[sve][acle] Add SVE BFloat16 extensions.

[fpetrogalli]

Summary

List of intrinsics:

svfloat32_t svbfdot[_f32](svfloat32_t op1, svbfloat16_t op2, svbfloat16_t op3)
svfloat32_t svbfdot[_n_f32](svfloat32_t op1, svbfloat16_t op2, bfloat16_t op3)
svfloat32_t svbfdot_lane[_f32](svfloat32_t op1, svbfloat16_t op2, svbfloat16_t op3, uint64_t imm_index)

svfloat32_t svbfmmla[_f32](svfloat32_t op1, svbfloat16_t op2, svbfloat16_t op3)

svfloat32_t svbfmlalb[_f32](svfloat32_t op1, svbfloat16_t op2, svbfloat16_t op3)
svfloat32_t svbfmlalb[_n_f32](svfloat32_t op1, svbfloat16_t op2, bfloat16_t op3)
svfloat32_t svbfmlalb_lane[_f32](svfloat32_t op1, svbfloat16_t op2, svbfloat16_t op3, uint64_t imm_index)

svfloat32_t svbfmlalt[_f32](svfloat32_t op1, svbfloat16_t op2, svbfloat16_t op3)
svfloat32_t svbfmlalt[_n_f32](svfloat32_t op1, svbfloat16_t op2, bfloat16_t op3)
svfloat32_t svbfmlalt_lane[_f32](svfloat32_t op1, svbfloat16_t op2, svbfloat16_t op3, uint64_t imm_index)

svbfloat16_t svcvt_bf16[_f32]_m(svbfloat16_t inactive, svbool_t pg, svfloat32_t op)
svbfloat16_t svcvt_bf16[_f32]_x(svbool_t pg, svfloat32_t op)
svbfloat16_t svcvt_bf16[_f32]_z(svbool_t pg, svfloat32_t op)

svbfloat16_t svcvtnt_bf16[_f32]_m(svbfloat16_t even, svbool_t pg, svfloat32_t op)
svbfloat16_t svcvtnt_bf16[_f32]_x(svbfloat16_t even, svbool_t pg, svfloat32_t op)

For reference, see section 7.2 of "Arm C Language Extensions for SVE - Version 00bet4"

[sdesmalen]

clang/include/clang/Basic/arm_sve.td

502

The types for these intrinsics are always svfloat32_t and svbfloat16_t, which given their semantics is unlikely to ever be extended to other types, so it's easier to make the LLVM IR non-overloaded (i.e. hardcoding llvm_nxv4f32_ty and llvm_nxv8bf16_ty) and using the IsOverloadNone flag for these builtins. Then you can express this builtin as:

def SVBFDOT: SInst<"svbfdot[_{0}]",  "MMdd", "b", MergeNone, "aarch64_sve_bfdot">;

and drop the need for the $ modifier.

506

Similar to the suggestion above to use "MMdd" for SVBFDOT, this could use "MMda" and you don't need the ~ modifier.

nit: add whitespace above this line.
nit: the rest of this file tries to align the columns, that makes this file a bit easier to read.

1050

nit: redundant comment (same for above)

clang/test/CodeGen/aarch64-sve-intrinsics/acle_sve_bfdot.c

28

Testing the edge cases 0 and 3 should be sufficient. (same for all other cases in this patch)

llvm/include/llvm/IR/IntrinsicsAArch64.td

1343

nit: SVE_bfloat is not very descriptive, maybe use SVE_4Vec_BF16 and SVE_4Vec_BF16_Indexed?

1806

nit: use fcvtbf instead of cvt => int_aarch64_sve_fcvtbf_bf16f32 ?

[fpetrogalli]

Thank you for the review @sdesmalen!

Francesco

llvm/include/llvm/IR/IntrinsicsAArch64.td

1806

Renamed to int_aarch64_sve_fcvt_bf16f32 and int_aarch64_sve_fcvtnt_bf16f32 respectively, because I think it wouldn't make sense to add the bf suffix to the cvtnt version of the intrinsic.

[sdesmalen]

LGTM

llvm/include/llvm/IR/IntrinsicsAArch64.td

1806

I meant to write int_aarch64_sve_bfcvt_bf16f32. This seems consistent with all other intrinsics (fcvt, fcvtzu, scvtf, etc.) that use the name of the instruction directly in the name of the intrinsic.

[sdesmalen]

llvm/include/llvm/IR/IntrinsicsAArch64.td

1345

nit: keep this on one line.

