[ARM] Fix conditions for lowering to S[LR]I

[PetreTudor]

Summary

Fixed wrong conditions for generating (S[LR]I X, Y, C2) from
(or (and X, BvecC1), (lsl Y, C2)). The optimisation is also
enabled by default now.

[dmgreen]

Looks good.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

7963–7964

I know you didn't write this part, but it's generally not a great idea to lower via intrinsics if it can be helped.

Can you add a ISel node, a lot like ARMISD::VSHLIMM, and plumb that through tablegen instead?

[PetreTudor]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

7963–7964

Sure! I will start working on this right away.

[PetreTudor]

Now lowering to S[LR]I via TableGen + ISel nodes.

[PetreTudor]

Fix styling problem caused by linter (inconsistent with local style).

[dmgreen]

Thanks. LGTM.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

3153–3157

You could possibly use uint64_t ShiftAmount = Op.getConstantOperandVal(3), simplify it a little. It won't handle the report_fatal_error, but this should have already been checked at a higher level, somewhere in clang to make sure the shift amount is a constant.

Same for the ElemSizeInBits check below. You could change them to asserts. Up to you.

[PetreTudor]

Simplified logic for converting intrinsics to AArch64ISD::VS[LR]I.

[mstorsjo]

This seems to cause misoptimizations for me:

$ cat test.c
void rgb16to24(const unsigned char *src, unsigned char *dst, int src_size) {
    unsigned char *d = dst;
    const unsigned short *s = (const unsigned short *)src;
    const unsigned short *end = s + src_size / 2;

    while (s < end) {
        unsigned short bgr = *s++;
        *d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
        *d++ = ((bgr&0x07E0)>>3) | ((bgr&0x07E0)>> 9);
        *d++ = ((bgr&0x001F)<<3) | ((bgr&0x001F)>> 2);
    }
}
$ clang -target aarch64-linux-gnu -S -O3 test.c -o test.s

There's quite a bit of difference in the output of the compiler for this function caused by this patch - I haven't yet traced it and tracked down exactly which bit of it is wrong, but the testsuite that contains this code reports unexpected values.

[mstorsjo]

Filed it as a bug at https://bugs.llvm.org/show_bug.cgi?id=45610 as well.

[efriedma]

llvm/test/CodeGen/AArch64/arm64-sli-sri-opt.ll

5

Please generate simple tests like this with update_llc_test_checks.py.

This transform is simply wrong, as written: sli is not the vector version of bfi.

[PetreTudor]

It appears that there were probably more things wrong with this optimization than just the condition.
I have reverted the patch for now so that I can take more time to make sure that I get this fix right.

llvm/lib/Target/AArch64/AArch64ISelLowering.h

enum NodeType : unsigned {
 
   // Vector shift by scalar (again)
   SQSHL_I,
   UQSHL_I,
   SQSHLU_I,
   SRSHR_I,
   URSHR_I,
 
+  // Vector shift by constant and insert
+  VSLI,
+  VSRI,
+
   // Vector comparisons
   CMEQ,
   CMGE,
   CMGT,
   CMHI,
   CMHS,
   FCMEQ,
   FCMGE,
enum NodeType : unsigned {
   /// need to re-interpret the data in SIMD vector registers in big-endian
   /// mode without emitting such REV instructions.
   NVCAST,
 
   SMULL,
   UMULL,
 
   // Reciprocal estimates and steps.
-  FRECPE, FRECPS,
-  FRSQRTE, FRSQRTS,
+  FRECPE,
+  FRECPS,
+  FRSQRTE,
+  FRSQRTS,
 
   SUNPKHI,
   SUNPKLO,
   UUNPKHI,
   UUNPKLO,
 
   CLASTA_N,
   CLASTB_N,

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

 using namespace llvm::PatternMatch;
 
 #define DEBUG_TYPE "aarch64-lower"
 
 STATISTIC(NumTailCalls, "Number of tail calls");
 STATISTIC(NumShiftInserts, "Number of vector shift inserts");
 STATISTIC(NumOptimizedImms, "Number of times immediates were optimized");
 
