[X86][AVX512] add comi with Sae

[AsafBadouh]

Summary

add builtin_ia32_vcomisd and builtin_ia32_vcomisd.
both have comparison modifier to handle.
isX86CondUnordered - check if the modifier is for ordered (comi) or unordered (ucomi).
TranslateX86ConstCondToX86CC - translate from intel const to X86 condcode.

[delena]

../llvm/lib/Target/X86/X86ISelLowering.cpp

4002 ↗

(On Diff #39190)

(imm & 0x8) >> 3

4129 ↗

(On Diff #39190)

I think the whole function should be something like this:

{ return imm & 0x7; }

[AsafBadouh]

refactor and combine the two functions into one that returns tuple<isOrdered, x86::codeCond>

[delena]

LGTM

[AsafBadouh]

fix failure in Ubuntu.
add compare constants to the header.

[delena]

../commit/llvm_clean/lib/Target/X86/X86ISelLowering.cpp

16434

X86CC - unsigned
X86CC - a tuple, please use another name in the next case

16447

Please add comments here.

16450

Align the line

../commit/llvm_clean/lib/Target/X86/X86IntrinsicsInfo.h

1811

It may fail on platforms, where these constants are defined.

[AsafBadouh]

../commit/llvm_clean/lib/Target/X86/X86IntrinsicsInfo.h

1811

I've tested it on Ubuntu OS and Win

[AsafBadouh]

../commit/llvm_clean/lib/Target/X86/X86IntrinsicsInfo.h

1811

sorry, It is raise warning on Win.
I'll think on different solution.

[delena]

LGTM

../commit/llvm_clean/include/llvm/IR/IntrinsicsX86.td

def int_x86_avx512_mask_palignr_512 :
         Intrinsic<[llvm_v64i8_ty],
                   [llvm_v64i8_ty, llvm_v64i8_ty, llvm_i32_ty, llvm_v64i8_ty,
                    llvm_i64_ty], [IntrNoMem]>;
 }
 
 // Compares
 let TargetPrefix = "x86" in {
   // 512-bit
+  def int_x86_avx512_vcomi_sd : GCCBuiltin<"__builtin_ia32_vcomisd">,
+              Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty,
+                         llvm_v2f64_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_x86_avx512_vcomi_ss : GCCBuiltin<"__builtin_ia32_vcomiss">,
+              Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty,
+                         llvm_v4f32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
   def int_x86_avx512_mask_pcmpeq_b_512 : GCCBuiltin<"__builtin_ia32_pcmpeqb512_mask">,
         Intrinsic<[llvm_i64_ty], [llvm_v64i8_ty, llvm_v64i8_ty, llvm_i64_ty],
                   [IntrNoMem]>;
   def int_x86_avx512_mask_pcmpeq_w_512 : GCCBuiltin<"__builtin_ia32_pcmpeqw512_mask">,
         Intrinsic<[llvm_i32_ty], [llvm_v32i16_ty, llvm_v32i16_ty, llvm_i32_ty],
                   [IntrNoMem]>;
   def int_x86_avx512_mask_pcmpeq_d_512 : GCCBuiltin<"__builtin_ia32_pcmpeqd512_mask">,
             Intrinsic<[llvm_i16_ty], [llvm_v16i32_ty, llvm_v16i32_ty, llvm_i16_ty],

../commit/llvm_clean/lib/Target/X86/X86ISelLowering.cpp

case CMP_MASK_SCALAR_CC: {
       return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
                          DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
                          DAG.getValueType(MVT::i1));
     }
     case COMI: { // Comparison intrinsics
       ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
       SDValue LHS = Op.getOperand(1);
       SDValue RHS = Op.getOperand(2);
       unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
       assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
       SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
       SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
                                   DAG.getConstant(X86CC, dl, MVT::i8), Cond);
       return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
     }
+    case COMI_RM: { // Comparison intrinsics with Sae
+      SDValue LHS = Op.getOperand(1);
+      SDValue RHS = Op.getOperand(2);
+      SDValue CC = Op.getOperand(3);
+      SDValue Sae = Op.getOperand(4);
+      auto X86CC = TranslateX86ConstCondToX86CC(CC);
+      unsigned comiOp = std::get<0>(X86CC) ? IntrData->Opc0 : IntrData->Opc1;
+      SDValue Cond;
+      if (cast<ConstantSDNode>(Sae)->getZExtValue() !=
+                          X86::STATIC_ROUNDING::CUR_DIRECTION)
+        Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS, Sae);
+      else
+        Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS);
+      SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+        DAG.getConstant(std::get<1>(X86CC), dl, MVT::i8), Cond);
+      return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
+    }
     case VSHIFT:
       return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
                                  Op.getOperand(1), Op.getOperand(2), DAG);
     case VSHIFT_MASK:
       return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
                                                       Op.getSimpleValueType(),
                                                       Op.getOperand(1),
                                                       Op.getOperand(2), DAG),

