@@ -28412,6 +28412,43 @@ SDValue X86TargetLowering::LowerRESET_FPENV(SDValue Op,
28412
28412
return createSetFPEnvNodes(Env, Chain, DL, MVT::i32, MMO, DAG, Subtarget);
28413
28413
}
28414
28414
28415
+ // Generate a GFNI gf2p8affine bitmask for vXi8 bitreverse/shift/rotate.
28416
+ uint64_t getGFNICtrlImm(unsigned Opcode, unsigned Amt = 0) {
28417
+ assert((Amt < 8) && "Shift/Rotation amount out of range");
28418
+ switch (Opcode) {
28419
+ case ISD::BITREVERSE:
28420
+ return 0x8040201008040201ULL;
28421
+ case ISD::SHL:
28422
+ return ((0x0102040810204080ULL >> (Amt)) &
28423
+ (0x0101010101010101ULL * (0xFF >> (Amt))));
28424
+ case ISD::SRL:
28425
+ return ((0x0102040810204080ULL << (Amt)) &
28426
+ (0x0101010101010101ULL * ((0xFF << (Amt)) & 0xFF)));
28427
+ case ISD::SRA:
28428
+ return (getGFNICtrlImm(ISD::SRL, Amt) |
28429
+ (0x8080808080808080ULL >> (64 - (8 * Amt))));
28430
+ case ISD::ROTL:
28431
+ return getGFNICtrlImm(ISD::SRL, 8 - Amt) | getGFNICtrlImm(ISD::SHL, Amt);
28432
+ case ISD::ROTR:
28433
+ return getGFNICtrlImm(ISD::SHL, 8 - Amt) | getGFNICtrlImm(ISD::SRL, Amt);
28434
+ }
28435
+ llvm_unreachable("Unsupported GFNI opcode");
28436
+ }
28437
+
28438
+ // Generate a GFNI gf2p8affine bitmask for vXi8 bitreverse/shift/rotate.
28439
+ SDValue getGFNICtrlMask(unsigned Opcode, SelectionDAG &DAG, const SDLoc &DL,
28440
+ MVT VT, unsigned Amt = 0) {
28441
+ assert(VT.getVectorElementType() == MVT::i8 &&
28442
+ (VT.getSizeInBits() % 64) == 0 && "Illegal GFNI control type");
28443
+ uint64_t Imm = getGFNICtrlImm(Opcode, Amt);
28444
+ SmallVector<SDValue> MaskBits;
28445
+ for (unsigned I = 0, E = VT.getSizeInBits(); I != E; I += 8) {
28446
+ uint64_t Bits = (Imm >> (I % 64)) & 255;
28447
+ MaskBits.push_back(DAG.getConstant(Bits, DL, MVT::i8));
28448
+ }
28449
+ return DAG.getBuildVector(VT, DL, MaskBits);
28450
+ }
28451
+
28415
28452
/// Lower a vector CTLZ using native supported vector CTLZ instruction.
28416
28453
//
28417
28454
// i8/i16 vector implemented using dword LZCNT vector instruction
@@ -29597,43 +29634,6 @@ SDValue X86TargetLowering::LowerWin64_INT128_TO_FP(SDValue Op,
29597
29634
return IsStrict ? DAG.getMergeValues({Result, Chain}, dl) : Result;
29598
29635
}
29599
29636
29600
- // Generate a GFNI gf2p8affine bitmask for vXi8 bitreverse/shift/rotate.
29601
- uint64_t getGFNICtrlImm(unsigned Opcode, unsigned Amt = 0) {
29602
- assert((Amt < 8) && "Shift/Rotation amount out of range");
29603
- switch (Opcode) {
29604
- case ISD::BITREVERSE:
29605
- return 0x8040201008040201ULL;
29606
- case ISD::SHL:
29607
- return ((0x0102040810204080ULL >> (Amt)) &
29608
- (0x0101010101010101ULL * (0xFF >> (Amt))));
29609
- case ISD::SRL:
29610
- return ((0x0102040810204080ULL << (Amt)) &
29611
- (0x0101010101010101ULL * ((0xFF << (Amt)) & 0xFF)));
29612
- case ISD::SRA:
29613
- return (getGFNICtrlImm(ISD::SRL, Amt) |
29614
- (0x8080808080808080ULL >> (64 - (8 * Amt))));
29615
- case ISD::ROTL:
29616
- return getGFNICtrlImm(ISD::SRL, 8 - Amt) | getGFNICtrlImm(ISD::SHL, Amt);
29617
- case ISD::ROTR:
29618
- return getGFNICtrlImm(ISD::SHL, 8 - Amt) | getGFNICtrlImm(ISD::SRL, Amt);
29619
- }
29620
- llvm_unreachable("Unsupported GFNI opcode");
29621
- }
29622
-
29623
- // Generate a GFNI gf2p8affine bitmask for vXi8 bitreverse/shift/rotate.
29624
- SDValue getGFNICtrlMask(unsigned Opcode, SelectionDAG &DAG, const SDLoc &DL, MVT VT,
29625
- unsigned Amt = 0) {
29626
- assert(VT.getVectorElementType() == MVT::i8 &&
29627
- (VT.getSizeInBits() % 64) == 0 && "Illegal GFNI control type");
29628
- uint64_t Imm = getGFNICtrlImm(Opcode, Amt);
29629
- SmallVector<SDValue> MaskBits;
29630
- for (unsigned I = 0, E = VT.getSizeInBits(); I != E; I += 8) {
29631
- uint64_t Bits = (Imm >> (I % 64)) & 255;
29632
- MaskBits.push_back(DAG.getConstant(Bits, DL, MVT::i8));
29633
- }
29634
- return DAG.getBuildVector(VT, DL, MaskBits);
29635
- }
29636
-
29637
29637
// Return true if the required (according to Opcode) shift-imm form is natively
29638
29638
// supported by the Subtarget
29639
29639
static bool supportedVectorShiftWithImm(EVT VT, const X86Subtarget &Subtarget,
0 commit comments