[SveEmitter] Add builtins for compares and ReverseCompare flag.

The IsReverseCompare flag tells CGBuiltin to swap the operands,
so that a LT/LE intrinsics can be expressed in terms of GE/GT
intrinsics.

This patch also adds builtins for the wide-variants of the compares.

Reviewers: SjoerdMeijer, efriedma, ctetreau

Reviewed By: efriedma

Tags: #clang

Differential Revision: https://reviews.llvm.org/D78747

Author: sdesmalen-arm

clang/include/clang/Basic/TargetBuiltins.h

@@ -239,6 +239,7 @@ namespace clang {
     bool isOverloadWhileRW() const { return Flags & IsOverloadWhileRW; }
     bool isOverloadCvt() const { return Flags & IsOverloadCvt; }
     bool isPrefetch() const { return Flags & IsPrefetch; }
+    bool isReverseCompare() const { return Flags & ReverseCompare; }
 
     uint64_t getBits() const { return Flags; }
     bool isFlagSet(uint64_t Flag) const { return Flags & Flag; }

clang/include/clang/Basic/arm_sve.td

@@ -70,6 +70,9 @@
 // q: 1/4 width elements, 4x element count
 // o: 4x width elements, 1/4 element count
 //
+// w: vector of element type promoted to 64bits, vector maintains
+//    signedness of its element type.
+// j: element type promoted to 64bits (splat to vector type)
 // i: constant uint64_t
 // k: int32_t
 // l: int64_t
@@ -179,6 +182,7 @@ def IsOverloadCvt             : FlagType<0x00800000>; // Use {typeof(operand0),
 def OverloadKindMask          : FlagType<0x00E00000>; // When the masked values are all '0', the default type is used as overload type.
 def IsByteIndexed             : FlagType<0x01000000>;
 def IsPrefetch                : FlagType<0x08000000>; // Contiguous prefetches.
+def ReverseCompare            : FlagType<0x20000000>; // Compare operands must be swapped.
 
 // These must be kept in sync with the flags in include/clang/Basic/TargetBuiltins.h
 class ImmCheckType<int val> {
@@ -610,6 +614,53 @@ def SVEXT        : SInst<"svext[_{d}]",       "dddi", "csilUcUsUiUlhfd", MergeNo
 // Shifts
 def SVASRD_M : SInst<"svasrd[_n_{d}]", "dPdi", "csil",            MergeOp1,  "aarch64_sve_asrd", [], [ImmCheck<2, ImmCheckShiftRight, 1>]>;
 