[fpetrogalli]

Formatting changes. NFC.

clang/include/clang/Basic/arm_sve.td

 
 // Load one octoword and replicate (scalar base)
 let ArchGuard = "defined(__ARM_FEATURE_SVE_MATMUL_FP64)" in {
   def SVLD1RO : SInst<"svld1ro[_{2}]", "dPc", "csilUcUsUiUlhfd", MergeNone, "aarch64_sve_ld1ro">;
 }
 let ArchGuard = "defined(__ARM_FEATURE_SVE_MATMUL_FP64) && defined(__ARM_FEATURE_BF16_SCALAR_ARITHMETIC)" in {
   def SVLD1RO_BF : SInst<"svld1ro[_{2}]", "dPc", "b", MergeNone, "aarch64_sve_ld1ro">;
 }
+
+let ArchGuard = "defined(__ARM_FEATURE_SVE_BF16)" in {
+  def SVBFDOT        : SInst<"svbfdot[_{0}]",        "MMdd",  "b", MergeNone, "aarch64_sve_bfdot",        [IsOverloadNone]>;
+  def SVBFMLALB      : SInst<"svbfmlalb[_{0}]",      "MMdd",  "b", MergeNone, "aarch64_sve_bfmlalb",      [IsOverloadNone]>;
+  def SVBFMLALT      : SInst<"svbfmlalt[_{0}]",      "MMdd",  "b", MergeNone, "aarch64_sve_bfmlalt",      [IsOverloadNone]>;
+  def SVBFMMLA       : SInst<"svbfmmla[_{0}]",       "MMdd",  "b", MergeNone, "aarch64_sve_bfmmla",       [IsOverloadNone]>;
+  def SVBFDOT_N      : SInst<"svbfdot[_n_{0}]",      "MMda",  "b", MergeNone, "aarch64_sve_bfdot",        [IsOverloadNone]>;
+  def SVBFMLAL_N     : SInst<"svbfmlalb[_n_{0}]",    "MMda",  "b", MergeNone, "aarch64_sve_bfmlalb",      [IsOverloadNone]>;
+  def SVBFMLALT_N    : SInst<"svbfmlalt[_n_{0}]",    "MMda",  "b", MergeNone, "aarch64_sve_bfmlalt",      [IsOverloadNone]>;
+  def SVBFDOT_LANE   : SInst<"svbfdot_lane[_{0}]",   "MMddn", "b", MergeNone, "aarch64_sve_bfdot_lane",   [IsOverloadNone], [ImmCheck<3, ImmCheck0_3>]>;
+  def SVBFMLALB_LANE : SInst<"svbfmlalb_lane[_{0}]", "MMddn", "b", MergeNone, "aarch64_sve_bfmlalb_lane", [IsOverloadNone], [ImmCheck<3, ImmCheck0_7>]>;
+  def SVBFMLALT_LANE : SInst<"svbfmlalt_lane[_{0}]", "MMddn", "b", MergeNone, "aarch64_sve_bfmlalt_lane", [IsOverloadNone], [ImmCheck<3, ImmCheck0_7>]>;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 // Stores
 
 // Store one vector (scalar base)
 def SVST1    : MInst<"svst1[_{d}]",  "vPpd", "csilUcUsUiUlhfd", [IsStore], MemEltTyDefault, "aarch64_sve_st1">;
 def SVST1B_S : MInst<"svst1b[_{d}]", "vPAd", "sil",             [IsStore], MemEltTyInt8,    "aarch64_sve_st1">;
 def SVST1B_U : MInst<"svst1b[_{d}]", "vPEd", "UsUiUl",          [IsStore], MemEltTyInt8,    "aarch64_sve_st1">;
 def SVST1H_S : MInst<"svst1h[_{d}]", "vPBd", "il",              [IsStore], MemEltTyInt16,   "aarch64_sve_st1">;

 defm SVFCVTZS_S16_F16 : SInstCvtMXZ<"svcvt_s16[_f16]", "ddPO", "dPO", "s",  "aarch64_sve_fcvtzs", [IsOverloadCvt]>;
 defm SVFCVTZS_S32_F16 : SInstCvtMXZ<"svcvt_s32[_f16]", "ddPO", "dPO", "i",  "aarch64_sve_fcvtzs_i32f16">;
 defm SVFCVTZS_S64_F16 : SInstCvtMXZ<"svcvt_s64[_f16]", "ddPO", "dPO", "l",  "aarch64_sve_fcvtzs_i64f16">;
 