../commit/llvm_clean/lib/Target/X86/X86InstrAVX512.td

defm VCVTPS2PHZ : avx512_cvtps2ph<v16i16x_info, v16f32_info, f256mem>,
                       EVEX, EVEX_V512, EVEX_CD8<32, CD8VH>;
   let Predicates = [HasVLX] in {
     defm VCVTPS2PHZ256 : avx512_cvtps2ph<v8i16x_info, v8f32x_info, f128mem>,
                         EVEX, EVEX_V256, EVEX_CD8<32, CD8VH>;
     defm VCVTPS2PHZ128 : avx512_cvtps2ph<v8i16x_info, v4f32x_info, f128mem>,
                         EVEX, EVEX_V128, EVEX_CD8<32, CD8VH>;
   }
 }
+
+//  Unordered/Ordered scalar fp compare with Sea and set EFLAGS
+multiclass avx512_ord_cmp_sae<bits<8> opc, X86VectorVTInfo _, SDNode OpNode,
+                            string OpcodeStr> {
+  def rb: AVX512<opc, MRMSrcReg, (outs), (ins _.RC:$src1, _.RC:$src2),
+                 !strconcat(OpcodeStr, "\t{{sae}, $src2, $src1|$src1, $src2, {sae}}"),
+                 [(set EFLAGS, (OpNode (_.VT _.RC:$src1), _.RC:$src2, 
+                                                        (i32 FROUND_NO_EXC)))],
+                 IIC_SSE_COMIS_RR>, EVEX, EVEX_B, VEX_LIG, EVEX_V128,
+                 Sched<[WriteFAdd]>;
+}
+
+let Defs = [EFLAGS], Predicates = [HasAVX512] in {
+  defm VUCOMISSZ : avx512_ord_cmp_sae<0x2E, v4f32x_info, X86ucomiSae, "vucomiss">,
+                                   AVX512PSIi8Base, EVEX_CD8<32, CD8VT1>;
+  defm VUCOMISDZ : avx512_ord_cmp_sae<0x2E, v2f64x_info, X86ucomiSae, "vucomisd">,
+                                   AVX512PDIi8Base, VEX_W, EVEX_CD8<64, CD8VT1>;
+  defm VCOMISSZ : avx512_ord_cmp_sae<0x2F, v4f32x_info, X86comiSae, "vcomiss">,
+                                   AVX512PSIi8Base, EVEX_CD8<32, CD8VT1>;
+  defm VCOMISDZ : avx512_ord_cmp_sae<0x2F, v2f64x_info, X86comiSae, "vcomisd">,
+                                   AVX512PDIi8Base, VEX_W, EVEX_CD8<64, CD8VT1>;
+}
+
 let Defs = [EFLAGS], Predicates = [HasAVX512] in {
   defm VUCOMISSZ : sse12_ord_cmp<0x2E, FR32X, X86cmp, f32, f32mem, loadf32,
                                  "ucomiss">, PS, EVEX, VEX_LIG,
                                  EVEX_CD8<32, CD8VT1>;
   defm VUCOMISDZ : sse12_ord_cmp<0x2E, FR64X, X86cmp, f64, f64mem, loadf64,
                                   "ucomisd">, PD, EVEX,
                                   VEX_LIG, VEX_W, EVEX_CD8<64, CD8VT1>;
   let Pattern = []<dag> in {

../commit/llvm_clean/lib/Target/X86/X86InstrFragmentsSIMD.td

 
 //===----------------------------------------------------------------------===//
 // SSE specific DAG Nodes.
 //===----------------------------------------------------------------------===//
 
 def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>,
                                        SDTCisFP<1>, SDTCisVT<3, i8>,
                                        SDTCisVec<1>]>;
+def SDTX86CmpTestSae : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, 
+                                     SDTCisSameAs<1, 2>, SDTCisInt<3>]>;
 
 def X86fmin    : SDNode<"X86ISD::FMIN",      SDTFPBinOp>;
 def X86fmax    : SDNode<"X86ISD::FMAX",      SDTFPBinOp>;
 