+////////////////////////////////////////////////////////////////////////////////
+// Integer comparisons
+
+def SVCMPEQ : SInst<"svcmpeq[_{d}]", "PPdd", "csilUcUsUiUl", MergeNone, "aarch64_sve_cmpeq">;
+def SVCMPNE : SInst<"svcmpne[_{d}]", "PPdd", "csilUcUsUiUl", MergeNone, "aarch64_sve_cmpne">;
+def SVCMPGE : SInst<"svcmpge[_{d}]", "PPdd", "csil",         MergeNone, "aarch64_sve_cmpge">;
+def SVCMPGT : SInst<"svcmpgt[_{d}]", "PPdd", "csil",         MergeNone, "aarch64_sve_cmpgt">;
+def SVCMPLE : SInst<"svcmple[_{d}]", "PPdd", "csil",         MergeNone, "aarch64_sve_cmpge", [ReverseCompare]>;
+def SVCMPLT : SInst<"svcmplt[_{d}]", "PPdd", "csil",         MergeNone, "aarch64_sve_cmpgt", [ReverseCompare]>;
+def SVCMPHI : SInst<"svcmpgt[_{d}]", "PPdd", "UcUsUiUl",     MergeNone, "aarch64_sve_cmphi">;
+def SVCMPHS : SInst<"svcmpge[_{d}]", "PPdd", "UcUsUiUl",     MergeNone, "aarch64_sve_cmphs">;
+def SVCMPLO : SInst<"svcmplt[_{d}]", "PPdd", "UcUsUiUl",     MergeNone, "aarch64_sve_cmphi", [ReverseCompare]>;
+def SVCMPLS : SInst<"svcmple[_{d}]", "PPdd", "UcUsUiUl",     MergeNone, "aarch64_sve_cmphs", [ReverseCompare]>;
+
+def SVCMPEQ_N : SInst<"svcmpeq[_n_{d}]", "PPda", "csilUcUsUiUl", MergeNone, "aarch64_sve_cmpeq">;
+def SVCMPNE_N : SInst<"svcmpne[_n_{d}]", "PPda", "csilUcUsUiUl", MergeNone, "aarch64_sve_cmpne">;
+def SVCMPGE_N : SInst<"svcmpge[_n_{d}]", "PPda", "csil",         MergeNone, "aarch64_sve_cmpge">;
+def SVCMPGT_N : SInst<"svcmpgt[_n_{d}]", "PPda", "csil",         MergeNone, "aarch64_sve_cmpgt">;
+def SVCMPLE_N : SInst<"svcmple[_n_{d}]", "PPda", "csil",         MergeNone, "aarch64_sve_cmpge", [ReverseCompare]>;
+def SVCMPLT_N : SInst<"svcmplt[_n_{d}]", "PPda", "csil",         MergeNone, "aarch64_sve_cmpgt", [ReverseCompare]>;
+def SVCMPHS_N : SInst<"svcmpge[_n_{d}]", "PPda", "UcUsUiUl",     MergeNone, "aarch64_sve_cmphs">;
+def SVCMPHI_N : SInst<"svcmpgt[_n_{d}]", "PPda", "UcUsUiUl",     MergeNone, "aarch64_sve_cmphi">;
+def SVCMPLS_N : SInst<"svcmple[_n_{d}]", "PPda", "UcUsUiUl",     MergeNone, "aarch64_sve_cmphs", [ReverseCompare]>;
+def SVCMPLO_N : SInst<"svcmplt[_n_{d}]", "PPda", "UcUsUiUl",     MergeNone, "aarch64_sve_cmphi", [ReverseCompare]>;
+
+def SVCMPEQ_WIDE : SInst<"svcmpeq_wide[_{d}]", "PPdw", "csi",    MergeNone, "aarch64_sve_cmpeq_wide">;
+def SVCMPNE_WIDE : SInst<"svcmpne_wide[_{d}]", "PPdw", "csi",    MergeNone, "aarch64_sve_cmpne_wide">;
+def SVCMPGE_WIDE : SInst<"svcmpge_wide[_{d}]", "PPdw", "csi",    MergeNone, "aarch64_sve_cmpge_wide">;
+def SVCMPGT_WIDE : SInst<"svcmpgt_wide[_{d}]", "PPdw", "csi",    MergeNone, "aarch64_sve_cmpgt_wide">;
+def SVCMPLE_WIDE : SInst<"svcmple_wide[_{d}]", "PPdw", "csi",    MergeNone, "aarch64_sve_cmple_wide">;
+def SVCMPLT_WIDE : SInst<"svcmplt_wide[_{d}]", "PPdw", "csi",    MergeNone, "aarch64_sve_cmplt_wide">;
+def SVCMPHI_WIDE : SInst<"svcmpgt_wide[_{d}]", "PPdw", "UcUsUi", MergeNone, "aarch64_sve_cmphi_wide">;
+def SVCMPHS_WIDE : SInst<"svcmpge_wide[_{d}]", "PPdw", "UcUsUi", MergeNone, "aarch64_sve_cmphs_wide">;
+def SVCMPLO_WIDE : SInst<"svcmplt_wide[_{d}]", "PPdw", "UcUsUi", MergeNone, "aarch64_sve_cmplo_wide">;
+def SVCMPLS_WIDE : SInst<"svcmple_wide[_{d}]", "PPdw", "UcUsUi", MergeNone, "aarch64_sve_cmpls_wide">;
+
+def SVCMPEQ_WIDE_N : SInst<"svcmpeq_wide[_n_{d}]", "PPdj", "csi",    MergeNone, "aarch64_sve_cmpeq_wide">;
+def SVCMPNE_WIDE_N : SInst<"svcmpne_wide[_n_{d}]", "PPdj", "csi",    MergeNone, "aarch64_sve_cmpne_wide">;
+def SVCMPGE_WIDE_N : SInst<"svcmpge_wide[_n_{d}]", "PPdj", "csi",    MergeNone, "aarch64_sve_cmpge_wide">;
+def SVCMPGT_WIDE_N : SInst<"svcmpgt_wide[_n_{d}]", "PPdj", "csi",    MergeNone, "aarch64_sve_cmpgt_wide">;
+def SVCMPLE_WIDE_N : SInst<"svcmple_wide[_n_{d}]", "PPdj", "csi",    MergeNone, "aarch64_sve_cmple_wide">;
+def SVCMPLT_WIDE_N : SInst<"svcmplt_wide[_n_{d}]", "PPdj", "csi",    MergeNone, "aarch64_sve_cmplt_wide">;