 // svcvt_s##_f32
 defm SVFCVTZS_S32_F32 : SInstCvtMXZ<"svcvt_s32[_f32]", "ddPM", "dPM", "i",  "aarch64_sve_fcvtzs", [IsOverloadCvt]>;
 defm SVFCVTZS_S64_F32 : SInstCvtMXZ<"svcvt_s64[_f32]", "ddPM", "dPM", "l",  "aarch64_sve_fcvtzs_i64f32">;
 
+let ArchGuard = "defined(__ARM_FEATURE_SVE_BF16)" in {
+  defm SVCVT_BF16_F32   : SInstCvtMXZ<"svcvt_bf16[_f32]",  "ddPM", "dPM", "b",  "aarch64_sve_fcvt_bf16f32">;
+  defm SVCVTNT_BF16_F32 : SInstCvtMX<"svcvtnt_bf16[_f32]", "ddPM", "dPM", "b",  "aarch64_sve_fcvtnt_bf16f32">;
+}
+
 // svcvt_s##_f64
 defm SVFCVTZS_S32_F64 : SInstCvtMXZ<"svcvt_s32[_f64]", "ttPd", "tPd", "d",  "aarch64_sve_fcvtzs_i32f64">;
 defm SVFCVTZS_S64_F64 : SInstCvtMXZ<"svcvt_s64[_f64]", "ddPN", "dPN", "l",  "aarch64_sve_fcvtzs", [IsOverloadCvt]>;
 
 // svcvt_u##_f16
 defm SVFCVTZU_U16_F16 : SInstCvtMXZ<"svcvt_u16[_f16]", "ddPO", "dPO", "Us", "aarch64_sve_fcvtzu", [IsOverloadCvt]>;
 defm SVFCVTZU_U32_F16 : SInstCvtMXZ<"svcvt_u32[_f16]", "ddPO", "dPO", "Ui", "aarch64_sve_fcvtzu_i32f16">;
 defm SVFCVTZU_U64_F16 : SInstCvtMXZ<"svcvt_u64[_f16]", "ddPO", "dPO", "Ul", "aarch64_sve_fcvtzu_i64f16">;

clang/test/CodeGen/aarch64-sve-intrinsics/acle_sve_bfdot.c

+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -D__ARM_FEATURE_BF16_SCALAR_ARITHMETIC -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +bf16 -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -D__ARM_FEATURE_BF16_SCALAR_ARITHMETIC -DSVE_OVERLOADED_FORMS -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +bf16 -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+
+#include <arm_sve.h>
+
+#ifdef SVE_OVERLOADED_FORMS
+// A simple used,unused... macro, long enough to represent any SVE builtin.
+#define SVE_ACLE_FUNC(A1, A2_UNUSED, A3, A4_UNUSED) A1##A3
+#else
+#define SVE_ACLE_FUNC(A1, A2, A3, A4) A1##A2##A3##A4
+#endif
+
+svfloat32_t test_bfdot_f32(svfloat32_t x, svbfloat16_t y, svbfloat16_t z) {
+  // CHECK-LABEL: @test_bfdot_f32(
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfdot(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %z)
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfdot, _f32, , )(x, y, z);
+}
+
+svfloat32_t test_bfdot_lane_0_f32(svfloat32_t x, svbfloat16_t y, svbfloat16_t z) {
+  // CHECK-LABEL: @test_bfdot_lane_0_f32(
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfdot.lane(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %z, i64 0)
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfdot_lane, _f32, , )(x, y, z, 0);
+}
+
+svfloat32_t test_bfdot_lane_3_f32(svfloat32_t x, svbfloat16_t y, svbfloat16_t z) {
+  // CHECK-LABEL: @test_bfdot_lane_3_f32(
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfdot.lane(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %z, i64 3)
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfdot_lane, _f32, , )(x, y, z, 3);
+}
+
+svfloat32_t test_bfdot_n_f32(svfloat32_t x, svbfloat16_t y, bfloat16_t z) {
+  // CHECK-LABEL: @test_bfdot_n_f32(
+  // CHECK: %[[SPLAT:.*]] = call <vscale x 8 x bfloat> @llvm.aarch64.sve.dup.x.nxv8bf16(bfloat %z)
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfdot(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %[[SPLAT]])
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfdot, _n_f32, , )(x, y, z);
+}