 // Commutative and Associative FMIN and FMAX.
 def X86fminc    : SDNode<"X86ISD::FMINC", SDTFPBinOp,
     [SDNPCommutative, SDNPAssociative]>;
 def X86fmaxc    : SDNode<"X86ISD::FMAXC", SDTFPBinOp,

 def X86frsqrt14s: SDNode<"X86ISD::FRSQRT",  SDTFPBinOp>;
 def X86frcp14s : SDNode<"X86ISD::FRCP",    SDTFPBinOp>;
 def X86fgetsign: SDNode<"X86ISD::FGETSIGNx86",SDTFPToIntOp>;
 def X86fhadd   : SDNode<"X86ISD::FHADD",     SDTFPBinOp>;
 def X86fhsub   : SDNode<"X86ISD::FHSUB",     SDTFPBinOp>;
 def X86hadd    : SDNode<"X86ISD::HADD",      SDTIntBinOp>;
 def X86hsub    : SDNode<"X86ISD::HSUB",      SDTIntBinOp>;
 def X86comi    : SDNode<"X86ISD::COMI",      SDTX86CmpTest>;
+def X86comiSae : SDNode<"X86ISD::COMI",      SDTX86CmpTestSae>;
 def X86ucomi   : SDNode<"X86ISD::UCOMI",     SDTX86CmpTest>;
+def X86ucomiSae: SDNode<"X86ISD::UCOMI",     SDTX86CmpTestSae>;
 def X86cmps    : SDNode<"X86ISD::FSETCC",     SDTX86Cmps>;
 //def X86cmpsd   : SDNode<"X86ISD::FSETCCsd",    SDTX86Cmpsd>;
 def X86cvtdq2pd: SDNode<"X86ISD::CVTDQ2PD",
                  SDTypeProfile<1, 1, [SDTCisVT<0, v2f64>,
                                       SDTCisVT<1, v4i32>]>>;
 def X86cvtudq2pd: SDNode<"X86ISD::CVTUDQ2PD",
                  SDTypeProfile<1, 1, [SDTCisVT<0, v2f64>,
                                       SDTCisVT<1, v4i32>]>>;

../commit/llvm_clean/lib/Target/X86/X86IntrinsicsInfo.h

 #define LLVM_LIB_TARGET_X86_X86INTRINSICSINFO_H
 
 namespace llvm {
 
 enum IntrinsicType {
   INTR_NO_TYPE,
   GATHER, SCATTER, PREFETCH, RDSEED, RDRAND, RDPMC, RDTSC, XTEST, ADX, FPCLASS, FPCLASSS,
   INTR_TYPE_1OP, INTR_TYPE_2OP, INTR_TYPE_2OP_IMM8, INTR_TYPE_3OP, INTR_TYPE_4OP,
-  CMP_MASK, CMP_MASK_CC,CMP_MASK_SCALAR_CC, VSHIFT, VSHIFT_MASK, COMI,
+  CMP_MASK, CMP_MASK_CC,CMP_MASK_SCALAR_CC, VSHIFT, VSHIFT_MASK, COMI, COMI_RM,
   INTR_TYPE_1OP_MASK, INTR_TYPE_1OP_MASK_RM,
   INTR_TYPE_2OP_MASK, INTR_TYPE_2OP_MASK_RM, INTR_TYPE_2OP_IMM8_MASK,
   INTR_TYPE_3OP_MASK, INTR_TYPE_3OP_MASK_RM, INTR_TYPE_3OP_IMM8_MASK,
   FMA_OP_MASK, FMA_OP_MASKZ, FMA_OP_MASK3, VPERM_3OP_MASK,
   VPERM_3OP_MASKZ, INTR_TYPE_SCALAR_MASK,
   INTR_TYPE_SCALAR_MASK_RM, INTR_TYPE_3OP_SCALAR_MASK_RM,
   COMPRESS_EXPAND_IN_REG, COMPRESS_TO_MEM,
   TRUNCATE_TO_MEM_VI8, TRUNCATE_TO_MEM_VI16, TRUNCATE_TO_MEM_VI32,