-static cl::opt<bool>
-EnableAArch64SlrGeneration("aarch64-shift-insert-generation", cl::Hidden,
-                           cl::desc("Allow AArch64 SLI/SRI formation"),
-                           cl::init(false));
-
 // FIXME: The necessary dtprel relocations don't seem to be supported
 // well in the GNU bfd and gold linkers at the moment. Therefore, by
 // default, for now, fall back to GeneralDynamic code generation.
 cl::opt<bool> EnableAArch64ELFLocalDynamicTLSGeneration(
     "aarch64-elf-ldtls-generation", cl::Hidden,
     cl::desc("Allow AArch64 Local Dynamic TLS code generation"),
     cl::init(false));
 
const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
   case AArch64ISD::TRN2:              return "AArch64ISD::TRN2";
   case AArch64ISD::REV16:             return "AArch64ISD::REV16";
   case AArch64ISD::REV32:             return "AArch64ISD::REV32";
   case AArch64ISD::REV64:             return "AArch64ISD::REV64";
   case AArch64ISD::EXT:               return "AArch64ISD::EXT";
   case AArch64ISD::VSHL:              return "AArch64ISD::VSHL";
   case AArch64ISD::VLSHR:             return "AArch64ISD::VLSHR";
   case AArch64ISD::VASHR:             return "AArch64ISD::VASHR";
+  case AArch64ISD::VSLI:              return "AArch64ISD::VSLI";
+  case AArch64ISD::VSRI:              return "AArch64ISD::VSRI";
   case AArch64ISD::CMEQ:              return "AArch64ISD::CMEQ";
   case AArch64ISD::CMGE:              return "AArch64ISD::CMGE";
   case AArch64ISD::CMGT:              return "AArch64ISD::CMGT";
   case AArch64ISD::CMHI:              return "AArch64ISD::CMHI";
   case AArch64ISD::CMHS:              return "AArch64ISD::CMHS";
   case AArch64ISD::FCMEQ:             return "AArch64ISD::FCMEQ";
   case AArch64ISD::FCMGE:             return "AArch64ISD::FCMGE";
   case AArch64ISD::FCMGT:             return "AArch64ISD::FCMGT";
case Intrinsic::eh_recoverfp: {
     SDValue IncomingFPOp = Op.getOperand(2);
     GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
     auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
     if (!Fn)
       report_fatal_error(
           "llvm.eh.recoverfp must take a function as the first argument");
     return IncomingFPOp;
   }
+
+  case Intrinsic::aarch64_neon_vsri:
+  case Intrinsic::aarch64_neon_vsli: {
+    EVT Ty = Op.getValueType();
+
+    if (!Ty.isVector())
+      report_fatal_error("Unexpected type for aarch64_neon_vsli");
+
+    uint64_t ShiftAmount = Op.getConstantOperandVal(3);
+    unsigned ElemSizeInBits = Ty.getScalarSizeInBits();
+    assert(ShiftAmount <= ElemSizeInBits);
+
+    bool IsShiftRight = IntNo == Intrinsic::aarch64_neon_vsri;
+    unsigned Opcode = IsShiftRight ? AArch64ISD::VSRI : AArch64ISD::VSLI;
+    return DAG.getNode(Opcode, dl, Ty, Op.getOperand(1), Op.getOperand(2),
+                       Op.getOperand(3));
+  }
   }
 }
 
 bool AArch64TargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
   return ExtVal.getValueType().isScalableVector();
 }
 
 // Custom lower trunc store for v4i8 vectors, since it is promoted to v4i16.
if (IID < Intrinsic::num_intrinsics)
       return IID;
     return Intrinsic::not_intrinsic;
   }
   }
 }
 
 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
-// BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
-// Also, logical shift right -> sri, with the same structure.
+// BUILD_VECTORs with constant element C1, C2 is a constant, and:
+//   - for the SLI case: C1 == Ones(ElemSizeInBits) >> (ElemSizeInBits - C2)
+//   - for the SRI case: C1 == Ones(ElemSizeInBits) << (ElemSizeInBits - C2)
 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
   EVT VT = N->getValueType(0);
 
   if (!VT.isVector())
     return SDValue();
 