clang/test/CodeGen/aarch64-sve-intrinsics/acle_sve_bfmlalb.c

+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -D__ARM_FEATURE_BF16_SCALAR_ARITHMETIC -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +bf16  -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -D__ARM_FEATURE_BF16_SCALAR_ARITHMETIC -DSVE_OVERLOADED_FORMS -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +bf16 -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+
+#include <arm_sve.h>
+
+#ifdef SVE_OVERLOADED_FORMS
+// A simple used,unused... macro, long enough to represent any SVE builtin.
+#define SVE_ACLE_FUNC(A1, A2_UNUSED, A3, A4_UNUSED) A1##A3
+#else
+#define SVE_ACLE_FUNC(A1, A2, A3, A4) A1##A2##A3##A4
+#endif
+
+svfloat32_t test_svbfmlalb_f32(svfloat32_t x, svbfloat16_t y, svbfloat16_t z) {
+  // CHECK-LABEL: @test_svbfmlalb_f32(
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %z)
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfmlalb, _f32, , )(x, y, z);
+}
+
+svfloat32_t test_bfmlalb_lane_0_f32(svfloat32_t x, svbfloat16_t y, svbfloat16_t z) {
+  // CHECK-LABEL: @test_bfmlalb_lane_0_f32(
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb.lane(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %z, i64 0)
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfmlalb_lane, _f32, , )(x, y, z, 0);
+}
+
+svfloat32_t test_bfmlalb_lane_7_f32(svfloat32_t x, svbfloat16_t y, svbfloat16_t z) {
+  // CHECK-LABEL: @test_bfmlalb_lane_7_f32(
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb.lane(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %z, i64 7)
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfmlalb_lane, _f32, , )(x, y, z, 7);
+}
+
+svfloat32_t test_bfmlalb_n_f32(svfloat32_t x, svbfloat16_t y, bfloat16_t z) {
+  // CHECK-LABEL: @test_bfmlalb_n_f32(
+  // CHECK: %[[SPLAT:.*]] = call <vscale x 8 x bfloat> @llvm.aarch64.sve.dup.x.nxv8bf16(bfloat %z)
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %[[SPLAT]])
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfmlalb, _n_f32, , )(x, y, z);
+}

clang/test/CodeGen/aarch64-sve-intrinsics/acle_sve_bfmlalt.c

+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -D__ARM_FEATURE_BF16_SCALAR_ARITHMETIC -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +bf16 -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -D__ARM_FEATURE_BF16_SCALAR_ARITHMETIC -DSVE_OVERLOADED_FORMS -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +bf16 -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+
+#include <arm_sve.h>
+
+#ifdef SVE_OVERLOADED_FORMS
+// A simple used,unused... macro, long enough to represent any SVE builtin.
+#define SVE_ACLE_FUNC(A1, A2_UNUSED, A3, A4_UNUSED) A1##A3
+#else
+#define SVE_ACLE_FUNC(A1, A2, A3, A4) A1##A2##A3##A4
+#endif
+
+svfloat32_t test_svbfmlalt_f32(svfloat32_t x, svbfloat16_t y, svbfloat16_t z) {
+  // CHECK-LABEL: @test_svbfmlalt_f32(
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %z)
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfmlalt, _f32, , )(x, y, z);
+}
+
+svfloat32_t test_bfmlalt_lane_0_f32(svfloat32_t x, svbfloat16_t y, svbfloat16_t z) {
+  // CHECK-LABEL: @test_bfmlalt_lane_0_f32(
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt.lane(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %z, i64 0)
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfmlalt_lane, _f32, , )(x, y, z, 0);
+}
+
+svfloat32_t test_bfmlalt_lane_7_f32(svfloat32_t x, svbfloat16_t y, svbfloat16_t z) {
+  // CHECK-LABEL: @test_bfmlalt_lane_7_f32(
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt.lane(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %z, i64 7)
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfmlalt_lane, _f32, , )(x, y, z, 7);
+}
+
+svfloat32_t test_bfmlalt_n_f32(svfloat32_t x, svbfloat16_t y, bfloat16_t z) {
+  // CHECK-LABEL: @test_bfmlalt_n_f32(
+  // CHECK: %[[SPLAT:.*]] = call <vscale x 8 x bfloat> @llvm.aarch64.sve.dup.x.nxv8bf16(bfloat %z)
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %[[SPLAT]])
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfmlalt, _n_f32, , )(x, y, z);
+}

clang/test/CodeGen/aarch64-sve-intrinsics/acle_sve_bfmmla.c

+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -triple aarch64-none-linux-gnu -target-feature +sve -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -DSVE_OVERLOADED_FORMS -triple aarch64-none-linux-gnu -target-feature +sve -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+
+#include <arm_sve.h>
+
+#ifdef SVE_OVERLOADED_FORMS
+// A simple used,unused... macro, long enough to represent any SVE builtin.
+#define SVE_ACLE_FUNC(A1, A2_UNUSED, A3, A4_UNUSED) A1##A3
+#else
+#define SVE_ACLE_FUNC(A1, A2, A3, A4) A1##A2##A3##A4
+#endif
+
+svfloat32_t test_bfmmla_f32(svfloat32_t x, svbfloat16_t y, svbfloat16_t z) {
+  // CHECK-LABEL: @test_bfmmla_f32(
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.bfmmla(<vscale x 4 x float> %x, <vscale x 8 x bfloat> %y, <vscale x 8 x bfloat> %z)
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svbfmmla, _f32, , )(x, y, z);
+}