   X86_INTRINSIC_DATA(avx512_rcp28_sd,   INTR_TYPE_SCALAR_MASK_RM, X86ISD::RCP28, 0),
   X86_INTRINSIC_DATA(avx512_rcp28_ss,   INTR_TYPE_SCALAR_MASK_RM, X86ISD::RCP28, 0),
   X86_INTRINSIC_DATA(avx512_rsqrt14_sd, INTR_TYPE_SCALAR_MASK, X86ISD::FRSQRT, 0),
   X86_INTRINSIC_DATA(avx512_rsqrt14_ss, INTR_TYPE_SCALAR_MASK, X86ISD::FRSQRT, 0),
   X86_INTRINSIC_DATA(avx512_rsqrt28_pd, INTR_TYPE_1OP_MASK_RM,X86ISD::RSQRT28, 0),
   X86_INTRINSIC_DATA(avx512_rsqrt28_ps, INTR_TYPE_1OP_MASK_RM,X86ISD::RSQRT28, 0),
   X86_INTRINSIC_DATA(avx512_rsqrt28_sd, INTR_TYPE_SCALAR_MASK_RM,X86ISD::RSQRT28, 0),
   X86_INTRINSIC_DATA(avx512_rsqrt28_ss, INTR_TYPE_SCALAR_MASK_RM,X86ISD::RSQRT28, 0),
+  X86_INTRINSIC_DATA(avx512_vcomi_sd, COMI_RM, X86ISD::COMI, X86ISD::UCOMI),
+  X86_INTRINSIC_DATA(avx512_vcomi_ss, COMI_RM, X86ISD::COMI, X86ISD::UCOMI),
   X86_INTRINSIC_DATA(avx_hadd_pd_256,   INTR_TYPE_2OP, X86ISD::FHADD, 0),
   X86_INTRINSIC_DATA(avx_hadd_ps_256,   INTR_TYPE_2OP, X86ISD::FHADD, 0),
   X86_INTRINSIC_DATA(avx_hsub_pd_256,   INTR_TYPE_2OP, X86ISD::FHSUB, 0),
   X86_INTRINSIC_DATA(avx_hsub_ps_256,   INTR_TYPE_2OP, X86ISD::FHSUB, 0),
   X86_INTRINSIC_DATA(avx_max_pd_256,    INTR_TYPE_2OP, X86ISD::FMAX, 0),
   X86_INTRINSIC_DATA(avx_max_ps_256,    INTR_TYPE_2OP, X86ISD::FMAX, 0),
   X86_INTRINSIC_DATA(avx_min_pd_256,    INTR_TYPE_2OP, X86ISD::FMIN, 0),
   X86_INTRINSIC_DATA(avx_min_ps_256,    INTR_TYPE_2OP, X86ISD::FMIN, 0),

 static void verifyIntrinsicTables() {
   assert(std::is_sorted(std::begin(IntrinsicsWithoutChain),
                         std::end(IntrinsicsWithoutChain)) &&
          std::is_sorted(std::begin(IntrinsicsWithChain),
                         std::end(IntrinsicsWithChain)) &&
          "Intrinsic data tables should be sorted by Intrinsic ID");
 }
 
