Skip to content

Commit 6f13753

Browse files
ghehgghehg
committed
Lower neon_vpadal_v, neon_vpadalq_v
1 parent a223ec2 commit 6f13753

File tree

2 files changed

+208
-1
lines changed

2 files changed

+208
-1
lines changed

‎clang/lib/CIR/CodeGen/CIRGenBuiltinAArch64.cpp‎

Lines changed: 8 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -3868,7 +3868,14 @@ CIRGenFunction::emitAArch64BuiltinExpr(unsigned BuiltinID, const CallExpr *E,
38683868
}
38693869
case NEON::BI__builtin_neon_vpadal_v:
38703870
case NEON::BI__builtin_neon_vpadalq_v: {
3871-
llvm_unreachable("NEON::BI__builtin_neon_vpadalq_v NYI");
3871+
cir::VectorType argTy = getHalfEltSizeTwiceNumElemsVecType(builder, vTy);
3872+
mlir::Location loc = getLoc(E->getExprLoc());
3873+
llvm::SmallVector<mlir::Value, 1> args = {Ops[1]};
3874+
mlir::Value tmp = emitNeonCall(
3875+
builder, {argTy}, args,
3876+
usgn ? "aarch64.neon.uaddlp" : "aarch64.neon.saddlp", vTy, loc);
3877+
mlir::Value addEnd = builder.createBitcast(Ops[0], vTy);
3878+
return builder.createAdd(tmp, addEnd);
38723879
}
38733880
case NEON::BI__builtin_neon_vpmin_v:
38743881
case NEON::BI__builtin_neon_vpminq_v:

‎clang/test/CIR/CodeGen/AArch64/neon-misc.c‎

Lines changed: 200 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1516,3 +1516,203 @@ uint64x2_t test_splatq_laneq_u64(uint64x2_t v) {
15161516
// LLVM: [[RES:%.*]] = shufflevector <2 x i64> [[TMP1]], <2 x i64> [[TMP1]], <2 x i32> zeroinitializer
15171517
// LLVM: ret <2 x i64> [[RES]]
15181518
}
1519+
1520+
int16x4_t test_vpadal_s8(int16x4_t a, int8x8_t b) {
1521+
return vpadal_s8(a, b);
1522+
1523+
// CIR-LABEL: vpadal_s8
1524+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.saddlp" {{%.*}} :
1525+
// CIR-SAME: (!cir.vector<!s8i x 8>) -> !cir.vector<!s16i x 4>
1526+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 8>), !cir.vector<!s16i x 4>
1527+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!s16i x 4>
1528+
1529+
// LLVM: {{.*}}test_vpadal_s8(<4 x i16>{{.*}}[[a:%.*]], <8 x i8>{{.*}}[[b:%.*]])
1530+
// LLVM: [[TMP0:%.*]] = bitcast <4 x i16> [[a]] to <8 x i8>
1531+
// LLVM: [[VPADAL_I:%.*]] = call <4 x i16> @llvm.aarch64.neon.saddlp.v4i16.v8i8(<8 x i8> [[b]])
1532+
// LLVM: [[TMP1:%.*]] = add <4 x i16> [[VPADAL_I]], [[a]]
1533+
// LLVM: ret <4 x i16> [[TMP1]]
1534+
}
1535+
1536+
int32x2_t test_vpadal_s16(int32x2_t a, int16x4_t b) {
1537+
return vpadal_s16(a, b);
1538+
1539+
// CIR-LABEL: vpadal_s16
1540+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.saddlp" {{%.*}} :
1541+
// CIR-SAME: (!cir.vector<!s16i x 4>) -> !cir.vector<!s32i x 2>
1542+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 8>), !cir.vector<!s32i x 2>
1543+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!s32i x 2>
1544+
1545+
// LLVM: {{.*}}test_vpadal_s16(<2 x i32>{{.*}}[[a:%.*]], <4 x i16>{{.*}}[[b:%.*]])
1546+
// LLVM: [[TMP0:%.*]] = bitcast <2 x i32> [[a]] to <8 x i8>
1547+
// LLVM: [[TMP1:%.*]] = bitcast <4 x i16> [[b]] to <8 x i8>
1548+
// LLVM: [[VPADAL1_I:%.*]] = call <2 x i32> @llvm.aarch64.neon.saddlp.v2i32.v4i16(<4 x i16> [[b]])
1549+
// LLVM: [[TMP2:%.*]] = add <2 x i32> [[VPADAL1_I]], [[a]]
1550+
// LLVM: ret <2 x i32> [[TMP2]]
1551+
}
1552+
1553+
int64x1_t test_vpadal_s32(int64x1_t a, int32x2_t b) {
1554+
return vpadal_s32(a, b);
1555+
1556+
// CIR-LABEL: vpadal_s32
1557+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.saddlp" {{%.*}} :
1558+
// CIR-SAME: (!cir.vector<!s32i x 2>) -> !cir.vector<!s64i x 1>
1559+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 8>), !cir.vector<!s64i x 1>
1560+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!s64i x 1>
1561+
1562+
// LLVM: {{.*}}test_vpadal_s32(<1 x i64>{{.*}}[[a:%.*]], <2 x i32>{{.*}}[[b:%.*]])
1563+
// LLVM: [[TMP0:%.*]] = bitcast <1 x i64> [[a]] to <8 x i8>
1564+
// LLVM: [[TMP1:%.*]] = bitcast <2 x i32> [[b]] to <8 x i8>
1565+
// LLVM: [[VPADAL1_I:%.*]] = call <1 x i64> @llvm.aarch64.neon.saddlp.v1i64.v2i32(<2 x i32> [[b]])
1566+
// LLVM: [[TMP2:%.*]] = add <1 x i64> [[VPADAL1_I]], [[a]]
1567+
// LLVM: ret <1 x i64> [[TMP2]]
1568+
}
1569+
1570+
uint16x4_t test_vpadal_u8(uint16x4_t a, uint8x8_t b) {
1571+
return vpadal_u8(a, b);
1572+
1573+
// CIR-LABEL: vpadal_u8
1574+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.uaddlp" {{%.*}} :
1575+
// CIR-SAME: (!cir.vector<!u8i x 8>) -> !cir.vector<!u16i x 4>
1576+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 8>), !cir.vector<!u16i x 4>
1577+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!u16i x 4>
1578+
1579+
// LLVM: {{.*}}test_vpadal_u8(<4 x i16>{{.*}}[[a:%.*]], <8 x i8>{{.*}}[[b:%.*]])
1580+
// LLVM: [[TMP0:%.*]] = bitcast <4 x i16> [[a]] to <8 x i8>
1581+
// LLVM: [[VPADAL_I:%.*]] = call <4 x i16> @llvm.aarch64.neon.uaddlp.v4i16.v8i8(<8 x i8> [[b]])