clang/test/CodeGen/aarch64-sve-intrinsics/acle_sve_cvt-bfloat.c

+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -D__ARM_FEATURE_BF16_SCALAR_ARITHMETIC -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +bf16 -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -D__ARM_FEATURE_BF16_SCALAR_ARITHMETIC -DSVE_OVERLOADED_FORMS -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +bf16 -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+
+#include <arm_sve.h>
+
+#ifdef SVE_OVERLOADED_FORMS
+// A simple used,unused... macro, long enough to represent any SVE builtin.
+#define SVE_ACLE_FUNC(A1, A2_UNUSED, A3, A4_UNUSED) A1##A3
+#else
+#define SVE_ACLE_FUNC(A1, A2, A3, A4) A1##A2##A3##A4
+#endif
+
+svbfloat16_t test_svcvt_bf16_f32_x(svbool_t pg, svfloat32_t op) {
+  // CHECK-LABEL: test_svcvt_bf16_f32_x
+  // CHECK: %[[PG:.*]] = call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> %pg)
+  // CHECK: %[[INTRINSIC:.*]] = call <vscale x 8 x bfloat> @llvm.aarch64.sve.fcvt.bf16f32(<vscale x 8 x bfloat> undef, <vscale x 8 x i1> %[[PG]], <vscale x 4 x float> %op)
+  // CHECK: ret <vscale x 8 x bfloat> %[[INTRINSIC]]
+  return SVE_ACLE_FUNC(svcvt_bf16, _f32, _x, )(pg, op);
+}
+
+svbfloat16_t test_svcvt_bf16_f32_z(svbool_t pg, svfloat32_t op) {
+  // CHECK-LABEL: test_svcvt_bf16_f32_z
+  // CHECK: %[[PG:.*]] = call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> %pg)
+  // CHECK: %[[INTRINSIC:.*]] = call <vscale x 8 x bfloat> @llvm.aarch64.sve.fcvt.bf16f32(<vscale x 8 x bfloat> zeroinitializer, <vscale x 8 x i1> %[[PG]], <vscale x 4 x float> %op)
+  // CHECK: ret <vscale x 8 x bfloat> %[[INTRINSIC]]
+  return SVE_ACLE_FUNC(svcvt_bf16, _f32, _z, )(pg, op);
+}
+
+svbfloat16_t test_svcvt_bf16_f32_m(svbfloat16_t inactive, svbool_t pg, svfloat32_t op) {
+  // CHECK-LABEL: test_svcvt_bf16_f32_m
+  // CHECK: %[[PG:.*]] = call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> %pg)
+  // CHECK: %[[INTRINSIC:.*]] = call <vscale x 8 x bfloat> @llvm.aarch64.sve.fcvt.bf16f32(<vscale x 8 x bfloat> %inactive, <vscale x 8 x i1> %[[PG]], <vscale x 4 x float> %op)
+  // CHECK: ret <vscale x 8 x bfloat> %[[INTRINSIC]]
+  return SVE_ACLE_FUNC(svcvt_bf16, _f32, _m, )(inactive, pg, op);
+}

