[X86][TD][vpmovm2 ] New TD pattern for the vpmovm2 instruction

Up until now, vpmovm2 instruction described its destination operand size
by the source operand size. This patch adds new pattern for the vpmovm2
instruction. The node describes new expansion of the destination (from
{128|256} to 512).

Differential Revision: https://reviews.llvm.org/D30654


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@298586 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/X86/X86ISelLowering.cpp

@@ -17747,17 +17747,10 @@ static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
 
   // SKX processor
   if ((InVTElt == MVT::i1) &&
-      (((Subtarget.hasBWI() && Subtarget.hasVLX() &&
-        VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
+      (((Subtarget.hasBWI() && VTElt.getSizeInBits() <= 16)) ||
 
-       ((Subtarget.hasBWI() && VT.is512BitVector() &&
-        VTElt.getSizeInBits() <= 16)) ||
+       ((Subtarget.hasDQI() && VTElt.getSizeInBits() >= 32))))
 
-       ((Subtarget.hasDQI() && Subtarget.hasVLX() &&
-        VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
-
-       ((Subtarget.hasDQI() && VT.is512BitVector() &&
-        VTElt.getSizeInBits() >= 32))))
     return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
 
   unsigned NumElts = VT.getVectorNumElements();
@@ -18002,7 +17995,8 @@ static SDValue LowerExtended1BitVectorLoad(SDValue Op,
   MVT VT = Op.getValueType().getSimpleVT();
   unsigned NumElts = VT.getVectorNumElements();
 
-  if ((Subtarget.hasVLX() && Subtarget.hasBWI() && Subtarget.hasDQI()) ||
+  if ((Subtarget.hasBWI() && NumElts >= 32) ||
+      (Subtarget.hasDQI() && NumElts < 16) ||
       NumElts == 16) {
     // Load and extend - everything is legal
     if (NumElts < 8) {
@@ -18031,7 +18025,7 @@ static SDValue LowerExtended1BitVectorLoad(SDValue Op,
 
   if (NumElts <= 8) {
     // A subset, assume that we have only AVX-512F
-    unsigned NumBitsToLoad = NumElts < 8 ? 8 : NumElts;
+    unsigned NumBitsToLoad = 8;
     MVT TypeToLoad = MVT::getIntegerVT(NumBitsToLoad);
     SDValue Load = DAG.getLoad(TypeToLoad, dl, Ld->getChain(),
                               Ld->getBasePtr(),

lib/Target/X86/X86InstrAVX512.td

@@ -7953,6 +7953,17 @@ def rr : AVX512XS8I<opc, MRMSrcReg, (outs Vec.RC:$dst), (ins Vec.KRC:$src),
                   [(set Vec.RC:$dst, (Vec.VT (X86vsext Vec.KRC:$src)))]>, EVEX;
 }
 
+// Use 512bit version to implement 128/256 bit in case NoVLX.
+multiclass avx512_convert_mask_to_vector_lowering<X86VectorVTInfo X86Info,
+                                                            X86VectorVTInfo _> {
+
+  def : Pat<(X86Info.VT (X86vsext (X86Info.KVT X86Info.KRC:$src))),
+            (X86Info.VT (EXTRACT_SUBREG
+                           (_.VT (!cast<Instruction>(NAME#"Zrr")
+                             (_.KVT (COPY_TO_REGCLASS X86Info.KRC:$src,_.KRC)))),
+                           X86Info.SubRegIdx))>;
+}
+
 multiclass cvt_mask_by_elt_width<bits<8> opc, AVX512VLVectorVTInfo VTInfo,
                                  string OpcodeStr, Predicate prd> {
 let Predicates = [prd] in
@@ -7962,20 +7973,17 @@ let Predicates = [prd] in
     defm Z256 : cvt_by_vec_width<opc, VTInfo.info256, OpcodeStr>, EVEX_V256;
     defm Z128 : cvt_by_vec_width<opc, VTInfo.info128, OpcodeStr>, EVEX_V128;
   }
-}
+let Predicates = [prd, NoVLX] in {
+   defm Z256_Alt :   avx512_convert_mask_to_vector_lowering<VTInfo.info256,VTInfo.info512>;
+   defm Z128_Alt :   avx512_convert_mask_to_vector_lowering<VTInfo.info128,VTInfo.info512>;
+  }
 