1582+
// LLVM: [[TMP1:%.*]] = add <4 x i16> [[VPADAL_I]], [[a]]
1583+
// LLVM: ret <4 x i16> [[TMP1]]
1584+
}
1585+
1586+
uint32x2_t test_vpadal_u16(uint32x2_t a, uint16x4_t b) {
1587+
return vpadal_u16(a, b);
1588+
1589+
// CIR-LABEL: vpadal_u16
1590+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.uaddlp" {{%.*}} :
1591+
// CIR-SAME: (!cir.vector<!u16i x 4>) -> !cir.vector<!u32i x 2>
1592+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 8>), !cir.vector<!u32i x 2>
1593+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!u32i x 2>
1594+
1595+
// LLVM: {{.*}}test_vpadal_u16(<2 x i32>{{.*}}[[a:%.*]], <4 x i16>{{.*}}[[b:%.*]])
1596+
// LLVM: [[TMP0:%.*]] = bitcast <2 x i32> [[a]] to <8 x i8>
1597+
// LLVM: [[TMP1:%.*]] = bitcast <4 x i16> [[b]] to <8 x i8>
1598+
// LLVM: [[VPADAL1_I:%.*]] = call <2 x i32> @llvm.aarch64.neon.uaddlp.v2i32.v4i16(<4 x i16> [[b]])
1599+
// LLVM: [[TMP2:%.*]] = add <2 x i32> [[VPADAL1_I]], [[a]]
1600+
// LLVM: ret <2 x i32> [[TMP2]]
1601+
}
1602+
1603+
uint64x1_t test_vpadal_u32(uint64x1_t a, uint32x2_t b) {
1604+
return vpadal_u32(a, b);
1605+
1606+
// CIR-LABEL: vpadal_u32
1607+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.uaddlp" {{%.*}} :
1608+
// CIR-SAME: (!cir.vector<!u32i x 2>) -> !cir.vector<!u64i x 1>
1609+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 8>), !cir.vector<!u64i x 1>
1610+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!u64i x 1>
1611+
1612+
// LLVM: {{.*}}test_vpadal_u32(<1 x i64>{{.*}}[[a:%.*]], <2 x i32>{{.*}}[[b:%.*]])
1613+
// LLVM: [[TMP0:%.*]] = bitcast <1 x i64> [[a]] to <8 x i8>
1614+
// LLVM: [[TMP1:%.*]] = bitcast <2 x i32> [[b]] to <8 x i8>
1615+
// LLVM: [[VPADAL1_I:%.*]] = call <1 x i64> @llvm.aarch64.neon.uaddlp.v1i64.v2i32(<2 x i32> [[b]])
1616+
// LLVM: [[TMP2:%.*]] = add <1 x i64> [[VPADAL1_I]], [[a]]
1617+
// LLVM: ret <1 x i64> [[TMP2]]
1618+
}
1619+
1620+
int16x8_t test_vpadalq_s8(int16x8_t a, int8x16_t b) {
1621+
return vpadalq_s8(a, b);
1622+
1623+
// CIR-LABEL: vpadalq_s8
1624+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.saddlp" {{%.*}} :
1625+
// CIR-SAME: (!cir.vector<!s8i x 16>) -> !cir.vector<!s16i x 8>
1626+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 16>), !cir.vector<!s16i x 8>
1627+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!s16i x 8>
1628+
1629+
// LLVM: {{.*}}test_vpadalq_s8(<8 x i16>{{.*}}[[a:%.*]], <16 x i8>{{.*}}[[b:%.*]])
1630+
// LLVM: [[TMP0:%.*]] = bitcast <8 x i16> [[a]] to <16 x i8>
1631+
// LLVM: [[VPADAL_I:%.*]] = call <8 x i16> @llvm.aarch64.neon.saddlp.v8i16.v16i8(<16 x i8> [[b]])
1632+
// LLVM: [[TMP1:%.*]] = add <8 x i16> [[VPADAL_I]], [[a]]
1633+
// LLVM: ret <8 x i16> [[TMP1]]
1634+
}
1635+
1636+
int32x4_t test_vpadalq_s16(int32x4_t a, int16x8_t b) {
1637+
return vpadalq_s16(a, b);
1638+
1639+
// CIR-LABEL: vpadalq_s16
1640+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.saddlp" {{%.*}} :
1641+
// CIR-SAME: (!cir.vector<!s16i x 8>) -> !cir.vector<!s32i x 4>
1642+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 16>), !cir.vector<!s32i x 4>
1643+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!s32i x 4>
1644+
1645+
// LLVM: {{.*}}test_vpadalq_s16(<4 x i32>{{.*}}[[a:%.*]], <8 x i16>{{.*}}[[b:%.*]])
1646+
// LLVM: [[TMP0:%.*]] = bitcast <4 x i32> [[a]] to <16 x i8>
1647+
// LLVM: [[TMP1:%.*]] = bitcast <8 x i16> [[b]] to <16 x i8>
1648+
// LLVM: [[VPADAL1_I:%.*]] = call <4 x i32> @llvm.aarch64.neon.saddlp.v4i32.v8i16(<8 x i16> [[b]])
1649+
// LLVM: [[TMP2:%.*]] = add <4 x i32> [[VPADAL1_I]], [[a]]
1650+
