : SDNode<"X86ISD::VALIGN", SDTShuff3OpI>;
@@ -348,7 +362,9 @@ def X86Movhlps : SDNode<"X86ISD::MOVHLPS", SDTShuff2Op>;
 def X86Movlps : SDNode<"X86ISD::MOVLPS", SDTShuff2Op>;
 def X86Movlpd : SDNode<"X86ISD::MOVLPD", SDTShuff2Op>;
 
-def SDTPack : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>, SDTCisSameAs<2, 1>]>;
+def SDTPack : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
+                                   SDTCisSameSizeAs<0,1>,
+                                   SDTCisSameAs<1,2>]>;
 def X86Packss : SDNode<"X86ISD::PACKSS", SDTPack>;
 def X86Packus : SDNode<"X86ISD::PACKUS", SDTPack>;
 
@@ -360,10 +376,20 @@ def X86vpmaddwd    : SDNode<"X86ISD::VPMADDWD"   , SDTPack>;
 
 def X86VPermilpv  : SDNode<"X86ISD::VPERMILPV", SDTShuff2OpM>;
 def X86VPermilpi  : SDNode<"X86ISD::VPERMILPI", SDTShuff2OpI>;
-def X86VPermv     : SDNode<"X86ISD::VPERMV",    SDTShuff2Op>;
+def X86VPermv     : SDNode<"X86ISD::VPERMV",
+                           SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisInt<1>,
+                                                SDTCisSameNumEltsAs<0,1>,
+                                                SDTCisSameSizeAs<0,1>,
+                                                SDTCisSameAs<0,2>]>>;
 def X86VPermi     : SDNode<"X86ISD::VPERMI",    SDTShuff2OpI>;
-def X86VPermv3    : SDNode<"X86ISD::VPERMV3",   SDTShuff3Op>;
-def X86VPermiv3   : SDNode<"X86ISD::VPERMIV3",  SDTShuff3Op>;
+def X86VPermt2     : SDNode<"X86ISD::VPERMV3",
+                    SDTypeProfile<1, 3, [SDTCisVec<0>,
+                                         SDTCisSameAs<0,1>, SDTCisInt<2>,
+                                         SDTCisVec<2>, SDTCisSameNumEltsAs<0, 2>,
+                                         SDTCisSameSizeAs<0,2>,
+                                         SDTCisSameAs<0,3>]>, []>;
+
+def X86VPermi2X   : SDNode<"X86ISD::VPERMIV3", SDTShuff3Op>;
 def X86vpternlog  : SDNode<"X86ISD::VPTERNLOG", SDTTernlog>;
 
 def X86VPerm2x128 : SDNode<"X86ISD::VPERM2X128", SDTShuff3OpI>;
@@ -373,20 +399,31 @@ def X86VRange      : SDNode<"X86ISD::VRANGE",    SDTFPBinOpImmRound>;
 def X86VReduce     : SDNode<"X86ISD::VREDUCE",   SDTFPUnaryOpImmRound>;
 def X86VRndScale   : SDNode<"X86ISD::VRNDSCALE", SDTFPUnaryOpImmRound>;
 def X86VGetMant    : SDNode<"X86ISD::VGETMANT",  SDTFPUnaryOpImmRound>;
-def X86Vfpclass    : SDNode<"X86ISD::VFPCLASS", 
+def X86Vfpclass    : SDNode<"X86ISD::VFPCLASS",
                        SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCVecEltisVT<0, i1>,
-                                            SDTCisVec<1>, SDTCisInt<2>]>, []>;
-def X86Vfpclasss   : SDNode<"X86ISD::VFPCLASS", SDTypeProfile<1, 2, [SDTCisInt<0>,
-                              SDTCisFP<1>, SDTCisInt<2>]>,[]>;
+                                            SDTCisVec<1>, SDTCisFP<1>,
+                                            SDTCisSameNumEltsAs<0,1>,
+                                            SDTCisVT<2, i32>]>, []>;
+def X86Vfpclasss   : SDNode<"X86ISD::VFPCLASSS",
+                       SDTypeProfile<1, 2, [SDTCisVT<0, i1>,
+                                            SDTCisFP<1>, SDTCisVT<2, i32>]>,[]>;
 