-multiclass avx512_convert_mask_to_vector<string OpcodeStr> {
-  defm NAME##B : cvt_mask_by_elt_width<0x28, avx512vl_i8_info,  OpcodeStr,
-                                       HasBWI>;
-  defm NAME##W : cvt_mask_by_elt_width<0x28, avx512vl_i16_info, OpcodeStr,
-                                       HasBWI>, VEX_W;
-  defm NAME##D : cvt_mask_by_elt_width<0x38, avx512vl_i32_info, OpcodeStr,
-                                       HasDQI>;
-  defm NAME##Q : cvt_mask_by_elt_width<0x38, avx512vl_i64_info, OpcodeStr,
-                                       HasDQI>, VEX_W;
 }
 
-defm VPMOVM2 : avx512_convert_mask_to_vector<"vpmovm2">;
+defm VPMOVM2B : cvt_mask_by_elt_width<0x28, avx512vl_i8_info, "vpmovm2" , HasBWI>;
+defm VPMOVM2W : cvt_mask_by_elt_width<0x28, avx512vl_i16_info, "vpmovm2", HasBWI> , VEX_W;
+defm VPMOVM2D : cvt_mask_by_elt_width<0x38, avx512vl_i32_info, "vpmovm2", HasDQI>;
+defm VPMOVM2Q : cvt_mask_by_elt_width<0x38, avx512vl_i64_info, "vpmovm2", HasDQI> , VEX_W;
 
 multiclass convert_vector_to_mask_common<bits<8> opc, X86VectorVTInfo _, string OpcodeStr > {
     def rr : AVX512XS8I<opc, MRMSrcReg, (outs _.KRC:$dst), (ins _.RC:$src),

test/CodeGen/X86/avx512-cvt.ll

@@ -1107,11 +1107,9 @@ define <16 x double> @sitofp_16i1_double(<16 x double> %a) {
 ; AVX512DQ-NEXT:    vxorpd %zmm2, %zmm2, %zmm2
 ; AVX512DQ-NEXT:    vcmpltpd %zmm1, %zmm2, %k0
 ; AVX512DQ-NEXT:    vcmpltpd %zmm0, %zmm2, %k1
-; AVX512DQ-NEXT:    vpmovm2q %k1, %zmm0
-; AVX512DQ-NEXT:    vpmovqd %zmm0, %ymm0
+; AVX512DQ-NEXT:    vpmovm2d %k1, %zmm0
 ; AVX512DQ-NEXT:    vcvtdq2pd %ymm0, %zmm0
-; AVX512DQ-NEXT:    vpmovm2q %k0, %zmm1
-; AVX512DQ-NEXT:    vpmovqd %zmm1, %ymm1
+; AVX512DQ-NEXT:    vpmovm2d %k0, %zmm1
 ; AVX512DQ-NEXT:    vcvtdq2pd %ymm1, %zmm1
 ; AVX512DQ-NEXT:    retq
   %cmpres = fcmp ogt <16 x double> %a, zeroinitializer
@@ -1150,8 +1148,7 @@ define <8 x double> @sitofp_8i1_double(<8 x double> %a) {
 ; AVX512DQ:       ## BB#0:
 ; AVX512DQ-NEXT:    vxorpd %zmm1, %zmm1, %zmm1
 ; AVX512DQ-NEXT:    vcmpltpd %zmm0, %zmm1, %k0
-; AVX512DQ-NEXT:    vpmovm2q %k0, %zmm0
-; AVX512DQ-NEXT:    vpmovqd %zmm0, %ymm0
+; AVX512DQ-NEXT:    vpmovm2d %k0, %zmm0
 ; AVX512DQ-NEXT:    vcvtdq2pd %ymm0, %zmm0
 ; AVX512DQ-NEXT:    retq
   %cmpres = fcmp ogt <8 x double> %a, zeroinitializer
@@ -1192,8 +1189,7 @@ define <8 x float> @sitofp_8i1_float(<8 x float> %a) {
 ; AVX512DQ-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
 ; AVX512DQ-NEXT:    vxorps %ymm1, %ymm1, %ymm1
 ; AVX512DQ-NEXT:    vcmpltps %zmm0, %zmm1, %k0
-; AVX512DQ-NEXT:    vpmovm2q %k0, %zmm0
-; AVX512DQ-NEXT:    vpmovqd %zmm0, %ymm0
+; AVX512DQ-NEXT:    vpmovm2d %k0, %zmm0
 ; AVX512DQ-NEXT:    vcvtdq2ps %ymm0, %ymm0
 ; AVX512DQ-NEXT:    retq
   %cmpres = fcmp ogt <8 x float> %a, zeroinitializer