// LLVM: ret <4 x i32> [[TMP2]]
1651+
}
1652+
1653+
int64x2_t test_vpadalq_s32(int64x2_t a, int32x4_t b) {
1654+
return vpadalq_s32(a, b);
1655+
1656+
// CIR-LABEL: vpadalq_s32
1657+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.saddlp" {{%.*}} :
1658+
// CIR-SAME: (!cir.vector<!s32i x 4>) -> !cir.vector<!s64i x 2>
1659+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 16>), !cir.vector<!s64i x 2>
1660+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!s64i x 2>
1661+
1662+
// LLVM: {{.*}}test_vpadalq_s32(<2 x i64>{{.*}}[[a:%.*]], <4 x i32>{{.*}}[[b:%.*]])
1663+
// LLVM: [[TMP0:%.*]] = bitcast <2 x i64> [[a]] to <16 x i8>
1664+
// LLVM: [[TMP1:%.*]] = bitcast <4 x i32> [[b]] to <16 x i8>
1665+
// LLVM: [[VPADAL1_I:%.*]] = call <2 x i64> @llvm.aarch64.neon.saddlp.v2i64.v4i32(<4 x i32> [[b]])
1666+
// LLVM: [[TMP2:%.*]] = add <2 x i64> [[VPADAL1_I]], [[a]]
1667+
// LLVM: ret <2 x i64> [[TMP2]]
1668+
}
1669+
1670+
uint16x8_t test_vpadalq_u8(uint16x8_t a, uint8x16_t b) {
1671+
return vpadalq_u8(a, b);
1672+
1673+
// CIR-LABEL: vpadalq_u8
1674+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.uaddlp" {{%.*}} :
1675+
// CIR-SAME: (!cir.vector<!u8i x 16>) -> !cir.vector<!u16i x 8>
1676+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 16>), !cir.vector<!u16i x 8>
1677+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!u16i x 8>
1678+
1679+
// LLVM: {{.*}}test_vpadalq_u8(<8 x i16>{{.*}}[[a:%.*]], <16 x i8>{{.*}}[[b:%.*]])
1680+
// LLVM: [[TMP0:%.*]] = bitcast <8 x i16> [[a]] to <16 x i8>
1681+
// LLVM: [[VPADAL_I:%.*]] = call <8 x i16> @llvm.aarch64.neon.uaddlp.v8i16.v16i8(<16 x i8> [[b]])
1682+
// LLVM: [[TMP1:%.*]] = add <8 x i16> [[VPADAL_I]], [[a]]
1683+
// LLVM: ret <8 x i16> [[TMP1]]
1684+
}
1685+
1686+
uint32x4_t test_vpadalq_u16(uint32x4_t a, uint16x8_t b) {
1687+
return vpadalq_u16(a, b);
1688+
1689+
// CIR-LABEL: vpadalq_u16
1690+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.uaddlp" {{%.*}} :
1691+
// CIR-SAME: (!cir.vector<!u16i x 8>) -> !cir.vector<!u32i x 4>
1692+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 16>), !cir.vector<!u32i x 4>
1693+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!u32i x 4>
1694+
1695+
// LLVM: {{.*}}test_vpadalq_u16(<4 x i32>{{.*}}[[a:%.*]], <8 x i16>{{.*}}[[b:%.*]])
1696+
// LLVM: [[TMP0:%.*]] = bitcast <4 x i32> [[a]] to <16 x i8>
1697+
// LLVM: [[TMP1:%.*]] = bitcast <8 x i16> [[b]] to <16 x i8>
1698+
// LLVM: [[VPADAL1_I:%.*]] = call <4 x i32> @llvm.aarch64.neon.uaddlp.v4i32.v8i16(<8 x i16> [[b]])
1699+
// LLVM: [[TMP2:%.*]] = add <4 x i32> [[VPADAL1_I]], [[a]]
1700+
// LLVM: ret <4 x i32> [[TMP2]]
1701+
}
1702+
1703+
uint64x2_t test_vpadalq_u32(uint64x2_t a, uint32x4_t b) {
1704+
return vpadalq_u32(a, b);
1705+
1706+
// CIR-LABEL: vpadalq_u32
1707+
// CIR: [[VPADAL_I:%.*]] = cir.llvm.intrinsic "aarch64.neon.uaddlp" {{%.*}} :
1708+
// CIR-SAME: (!cir.vector<!u32i x 4>) -> !cir.vector<!u64i x 2>
1709+
// CIR: [[a:%.*]] = cir.cast(bitcast, {{%.*}} : !cir.vector<!s8i x 16>), !cir.vector<!u64i x 2>
1710+
// CIR: {{%.*}} = cir.binop(add, [[VPADAL_I]], [[a]]) : !cir.vector<!u64i x 2>
1711+
1712+
// LLVM: {{.*}}test_vpadalq_u32(<2 x i64>{{.*}}[[a:%.*]], <4 x i32>{{.*}}[[b:%.*]])
1713+
// LLVM: [[TMP0:%.*]] = bitcast <2 x i64> [[a]] to <16 x i8>
1714+
// LLVM: [[TMP1:%.*]] = bitcast <4 x i32> [[b]] to <16 x i8>
1715+
// LLVM: [[VPADAL1_I:%.*]] = call <2 x i64> @llvm.aarch64.neon.uaddlp.v2i64.v4i32(<4 x i32> [[b]])
1716+
// LLVM: [[TMP2:%.*]] = add <2 x i64> [[VPADAL1_I]], [[a]]
1717+
// LLVM: ret <2 x i64> [[TMP2]]
1718+
}

0 commit comments

Comments
 (0)