+// X86 specific compare constants.
+// They must be kept in synch with avxintrin.h
+#define _CMP_EQ_OQ    0x00 /* Equal (ordered, non-signaling)  */
+#define _CMP_LT_OS    0x01 /* Less-than (ordered, signaling)  */
+#define _CMP_LE_OS    0x02 /* Less-than-or-equal (ordered, signaling)  */
+#define _CMP_UNORD_Q  0x03 /* Unordered (non-signaling)  */
+#define _CMP_NEQ_UQ   0x04 /* Not-equal (unordered, non-signaling)  */
+#define _CMP_NLT_US   0x05 /* Not-less-than (unordered, signaling)  */
+#define _CMP_NLE_US   0x06 /* Not-less-than-or-equal (unordered, signaling)  */
+#define _CMP_ORD_Q    0x07 /* Ordered (nonsignaling)   */
+#define _CMP_EQ_UQ    0x08 /* Equal (unordered, non-signaling)  */
+#define _CMP_NGE_US   0x09 /* Not-greater-than-or-equal (unord, signaling)  */
+#define _CMP_NGT_US   0x0a /* Not-greater-than (unordered, signaling)  */
+#define _CMP_FALSE_OQ 0x0b /* False (ordered, non-signaling)  */
+#define _CMP_NEQ_OQ   0x0c /* Not-equal (ordered, non-signaling)  */
+#define _CMP_GE_OS    0x0d /* Greater-than-or-equal (ordered, signaling)  */
+#define _CMP_GT_OS    0x0e /* Greater-than (ordered, signaling)  */
+#define _CMP_TRUE_UQ  0x0f /* True (unordered, non-signaling)  */
+#define _CMP_EQ_OS    0x10 /* Equal (ordered, signaling)  */
+#define _CMP_LT_OQ    0x11 /* Less-than (ordered, non-signaling)  */
+#define _CMP_LE_OQ    0x12 /* Less-than-or-equal (ordered, non-signaling)  */
+#define _CMP_UNORD_S  0x13 /* Unordered (signaling)  */
+#define _CMP_NEQ_US   0x14 /* Not-equal (unordered, signaling)  */
+#define _CMP_NLT_UQ   0x15 /* Not-less-than (unordered, non-signaling)  */
+#define _CMP_NLE_UQ   0x16 /* Not-less-than-or-equal (unord, non-signaling)  */
+#define _CMP_ORD_S    0x17 /* Ordered (signaling)  */
+#define _CMP_EQ_US    0x18 /* Equal (unordered, signaling)  */
+#define _CMP_NGE_UQ   0x19 /* Not-greater-than-or-equal (unord, non-sign)  */
+#define _CMP_NGT_UQ   0x1a /* Not-greater-than (unordered, non-signaling)  */
+#define _CMP_FALSE_OS 0x1b /* False (ordered, signaling)  */
+#define _CMP_NEQ_OS   0x1c /* Not-equal (ordered, signaling)  */
+#define _CMP_GE_OQ    0x1d /* Greater-than-or-equal (ordered, non-signaling)  */
+#define _CMP_GT_OQ    0x1e /* Greater-than (ordered, non-signaling)  */
+#define _CMP_TRUE_US  0x1f /* True (unordered, signaling)  */
+
+/*
+* Get comparison modifier from _mm_comi_round_sd/ss intrinsic
+* Return tuple <isOrdered, X86 condcode>
+*/
+static std::tuple<bool,unsigned> TranslateX86ConstCondToX86CC(SDValue &imm) {
+  ConstantSDNode *CImm = dyn_cast<ConstantSDNode>(imm);
+  unsigned IntImm = CImm->getZExtValue();
+  // On a floating point condition, the flags are set as follows:
+  // ZF  PF  CF   op
+  //  0 | 0 | 0 | X > Y
+  //  0 | 0 | 1 | X < Y
+  //  1 | 0 | 0 | X == Y
+  //  1 | 1 | 1 | unordered
+  switch (IntImm) {
+  default: llvm_unreachable("Invalid floating point compare value for Comi!");
+  case _CMP_EQ_OQ:      // 0x00 - Equal (ordered, nonsignaling)
+  case _CMP_EQ_OS:      // 0x10 - Equal (ordered, signaling)
+    return std::make_tuple(true, X86::COND_E);
+  case _CMP_EQ_UQ:      // 0x08 - Equal (unordered, non-signaling)
+  case _CMP_EQ_US:      // 0x18 - Equal (unordered, signaling)
+    return std::make_tuple(false , X86::COND_E);
+  case _CMP_LT_OS:      // 0x01 - Less-than (ordered, signaling)
+  case _CMP_LT_OQ:      // 0x11 - Less-than (ordered, nonsignaling)
+    return std::make_tuple(true, X86::COND_B);
+  case _CMP_NGE_US:     // 0x09 - Not-greater-than-or-equal (unordered, signaling)
+  case _CMP_NGE_UQ:     // 0x19 - Not-greater-than-or-equal (unordered, nonsignaling)
+    return std::make_tuple(false , X86::COND_B);
+  case _CMP_LE_OS:      // 0x02 - Less-than-or-equal (ordered, signaling)
+  case _CMP_LE_OQ:      // 0x12 - Less-than-or-equal (ordered, nonsignaling)
+    return std::make_tuple(true, X86::COND_BE);
+  case _CMP_NGT_US:     // 0x0A - Not-greater-than (unordered, signaling)
+  case _CMP_NGT_UQ:     // 0x1A - Not-greater-than (unordered, nonsignaling)
+    return std::make_tuple(false, X86::COND_BE);
+  case _CMP_GT_OS:      // 0x0E - Greater-than (ordered, signaling)
+  case _CMP_GT_OQ:      // 0x1E - Greater-than (ordered, nonsignaling)
+    return std::make_tuple(true, X86::COND_A);
+  case _CMP_NLE_US:     // 0x06 - Not-less-than-or-equal (unordered,signaling)
+  case _CMP_NLE_UQ:     // 0x16 - Not-less-than-or-equal (unordered, nonsignaling)
+    return std::make_tuple(false, X86::COND_A);
+  case _CMP_GE_OS:      // 0x0D - Greater-than-or-equal (ordered, signaling)
+  case _CMP_GE_OQ:      // 0x1D - Greater-than-or-equal (ordered, nonsignaling)
+    return std::make_tuple(true, X86::COND_AE);
+  case _CMP_NLT_US:     // 0x05 - Not-less-than (unordered, signaling)
+  case _CMP_NLT_UQ:     // 0x15 - Not-less-than (unordered, nonsignaling)
+    return std::make_tuple(false, X86::COND_AE);
+  case _CMP_NEQ_OQ:     // 0x0C - Not-equal (ordered, non-signaling)
+  case _CMP_NEQ_OS:     // 0x1C - Not-equal (ordered, signaling)
+    return std::make_tuple(true, X86::COND_NE);
+  case _CMP_NEQ_UQ:     // 0x04 - Not-equal (unordered, nonsignaling)
+  case _CMP_NEQ_US:     // 0x14 - Not-equal (unordered, signaling)
+    return std::make_tuple(false, X86::COND_NE);
+  }
+}
+
 } // End llvm namespace
 