clang/test/CodeGen/aarch64-sve-intrinsics/acle_sve_cvtnt.c

+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -D__ARM_FEATURE_BF16_SCALAR_ARITHMETIC -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +bf16 -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_BF16 -D__ARM_FEATURE_BF16_SCALAR_ARITHMETIC -DSVE_OVERLOADED_FORMS -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +bf16 -fallow-half-arguments-and-returns -S -O1 -Werror -Wall -emit-llvm -o - %s | FileCheck %s
+
+#include <arm_sve.h>
+
+#ifdef SVE_OVERLOADED_FORMS
+// A simple used,unused... macro, long enough to represent any SVE builtin.
+#define SVE_ACLE_FUNC(A1, A2_UNUSED, A3, A4_UNUSED) A1##A3
+#else
+#define SVE_ACLE_FUNC(A1, A2, A3, A4) A1##A2##A3##A4
+#endif
+
+svbfloat16_t test_svcvtnt_bf16_f32_x(svbool_t pg, svfloat32_t op) {
+  // CHECK-LABEL: test_svcvtnt_bf16_f32_x
+  // CHECK: %[[PG:.*]] = call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> %pg)
+  // CHECK: %[[INTRINSIC:.*]] = call <vscale x 8 x bfloat> @llvm.aarch64.sve.fcvtnt.bf16f32(<vscale x 8 x bfloat> undef, <vscale x 8 x i1> %[[PG]], <vscale x 4 x float> %op)
+  // CHECK: ret <vscale x 8 x bfloat> %[[INTRINSIC]]
+  return SVE_ACLE_FUNC(svcvtnt_bf16, _f32, _x, )(pg, op);
+}
+
+svbfloat16_t test_svcvtnt_bf16_f32_m(svbfloat16_t inactive, svbool_t pg, svfloat32_t op) {
+  // CHECK-LABEL: test_svcvtnt_bf16_f32_m
+  // CHECK: %[[PG:.*]] = call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> %pg)
+  // CHECK: %[[INTRINSIC:.*]] = call <vscale x 8 x bfloat> @llvm.aarch64.sve.fcvtnt.bf16f32(<vscale x 8 x bfloat> %inactive, <vscale x 8 x i1> %[[PG]], <vscale x 4 x float> %op)
+  // CHECK: ret <vscale x 8 x bfloat> %[[INTRINSIC]]
+  return SVE_ACLE_FUNC(svcvtnt_bf16, _f32, _m, )(inactive, pg, op);
+}

clang/utils/TableGen/SveEmitter.cpp

public:
   bool isSigned() const { return Signed; }
   bool isImmediate() const { return Immediate; }
   bool isScalar() const { return NumVectors == 0; }
   bool isVector() const { return NumVectors > 0; }
   bool isScalableVector() const { return isVector() && IsScalable; }
   bool isChar() const { return ElementBitwidth == 8; }
   bool isVoid() const { return Void & !Pointer; }
   bool isDefault() const { return DefaultType; }
-  bool isFloat() const { return Float; }
-  bool isBFloat() const { return BFloat; }
+  bool isFloat() const { return Float && !BFloat; }
+  bool isBFloat() const { return BFloat && !Float; }
   bool isFloatingPoint() const { return Float || BFloat; }
   bool isInteger() const { return !isFloatingPoint() && !Predicate; }
   bool isScalarPredicate() const {
     return !isFloatingPoint() && Predicate && NumVectors == 0;
   }
   bool isPredicateVector() const { return Predicate; }
   bool isPredicatePattern() const { return PredicatePattern; }
   bool isPrefetchOp() const { return PrefetchOp; }
case 'f':
       ElementBitwidth = 32;
       break;
     case 'd':
       Float = true;
       ElementBitwidth = 64;
       break;
     case 'b':
       BFloat = true;
+      Float = false;
       ElementBitwidth = 16;
       break;
     default:
       llvm_unreachable("Unhandled type code!");
     }
   }
   assert(ElementBitwidth != ~0U && "Bad element bitwidth!");
 }
case 'm':
     Float = false;
     ElementBitwidth = Bitwidth = 32;
     NumVectors = 0;
     break;
   case 'n':
     Predicate = false;
     Signed = false;
     Float = false;
+    BFloat = false;
     ElementBitwidth = Bitwidth = 64;
     NumVectors = 0;
     break;
   case 'w':
     ElementBitwidth = 64;
     break;
   case 'j':
     ElementBitwidth = Bitwidth = 64;
void SVEType::applyModifier(char Mod) {
   case 'O':
     Predicate = false;
     Float = true;
     ElementBitwidth = 16;
     break;
   case 'M':
     Predicate = false;
     Float = true;
+    BFloat = false;
     ElementBitwidth = 32;
     break;
   case 'N':
     Predicate = false;
     Float = true;
     ElementBitwidth = 64;
     break;
   case 'Q':

llvm/include/llvm/IR/IntrinsicsAArch64.td

: Intrinsic<[],
                 ],
                 [IntrInaccessibleMemOrArgMemOnly, ImmArg<ArgIndex<3>>]>;
 
 class SVE_MatMul_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
                 [LLVMMatchType<0>, LLVMSubdivide4VectorType<0>, LLVMSubdivide4VectorType<0>],
                 [IntrNoMem]>;
 