   SDLoc DL(N);
 
static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
   if (!C2node)
     return SDValue();
 
   // Is the and mask vector all constant?
   uint64_t C1;
   if (!isAllConstantBuildVector(And.getOperand(1), C1))
     return SDValue();
 
-  // Is C1 == ~C2, taking into account how much one can shift elements of a
-  // particular size?
+  // Is C1 == Ones(ElemSizeInBits) << (ElemSizeInBits - C2) or
+  // C1 == Ones(ElemSizeInBits) >> (ElemSizeInBits - C2), taking into account
+  // how much one can shift elements of a particular size?
   uint64_t C2 = C2node->getZExtValue();
   unsigned ElemSizeInBits = VT.getScalarSizeInBits();
   if (C2 > ElemSizeInBits)
     return SDValue();
   unsigned ElemMask = (1 << ElemSizeInBits) - 1;
-  if ((C1 & ElemMask) != (~C2 & ElemMask))
+  if (IsShiftRight) {
+    if ((C1 & ElemMask) != ((ElemMask << (ElemSizeInBits - C2)) & ElemMask))
+      return SDValue();
+  } else {
+    if ((C1 & ElemMask) != ((ElemMask >> (ElemSizeInBits - C2)) & ElemMask))
       return SDValue();
+  }
 
   SDValue X = And.getOperand(0);
   SDValue Y = Shift.getOperand(0);
 
-  unsigned Intrin =
-      IsShiftRight ? Intrinsic::aarch64_neon_vsri : Intrinsic::aarch64_neon_vsli;
-  SDValue ResultSLI =
-      DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
-                  DAG.getConstant(Intrin, DL, MVT::i32), X, Y,
-                  Shift.getOperand(1));
+  unsigned Inst = IsShiftRight ? AArch64ISD::VSRI : AArch64ISD::VSLI;
+  SDValue ResultSLI = DAG.getNode(Inst, DL, VT, X, Y, Shift.getOperand(1));
 
   LLVM_DEBUG(dbgs() << "aarch64-lower: transformed: \n");
   LLVM_DEBUG(N->dump(&DAG));
   LLVM_DEBUG(dbgs() << "into: \n");
   LLVM_DEBUG(ResultSLI->dump(&DAG));
 
   ++NumShiftInserts;
   return ResultSLI;
 }
 
 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
                                              SelectionDAG &DAG) const {
   // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
-  if (EnableAArch64SlrGeneration) {
-    if (SDValue Res = tryLowerToSLI(Op.getNode(), DAG))
-      return Res;
-  }
 
   EVT VT = Op.getValueType();
 
   SDValue LHS = Op.getOperand(0);
   BuildVectorSDNode *BVN =
       dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
   if (!BVN) {
     // OR commutes, so try swapping the operands.

llvm/lib/Target/AArch64/AArch64InstrInfo.td

 def SDT_AArch64MOVIshift : SDTypeProfile<1, 2, [SDTCisInt<1>, SDTCisInt<2>]>;
 def SDT_AArch64vecimm : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
                                            SDTCisInt<2>, SDTCisInt<3>]>;
 def SDT_AArch64UnaryVec: SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisSameAs<0,1>]>;
 def SDT_AArch64ExtVec: SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
                                           SDTCisSameAs<0,2>, SDTCisInt<3>]>;
 def SDT_AArch64vshift : SDTypeProfile<1, 2, [SDTCisSameAs<0,1>, SDTCisInt<2>]>;
 
+def SDT_AArch64vshiftinsert : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisInt<3>,
+                                                 SDTCisSameAs<0,1>,
+                                                 SDTCisSameAs<0,2>]>;
+
 def SDT_AArch64unvec : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisSameAs<0,1>]>;
 def SDT_AArch64fcmpz : SDTypeProfile<1, 1, []>;
 def SDT_AArch64fcmp  : SDTypeProfile<1, 2, [SDTCisSameAs<1,2>]>;
 def SDT_AArch64binvec : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
                                            SDTCisSameAs<0,2>]>;
 def SDT_AArch64trivec : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
                                            SDTCisSameAs<0,2>,
                                            SDTCisSameAs<0,3>]>;