+def SVCMPHS_WIDE_N : SInst<"svcmpge_wide[_n_{d}]", "PPdj", "UcUsUi", MergeNone, "aarch64_sve_cmphs_wide">;
+def SVCMPHI_WIDE_N : SInst<"svcmpgt_wide[_n_{d}]", "PPdj", "UcUsUi", MergeNone, "aarch64_sve_cmphi_wide">;
+def SVCMPLO_WIDE_N : SInst<"svcmplt_wide[_n_{d}]", "PPdj", "UcUsUi", MergeNone, "aarch64_sve_cmplo_wide">;
+def SVCMPLS_WIDE_N : SInst<"svcmple_wide[_n_{d}]", "PPdj", "UcUsUi", MergeNone, "aarch64_sve_cmpls_wide">;
+
 ////////////////////////////////////////////////////////////////////////////////
 // SVE2 - Narrowing DSP operations
 let ArchGuard = "defined(__ARM_FEATURE_SVE2)" in {
@@ -645,6 +696,36 @@ def SVTMAD  : SInst<"svtmad[_{d}]",  "dddi", "hfd", MergeNone, "aarch64_sve_ftma
 def SVMLA_LANE  : SInst<"svmla_lane[_{d}]",  "ddddi",  "hfd", MergeNone, "aarch64_sve_fmla_lane", [], [ImmCheck<3, ImmCheckLaneIndex, 2>]>;
 def SVCMLA_LANE : SInst<"svcmla_lane[_{d}]", "ddddii", "hf",  MergeNone, "aarch64_sve_fcmla_lane", [], [ImmCheck<3, ImmCheckLaneIndexCompRotate, 2>,
                                                                                                         ImmCheck<4, ImmCheckComplexRotAll90>]>;
+
+////////////////////////////////////////////////////////////////////////////////
+// Floating-point comparisons
+
+def SVACGE  : SInst<"svacge[_{d}]",  "PPdd", "hfd", MergeNone, "aarch64_sve_facge">;
+def SVACGT  : SInst<"svacgt[_{d}]",  "PPdd", "hfd", MergeNone, "aarch64_sve_facgt">;
+def SVACLE  : SInst<"svacle[_{d}]",  "PPdd", "hfd", MergeNone, "aarch64_sve_facge", [ReverseCompare]>;
+def SVACLT  : SInst<"svaclt[_{d}]",  "PPdd", "hfd", MergeNone, "aarch64_sve_facgt", [ReverseCompare]>;
+def SVCMPUO : SInst<"svcmpuo[_{d}]", "PPdd", "hfd", MergeNone, "aarch64_sve_fcmpuo">;
+
+def SVACGE_N  : SInst<"svacge[_n_{d}]",  "PPda", "hfd", MergeNone, "aarch64_sve_facge">;
+def SVACGT_N  : SInst<"svacgt[_n_{d}]",  "PPda", "hfd", MergeNone, "aarch64_sve_facgt">;
+def SVACLE_N  : SInst<"svacle[_n_{d}]",  "PPda", "hfd", MergeNone, "aarch64_sve_facge", [ReverseCompare]>;
+def SVACLT_N  : SInst<"svaclt[_n_{d}]",  "PPda", "hfd", MergeNone, "aarch64_sve_facgt", [ReverseCompare]>;
+def SVCMPUO_N : SInst<"svcmpuo[_n_{d}]", "PPda", "hfd", MergeNone, "aarch64_sve_fcmpuo">;
+
+def SVCMPEQ_F : SInst<"svcmpeq[_{d}]", "PPdd", "hfd", MergeNone, "aarch64_sve_fcmpeq">;
+def SVCMPNE_F : SInst<"svcmpne[_{d}]", "PPdd", "hfd", MergeNone, "aarch64_sve_fcmpne">;
+def SVCMPGE_F : SInst<"svcmpge[_{d}]", "PPdd", "hfd", MergeNone, "aarch64_sve_fcmpge">;
+def SVCMPGT_F : SInst<"svcmpgt[_{d}]", "PPdd", "hfd", MergeNone, "aarch64_sve_fcmpgt">;
+def SVCMPLE_F : SInst<"svcmple[_{d}]", "PPdd", "hfd", MergeNone, "aarch64_sve_fcmpge", [ReverseCompare]>;
+def SVCMPLT_F : SInst<"svcmplt[_{d}]", "PPdd", "hfd", MergeNone, "aarch64_sve_fcmpgt", [ReverseCompare]>;
+
+def SVCMPEQ_F_N : SInst<"svcmpeq[_n_{d}]", "PPda", "hfd", MergeNone, "aarch64_sve_fcmpeq">;
+def SVCMPNE_F_N : SInst<"svcmpne[_n_{d}]", "PPda", "hfd", MergeNone, "aarch64_sve_fcmpne">;
+def SVCMPGE_F_N : SInst<"svcmpge[_n_{d}]", "PPda", "hfd", MergeNone, "aarch64_sve_fcmpge">;
+def SVCMPGT_F_N : SInst<"svcmpgt[_n_{d}]", "PPda", "hfd", MergeNone, "aarch64_sve_fcmpgt">;
+def SVCMPLE_F_N : SInst<"svcmple[_n_{d}]", "PPda", "hfd", MergeNone, "aarch64_sve_fcmpge", [ReverseCompare]>;
+def SVCMPLT_F_N : SInst<"svcmplt[_n_{d}]", "PPda", "hfd", MergeNone, "aarch64_sve_fcmpgt", [ReverseCompare]>;
+
 ////////////////////////////////////////////////////////////////////////////////
 // Floating-point conversions
 

clang/lib/CodeGen/CGBuiltin.cpp

@@ -7892,6 +7892,9 @@ Value *CodeGenFunction::EmitAArch64SVEBuiltinExpr(unsigned BuiltinID,
       Ops[OpNo] = EmitSVEDupX(Ops[OpNo]);
     }
 
+    if (TypeFlags.isReverseCompare())
+      std::swap(Ops[1], Ops[2]);
+
     // Predicated intrinsics with _z suffix need a select w/ zeroinitializer.
     if (TypeFlags.getMergeType() == SVETypeFlags::MergeZero) {
       llvm::Type *OpndTy = Ops[1]->getType();