 #endif

../commit/llvm_clean/test/CodeGen/X86/avx512-intrinsics.ll

 ; CHECK-NEXT:    vpaddq %zmm0, %zmm3, %zmm0
 ; CHECK-NEXT:    retq
   %res = call <8 x i64> @llvm.x86.avx512.maskz.pternlog.q.512(<8 x i64> %x0, <8 x i64> %x1, <8 x i64> %x2, i32 33, i8 %x4)
   %res1 = call <8 x i64> @llvm.x86.avx512.maskz.pternlog.q.512(<8 x i64> %x0, <8 x i64> %x1, <8 x i64> %x2, i32 33, i8 -1)
   %res2 = add <8 x i64> %res, %res1
   ret <8 x i64> %res2
 }
 
+define i32 @test_x86_avx512_comi_sd_eq_sae(<2 x double> %a0, <2 x double> %a1) {
+; CHECK-LABEL: test_x86_avx512_comi_sd_eq_sae
+; CHECK: vcomisd {sae}, %xmm1, %xmm0
+; CHECK-NEXT: sete  %al
+  %res = call i32 @llvm.x86.avx512.vcomi.sd(<2 x double> %a0, <2 x double> %a1, i32 0, i32 8) 
+  ret i32 %res
+}
+
+define i32 @test_x86_avx512_ucomi_sd_eq_sae(<2 x double> %a0, <2 x double> %a1) {
+; CHECK-LABEL: test_x86_avx512_ucomi_sd_eq_sae
+; CHECK: vucomisd {sae}, %xmm1, %xmm0
+; CHECK-NEXT: sete %al
+  %res = call i32 @llvm.x86.avx512.vcomi.sd(<2 x double> %a0, <2 x double> %a1, i32 8, i32 8) 
+  ret i32 %res
+}
+
+define i32 @test_x86_avx512_comi_sd_eq(<2 x double> %a0, <2 x double> %a1) {
+; CHECK-LABEL: test_x86_avx512_comi_sd_eq
+; CHECK: vcomisd %xmm1, %xmm0
+; CHECK-NEXT: sete %al
+  %res = call i32 @llvm.x86.avx512.vcomi.sd(<2 x double> %a0, <2 x double> %a1, i32 0, i32 4) 
+  ret i32 %res
+}
+
+define i32 @test_x86_avx512_ucomi_sd_eq(<2 x double> %a0, <2 x double> %a1) {
+; CHECK-LABEL: test_x86_avx512_ucomi_sd_eq
+; CHECK: vucomisd %xmm1, %xmm0
+; CHECK-NEXT: sete %al
+  %res = call i32 @llvm.x86.avx512.vcomi.sd(<2 x double> %a0, <2 x double> %a1, i32 8, i32 4) 
+  ret i32 %res
+}
+
+define i32 @test_x86_avx512_comi_sd_lt_sae(<2 x double> %a0, <2 x double> %a1) {
+; CHECK-LABEL: test_x86_avx512_comi_sd_lt_sae
+; CHECK: vcomisd {sae}, %xmm1, %xmm0
+; CHECK-NEXT: sbbl %eax, %eax
+  %res = call i32 @llvm.x86.avx512.vcomi.sd(<2 x double> %a0, <2 x double> %a1, i32 1, i32 8) 
+  ret i32 %res
+}
+
+define i32 @test_x86_avx512_ucomi_sd_lt_sae(<2 x double> %a0, <2 x double> %a1) {
+; CHECK-LABEL: test_x86_avx512_ucomi_sd_lt_sae
+; CHECK: vucomisd {sae}, %xmm1, %xmm0
+; CHECK-NEXT: sbbl %eax, %eax
+  %res = call i32 @llvm.x86.avx512.vcomi.sd(<2 x double> %a0, <2 x double> %a1, i32 9, i32 8) 
+  ret i32 %res
+}
+
+define i32 @test_x86_avx512_comi_sd_lt(<2 x double> %a0, <2 x double> %a1) {
+; CHECK-LABEL: test_x86_avx512_comi_sd_lt
+; CHECK: vcomisd %xmm1, %xmm0
+; CHECK-NEXT: sbbl %eax, %eax
+  %res = call i32 @llvm.x86.avx512.vcomi.sd(<2 x double> %a0, <2 x double> %a1, i32 1, i32 4) 
+  ret i32 %res
+}
+
+define i32 @test_x86_avx512_ucomi_sd_lt(<2 x double> %a0, <2 x double> %a1) {
+; CHECK-LABEL: test_x86_avx512_ucomi_sd_lt
+; CHECK: vucomisd %xmm1, %xmm0
+; CHECK-NEXT: sbbl %eax, %eax
+  %res = call i32 @llvm.x86.avx512.vcomi.sd(<2 x double> %a0, <2 x double> %a1, i32 9, i32 4) 
+  ret i32 %res
+}
+
+declare i32 @llvm.x86.avx512.vcomi.sd(<2 x double>, <2 x double>, i32, i32) 
+
+define i32 @test_x86_avx512_ucomi_ss_lt(<4 x float> %a0, <4 x float> %a1) {
+; CHECK-LABEL: test_x86_avx512_ucomi_ss_lt
+; CHECK: vucomiss %xmm1, %xmm0
+; CHECK-NEXT: sbbl %eax, %eax
+  %res = call i32 @llvm.x86.avx512.vcomi.ss(<4 x float> %a0, <4 x float> %a1, i32 9, i32 4) 
+  ret i32 %res
+}
+
+declare i32 @llvm.x86.avx512.vcomi.ss(<4 x float>, <4 x float>, i32, i32)