 def AArch64vashr : SDNode<"AArch64ISD::VASHR", SDT_AArch64vshift>;
 def AArch64vlshr : SDNode<"AArch64ISD::VLSHR", SDT_AArch64vshift>;
 def AArch64vshl : SDNode<"AArch64ISD::VSHL", SDT_AArch64vshift>;
 def AArch64sqshli : SDNode<"AArch64ISD::SQSHL_I", SDT_AArch64vshift>;
 def AArch64uqshli : SDNode<"AArch64ISD::UQSHL_I", SDT_AArch64vshift>;
 def AArch64sqshlui : SDNode<"AArch64ISD::SQSHLU_I", SDT_AArch64vshift>;
 def AArch64srshri : SDNode<"AArch64ISD::SRSHR_I", SDT_AArch64vshift>;
 def AArch64urshri : SDNode<"AArch64ISD::URSHR_I", SDT_AArch64vshift>;
+def AArch64vsli : SDNode<"AArch64ISD::VSLI", SDT_AArch64vshiftinsert>;
+def AArch64vsri : SDNode<"AArch64ISD::VSRI", SDT_AArch64vshiftinsert>;
 
 def AArch64not: SDNode<"AArch64ISD::NOT", SDT_AArch64unvec>;
 def AArch64bit: SDNode<"AArch64ISD::BIT", SDT_AArch64trivec>;
 def AArch64bsp: SDNode<"AArch64ISD::BSP", SDT_AArch64trivec>;
 
 def AArch64cmeq: SDNode<"AArch64ISD::CMEQ", SDT_AArch64binvec>;
 def AArch64cmge: SDNode<"AArch64ISD::CMGE", SDT_AArch64binvec>;
 def AArch64cmgt: SDNode<"AArch64ISD::CMGT", SDT_AArch64binvec>;

 defm FCVTZU:SIMDVectorRShiftSD<1, 0b11111, "fcvtzu", int_aarch64_neon_vcvtfp2fxu>;
 defm SCVTF: SIMDVectorRShiftToFP<0, 0b11100, "scvtf",
                                    int_aarch64_neon_vcvtfxs2fp>;
 defm RSHRN   : SIMDVectorRShiftNarrowBHS<0, 0b10001, "rshrn",
                                          int_aarch64_neon_rshrn>;
 defm SHL     : SIMDVectorLShiftBHSD<0, 0b01010, "shl", AArch64vshl>;
 defm SHRN    : SIMDVectorRShiftNarrowBHS<0, 0b10000, "shrn",
                           BinOpFrag<(trunc (AArch64vashr node:$LHS, node:$RHS))>>;
-defm SLI     : SIMDVectorLShiftBHSDTied<1, 0b01010, "sli", int_aarch64_neon_vsli>;
-def : Pat<(v1i64 (int_aarch64_neon_vsli (v1i64 FPR64:$Rd), (v1i64 FPR64:$Rn),
+defm SLI     : SIMDVectorLShiftBHSDTied<1, 0b01010, "sli", AArch64vsli>;
+def : Pat<(v1i64 (AArch64vsli (v1i64 FPR64:$Rd), (v1i64 FPR64:$Rn),
                                       (i32 vecshiftL64:$imm))),
           (SLId FPR64:$Rd, FPR64:$Rn, vecshiftL64:$imm)>;
 defm SQRSHRN : SIMDVectorRShiftNarrowBHS<0, 0b10011, "sqrshrn",
                                          int_aarch64_neon_sqrshrn>;
 defm SQRSHRUN: SIMDVectorRShiftNarrowBHS<1, 0b10001, "sqrshrun",
                                          int_aarch64_neon_sqrshrun>;
 defm SQSHLU : SIMDVectorLShiftBHSD<1, 0b01100, "sqshlu", AArch64sqshlui>;
 defm SQSHL  : SIMDVectorLShiftBHSD<0, 0b01110, "sqshl", AArch64sqshli>;
 defm SQSHRN  : SIMDVectorRShiftNarrowBHS<0, 0b10010, "sqshrn",
                                          int_aarch64_neon_sqshrn>;
 defm SQSHRUN : SIMDVectorRShiftNarrowBHS<1, 0b10000, "sqshrun",
                                          int_aarch64_neon_sqshrun>;
-defm SRI     : SIMDVectorRShiftBHSDTied<1, 0b01000, "sri", int_aarch64_neon_vsri>;
-def : Pat<(v1i64 (int_aarch64_neon_vsri (v1i64 FPR64:$Rd), (v1i64 FPR64:$Rn),
+defm SRI     : SIMDVectorRShiftBHSDTied<1, 0b01000, "sri", AArch64vsri>;
+def : Pat<(v1i64 (AArch64vsri (v1i64 FPR64:$Rd), (v1i64 FPR64:$Rn),
                                       (i32 vecshiftR64:$imm))),
           (SRId FPR64:$Rd, FPR64:$Rn, vecshiftR64:$imm)>;
 defm SRSHR   : SIMDVectorRShiftBHSD<0, 0b00100, "srshr", AArch64srshri>;
 defm SRSRA   : SIMDVectorRShiftBHSDTied<0, 0b00110, "srsra",
                  TriOpFrag<(add node:$LHS,
                                 (AArch64srshri node:$MHS, node:$RHS))> >;
 defm SSHLL   : SIMDVectorLShiftLongBHSD<0, 0b10100, "sshll",
                 BinOpFrag<(AArch64vshl (sext node:$LHS), node:$RHS)>>;