+class SVE_4Vec_BF16
+    : Intrinsic<[llvm_nxv4f32_ty],
+                [llvm_nxv4f32_ty, llvm_nxv8bf16_ty, llvm_nxv8bf16_ty],
+                [IntrNoMem]>;
+
+class SVE_4Vec_BF16_Indexed
+    : Intrinsic<[llvm_nxv4f32_ty],
+                [llvm_nxv4f32_ty, llvm_nxv8bf16_ty, llvm_nxv8bf16_ty, llvm_i64_ty],
+                [IntrNoMem, ImmArg<ArgIndex<3>>]>;
+
 //
 // Vector tuple creation intrinsics (ACLE)
 //
 
 def int_aarch64_sve_tuple_create2 : AdvSIMD_SVE_Create_2Vector_Tuple;
 def int_aarch64_sve_tuple_create3 : AdvSIMD_SVE_Create_3Vector_Tuple;
 def int_aarch64_sve_tuple_create4 : AdvSIMD_SVE_Create_4Vector_Tuple;
 

 def int_aarch64_sve_fcmpne : AdvSIMD_SVE_Compare_Intrinsic;
 def int_aarch64_sve_fcmpuo : AdvSIMD_SVE_Compare_Intrinsic;
 
 def int_aarch64_sve_fcvtzs_i32f16   : Builtin_SVCVT<"svcvt_s32_f16_m", llvm_nxv4i32_ty, llvm_nxv4i1_ty, llvm_nxv8f16_ty>;
 def int_aarch64_sve_fcvtzs_i32f64   : Builtin_SVCVT<"svcvt_s32_f64_m", llvm_nxv4i32_ty, llvm_nxv2i1_ty, llvm_nxv2f64_ty>;
 def int_aarch64_sve_fcvtzs_i64f16   : Builtin_SVCVT<"svcvt_s64_f16_m", llvm_nxv2i64_ty, llvm_nxv2i1_ty, llvm_nxv8f16_ty>;
 def int_aarch64_sve_fcvtzs_i64f32   : Builtin_SVCVT<"svcvt_s64_f32_m", llvm_nxv2i64_ty, llvm_nxv2i1_ty, llvm_nxv4f32_ty>;
 
+def int_aarch64_sve_fcvt_bf16f32    : Builtin_SVCVT<"svcvt_bf16_f32_m",   llvm_nxv8bf16_ty, llvm_nxv8i1_ty, llvm_nxv4f32_ty>;
+def int_aarch64_sve_fcvtnt_bf16f32  : Builtin_SVCVT<"svcvtnt_bf16_f32_m", llvm_nxv8bf16_ty, llvm_nxv8i1_ty, llvm_nxv4f32_ty>;
+
 def int_aarch64_sve_fcvtzu_i32f16   : Builtin_SVCVT<"svcvt_u32_f16_m", llvm_nxv4i32_ty, llvm_nxv4i1_ty, llvm_nxv8f16_ty>;
 def int_aarch64_sve_fcvtzu_i32f64   : Builtin_SVCVT<"svcvt_u32_f64_m", llvm_nxv4i32_ty, llvm_nxv2i1_ty, llvm_nxv2f64_ty>;
 def int_aarch64_sve_fcvtzu_i64f16   : Builtin_SVCVT<"svcvt_u64_f16_m", llvm_nxv2i64_ty, llvm_nxv2i1_ty, llvm_nxv8f16_ty>;
 def int_aarch64_sve_fcvtzu_i64f32   : Builtin_SVCVT<"svcvt_u64_f32_m", llvm_nxv2i64_ty, llvm_nxv2i1_ty, llvm_nxv4f32_ty>;
 
 def int_aarch64_sve_fcvt_f16f32     : Builtin_SVCVT<"svcvt_f16_f32_m", llvm_nxv8f16_ty, llvm_nxv4i1_ty, llvm_nxv4f32_ty>;
 def int_aarch64_sve_fcvt_f16f64     : Builtin_SVCVT<"svcvt_f16_f64_m", llvm_nxv8f16_ty, llvm_nxv2i1_ty, llvm_nxv2f64_ty>;
 def int_aarch64_sve_fcvt_f32f64     : Builtin_SVCVT<"svcvt_f32_f64_m", llvm_nxv4f32_ty, llvm_nxv2i1_ty, llvm_nxv2f64_ty>;

 def int_aarch64_sve_usdot_lane : AdvSIMD_SVE_DOT_Indexed_Intrinsic;
 def int_aarch64_sve_sudot_lane : AdvSIMD_SVE_DOT_Indexed_Intrinsic;
 
 //
 // SVE ACLE: 7.4/5. FP64/FP32 matrix multiply extensions
 //
 def int_aarch64_sve_fmmla : AdvSIMD_3VectorArg_Intrinsic;
 