../commit/llvm_clean/test/MC/X86/avx512-encodings.s

 // CHECK: vmovd  %xmm5, -512(%rdx)
 // CHECK:  encoding: [0xc5,0xf9,0x7e,0xaa,0x00,0xfe,0xff,0xff]
           vmovd  %xmm5, -512(%rdx)
 
 // CHECK: vmovd  %xmm5, -516(%rdx)
 // CHECK:  encoding: [0xc5,0xf9,0x7e,0xaa,0xfc,0xfd,0xff,0xff]
           vmovd  %xmm5, -516(%rdx)
 
+// CHECK: vcomisd %xmm21, %xmm23
+// CHECK:  encoding: [0x62,0xa1,0xfd,0x08,0x2f,0xfd]
+          vcomisd %xmm21, %xmm23
+
+// CHECK: vcomisd {sae}, %xmm21, %xmm23
+// CHECK:  encoding: [0x62,0xa1,0xfd,0x18,0x2f,0xfd]
+          vcomisd {sae}, %xmm21, %xmm23
+
+// CHECK: vcomisd (%rcx), %xmm23
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x08,0x2f,0x39]
+          vcomisd (%rcx), %xmm23
+
+// CHECK: vcomisd 291(%rax,%r14,8), %xmm23
+// CHECK:  encoding: [0x62,0xa1,0xfd,0x08,0x2f,0xbc,0xf0,0x23,0x01,0x00,0x00]
+          vcomisd 291(%rax,%r14,8), %xmm23
+
+// CHECK: vcomisd 1016(%rdx), %xmm23
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x08,0x2f,0x7a,0x7f]
+          vcomisd 1016(%rdx), %xmm23
+
+// CHECK: vcomisd 1024(%rdx), %xmm23
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x08,0x2f,0xba,0x00,0x04,0x00,0x00]
+          vcomisd 1024(%rdx), %xmm23
+
+// CHECK: vcomisd -1024(%rdx), %xmm23
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x08,0x2f,0x7a,0x80]
+          vcomisd -1024(%rdx), %xmm23
+
+// CHECK: vcomisd -1032(%rdx), %xmm23
+// CHECK:  encoding: [0x62,0xe1,0xfd,0x08,0x2f,0xba,0xf8,0xfb,0xff,0xff]
+          vcomisd -1032(%rdx), %xmm23
+
+// CHECK: vcomiss %xmm28, %xmm14
+// CHECK:  encoding: [0x62,0x11,0x7c,0x08,0x2f,0xf4]
+          vcomiss %xmm28, %xmm14
+
+// CHECK: vcomiss {sae}, %xmm28, %xmm14
+// CHECK:  encoding: [0x62,0x11,0x7c,0x18,0x2f,0xf4]
+          vcomiss {sae}, %xmm28, %xmm14
+
+// CHECK: vcomiss (%rcx), %xmm14
+// CHECK:  encoding: [0xc5,0x78,0x2f,0x31]
+          vcomiss (%rcx), %xmm14
+
+// CHECK: vcomiss 291(%rax,%r14,8), %xmm14
+// CHECK:  encoding: [0xc4,0x21,0x78,0x2f,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vcomiss 291(%rax,%r14,8), %xmm14
+
+// CHECK: vcomiss 508(%rdx), %xmm14
+// CHECK:  encoding: [0xc5,0x78,0x2f,0xb2,0xfc,0x01,0x00,0x00]
+          vcomiss 508(%rdx), %xmm14
+
+// CHECK: vcomiss 512(%rdx), %xmm14
+// CHECK:  encoding: [0xc5,0x78,0x2f,0xb2,0x00,0x02,0x00,0x00]
+          vcomiss 512(%rdx), %xmm14
+
+// CHECK: vcomiss -512(%rdx), %xmm14
+// CHECK:  encoding: [0xc5,0x78,0x2f,0xb2,0x00,0xfe,0xff,0xff]
+          vcomiss -512(%rdx), %xmm14