 def X86SubVBroadcast : SDNode<"X86ISD::SUBV_BROADCAST",
                     SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
                                          SDTCisSubVecOfVec<1, 0>]>, []>;
+// SDTCisSubVecOfVec restriction cannot be applied for 128 bit version of VBROADCASTI32x2.
+def X86SubV32x2Broadcast : SDNode<"X86ISD::SUBV_BROADCAST",
+                    SDTypeProfile<1, 1, [SDTCisVec<0>,
+                                         SDTCisSameAs<0,1>]>, []>;
+
 def X86VBroadcast : SDNode<"X86ISD::VBROADCAST", SDTVBroadcast>;
+def X86VBroadcastm : SDNode<"X86ISD::VBROADCASTM", SDTVBroadcastm>;
 def X86Vinsert   : SDNode<"X86ISD::VINSERT",  SDTypeProfile<1, 3,
-                              [SDTCisSameAs<0, 1>, SDTCisPtrTy<3>]>, []>;
+                              [SDTCisSameAs<0, 1>, SDTCisEltOfVec<2, 1>,
+                               SDTCisPtrTy<3>]>, []>;
 def X86Vextract   : SDNode<"X86ISD::VEXTRACT",  SDTypeProfile<1, 2,
-                              [SDTCisVec<1>, SDTCisPtrTy<2>]>, []>;
+                              [SDTCisEltOfVec<0, 1>, SDTCisVec<1>,
+                               SDTCisPtrTy<2>]>, []>;
 
 def X86Blendi    : SDNode<"X86ISD::BLENDI",   SDTBlend>;
 
@@ -445,7 +482,8 @@ def X86expand  : SDNode<"X86ISD::EXPAND", SDTypeProfile<1, 1,
                               [SDTCisSameAs<0, 1>, SDTCisVec<1>]>, []>;
 
 def SDTintToFPRound: SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisFP<0>,
-                               SDTCisSameAs<0,1>, SDTCisInt<2>, SDTCisInt<3>]>;
+                                          SDTCisSameAs<0,1>, SDTCisInt<2>,
+                                          SDTCisVT<3, i32>]>;
 
 def SDTDoubleToInt: SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
                                          SDTCisInt<0>, SDTCVecEltisVT<1, f64>]>;
@@ -511,24 +549,29 @@ def X86cvtph2ps     : SDNode<"ISD::FP16_TO_FP",
                               SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
                                                    SDTCVecEltisVT<0, f32>,
                                                    SDTCVecEltisVT<1, i16>,
-                                                   SDTCisFP<0>, SDTCisInt<2>]> >;
+                                                   SDTCisFP<0>,
+                                                   SDTCisVT<2, i32>]> >;
 
 def X86cvtps2ph   : SDNode<"ISD::FP_TO_FP16",
                         SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisVec<1>,
                                              SDTCVecEltisVT<0, i16>,
                                              SDTCVecEltisVT<1, f32>,
-                                             SDTCisFP<1>, SDTCisInt<2>, SDTCisInt<3>]> >;
+                                             SDTCisFP<1>, SDTCisVT<2, i32>,
+                                             SDTCisVT<3, i32>]> >;
 def X86vfpextRnd  : SDNode<"X86ISD::VFPEXT",
                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
                                              SDTCisFP<0>, SDTCisFP<1>,
+                                             SDTCVecEltisVT<0, f64>,
+                                             SDTCVecEltisVT<1, f32>,
                                              SDTCisOpSmallerThanOp<1, 0>,
-                                             SDTCisInt<2>]>>;
+                                             SDTCisVT<2, i32>]>>;
 def X86vfproundRnd: SDNode<"X86ISD::VFPROUND",
                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
                                              SDTCisFP<0>, SDTCisFP<1>,
                                              SDTCVecEltisVT<0, f32>,
                                              SDTCVecEltisVT<1, f64>,
-                                             SDTCisInt<2>]>>;
+                                             SDTCisOpSmallerThanOp<0, 1>,
+                                             SDTCisVT<2, i32>]>>;
 
 //===----------------------------------------------------------------------===//
 // SSE Complex Patterns