llvm/test/CodeGen/AArch64/arm64-sli-sri-opt.ll

-; RUN: llc < %s -aarch64-shift-insert-generation=true -mtriple=arm64-eabi -aarch64-neon-syntax=apple | FileCheck %s
+; RUN: llc < %s -mtriple=arm64-eabi -aarch64-neon-syntax=apple | FileCheck %s
 
 define void @testLeftGood(<16 x i8> %src1, <16 x i8> %src2, <16 x i8>* %dest) nounwind {
 ; CHECK-LABEL: testLeftGood:
 ; CHECK: sli.16b v0, v1, #3
-  %and.i = and <16 x i8> %src1, <i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252>
+  %and.i = and <16 x i8> %src1, <i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7>
   %vshl_n = shl <16 x i8> %src2, <i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3>
   %result = or <16 x i8> %and.i, %vshl_n
   store <16 x i8> %result, <16 x i8>* %dest, align 16
   ret void
 }
 
 define void @testLeftBad(<16 x i8> %src1, <16 x i8> %src2, <16 x i8>* %dest) nounwind {
 ; CHECK-LABEL: testLeftBad:
 ; CHECK-NOT: sli
   %and.i = and <16 x i8> %src1, <i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165>
   %vshl_n = shl <16 x i8> %src2, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
   %result = or <16 x i8> %and.i, %vshl_n
   store <16 x i8> %result, <16 x i8>* %dest, align 16
   ret void
 }
 
 define void @testRightGood(<16 x i8> %src1, <16 x i8> %src2, <16 x i8>* %dest) nounwind {
 ; CHECK-LABEL: testRightGood:
 ; CHECK: sri.16b v0, v1, #3
-  %and.i = and <16 x i8> %src1, <i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252, i8 252>
+  %and.i = and <16 x i8> %src1, <i8 224, i8 224, i8 224, i8 224, i8 224, i8 224, i8 224, i8 224, i8 224, i8 224, i8 224, i8 224, i8 224, i8 224, i8 224, i8 224>
   %vshl_n = lshr <16 x i8> %src2, <i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3>
   %result = or <16 x i8> %and.i, %vshl_n
   store <16 x i8> %result, <16 x i8>* %dest, align 16
   ret void
 }
 
 define void @testRightBad(<16 x i8> %src1, <16 x i8> %src2, <16 x i8>* %dest) nounwind {
 ; CHECK-LABEL: testRightBad:
 ; CHECK-NOT: sri
   %and.i = and <16 x i8> %src1, <i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165, i8 165>
   %vshl_n = lshr <16 x i8> %src2, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>
   %result = or <16 x i8> %and.i, %vshl_n
   store <16 x i8> %result, <16 x i8>* %dest, align 16
   ret void
 }