+
+// CHECK: vcomiss -516(%rdx), %xmm14
+// CHECK:  encoding: [0xc5,0x78,0x2f,0xb2,0xfc,0xfd,0xff,0xff]
+          vcomiss -516(%rdx), %xmm14
+
+// CHECK: vucomisd %xmm10, %xmm11
+// CHECK:  encoding: [0xc4,0x41,0x79,0x2e,0xda]
+          vucomisd %xmm10, %xmm11
+
+// CHECK: vucomisd {sae}, %xmm10, %xmm11
+// CHECK:  encoding: [0x62,0x51,0xfd,0x18,0x2e,0xda]
+          vucomisd {sae}, %xmm10, %xmm11
+
+// CHECK: vucomisd (%rcx), %xmm11
+// CHECK:  encoding: [0xc5,0x79,0x2e,0x19]
+          vucomisd (%rcx), %xmm11
+
+// CHECK: vucomisd 291(%rax,%r14,8), %xmm11
+// CHECK:  encoding: [0xc4,0x21,0x79,0x2e,0x9c,0xf0,0x23,0x01,0x00,0x00]
+          vucomisd 291(%rax,%r14,8), %xmm11
+
+// CHECK: vucomisd 1016(%rdx), %xmm11
+// CHECK:  encoding: [0xc5,0x79,0x2e,0x9a,0xf8,0x03,0x00,0x00]
+          vucomisd 1016(%rdx), %xmm11
+
+// CHECK: vucomisd 1024(%rdx), %xmm11
+// CHECK:  encoding: [0xc5,0x79,0x2e,0x9a,0x00,0x04,0x00,0x00]
+          vucomisd 1024(%rdx), %xmm11
+
+// CHECK: vucomisd -1024(%rdx), %xmm11
+// CHECK:  encoding: [0xc5,0x79,0x2e,0x9a,0x00,0xfc,0xff,0xff]
+          vucomisd -1024(%rdx), %xmm11
+
+// CHECK: vucomisd -1032(%rdx), %xmm11
+// CHECK:  encoding: [0xc5,0x79,0x2e,0x9a,0xf8,0xfb,0xff,0xff]
+          vucomisd -1032(%rdx), %xmm11
+
+// CHECK: vucomiss %xmm11, %xmm22
+// CHECK:  encoding: [0x62,0xc1,0x7c,0x08,0x2e,0xf3]
+          vucomiss %xmm11, %xmm22
+
+// CHECK: vucomiss {sae}, %xmm11, %xmm22
+// CHECK:  encoding: [0x62,0xc1,0x7c,0x18,0x2e,0xf3]
+          vucomiss {sae}, %xmm11, %xmm22
+
+// CHECK: vucomiss (%rcx), %xmm22
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x08,0x2e,0x31]
+          vucomiss (%rcx), %xmm22
+
+// CHECK: vucomiss 291(%rax,%r14,8), %xmm22
+// CHECK:  encoding: [0x62,0xa1,0x7c,0x08,0x2e,0xb4,0xf0,0x23,0x01,0x00,0x00]
+          vucomiss 291(%rax,%r14,8), %xmm22
+
+// CHECK: vucomiss 508(%rdx), %xmm22
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x08,0x2e,0x72,0x7f]
+          vucomiss 508(%rdx), %xmm22
+
+// CHECK: vucomiss 512(%rdx), %xmm22
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x08,0x2e,0xb2,0x00,0x02,0x00,0x00]
+          vucomiss 512(%rdx), %xmm22
+
+// CHECK: vucomiss -512(%rdx), %xmm22
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x08,0x2e,0x72,0x80]
+          vucomiss -512(%rdx), %xmm22
+
+// CHECK: vucomiss -516(%rdx), %xmm22
+// CHECK:  encoding: [0x62,0xe1,0x7c,0x08,0x2e,0xb2,0xfc,0xfd,0xff,0xff]
+          vucomiss -516(%rdx), %xmm22