+//
+// SVE ACLE: 7.2. BFloat16 extensions
+//
+
+def int_aarch64_sve_bfdot   : SVE_4Vec_BF16;
+def int_aarch64_sve_bfmlalb : SVE_4Vec_BF16;
+def int_aarch64_sve_bfmlalt : SVE_4Vec_BF16;
+
+def int_aarch64_sve_bfmmla  : SVE_4Vec_BF16;
+
+def int_aarch64_sve_bfdot_lane   : SVE_4Vec_BF16_Indexed;
+def int_aarch64_sve_bfmlalb_lane : SVE_4Vec_BF16_Indexed;
+def int_aarch64_sve_bfmlalt_lane : SVE_4Vec_BF16_Indexed;
 }
 
 //
 // SVE2 - Contiguous conflict detection
 //
 
 def int_aarch64_sve_whilerw_b : SVE2_CONFLICT_DETECT_Intrinsic;
 def int_aarch64_sve_whilerw_h : SVE2_CONFLICT_DETECT_Intrinsic;
 def int_aarch64_sve_whilerw_s : SVE2_CONFLICT_DETECT_Intrinsic;
 def int_aarch64_sve_whilerw_d : SVE2_CONFLICT_DETECT_Intrinsic;
 def int_aarch64_sve_whilewr_b : SVE2_CONFLICT_DETECT_Intrinsic;
 def int_aarch64_sve_whilewr_h : SVE2_CONFLICT_DETECT_Intrinsic;
 def int_aarch64_sve_whilewr_s : SVE2_CONFLICT_DETECT_Intrinsic;
 def int_aarch64_sve_whilewr_d : SVE2_CONFLICT_DETECT_Intrinsic;

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

multiclass sve_prefetch<SDPatternOperator prefetch, ValueType PredTy, Instruction RegImmInst, Instruction RegRegInst, int scale, ComplexPattern AddrCP> {
   defm FRINTZ_ZPmZ : sve_fp_2op_p_zd_HSD<0b00011, "frintz", int_aarch64_sve_frintz>;
   defm FRINTA_ZPmZ : sve_fp_2op_p_zd_HSD<0b00100, "frinta", int_aarch64_sve_frinta>;
   defm FRINTX_ZPmZ : sve_fp_2op_p_zd_HSD<0b00110, "frintx", int_aarch64_sve_frintx>;
   defm FRINTI_ZPmZ : sve_fp_2op_p_zd_HSD<0b00111, "frinti", int_aarch64_sve_frinti>;
   defm FRECPX_ZPmZ : sve_fp_2op_p_zd_HSD<0b01100, "frecpx", int_aarch64_sve_frecpx>;
   defm FSQRT_ZPmZ  : sve_fp_2op_p_zd_HSD<0b01101, "fsqrt",  int_aarch64_sve_fsqrt>;
 
   let Predicates = [HasBF16, HasSVE] in {
-    def BFDOT_ZZZ    : sve_bfloat_dot<"bfdot">;
-    def BFDOT_ZZI    : sve_bfloat_dot_indexed<"bfdot">;
-    def BFMMLA_ZZZ  : sve_bfloat_matmul<"bfmmla">;
-    def BFMMLA_B_ZZZ  : sve_bfloat_matmul_longvecl<0b0, "bfmlalb">;
-    def BFMMLA_T_ZZZ  : sve_bfloat_matmul_longvecl<0b1, "bfmlalt">;
-    def BFMMLA_B_ZZI  : sve_bfloat_matmul_longvecl_idx<0b0, "bfmlalb">;
-    def BFMMLA_T_ZZI  : sve_bfloat_matmul_longvecl_idx<0b1, "bfmlalt">;
-    def BFCVT_ZPmZ   : sve_bfloat_convert<0b1, "bfcvt">;
-    def BFCVTNT_ZPmZ : sve_bfloat_convert<0b0, "bfcvtnt">;
+    defm BFDOT_ZZZ    : sve_bfloat_dot<"bfdot", int_aarch64_sve_bfdot>;
+    defm BFDOT_ZZI    : sve_bfloat_dot_indexed<"bfdot", int_aarch64_sve_bfdot_lane>;
+    defm BFMMLA_ZZZ   : sve_bfloat_matmul<"bfmmla", int_aarch64_sve_bfmmla>;
+    defm BFMMLA_B_ZZZ : sve_bfloat_matmul_longvecl<0b0, "bfmlalb", int_aarch64_sve_bfmlalb>;
+    defm BFMMLA_T_ZZZ : sve_bfloat_matmul_longvecl<0b1, "bfmlalt", int_aarch64_sve_bfmlalt>;
+    defm BFMMLA_B_ZZI : sve_bfloat_matmul_longvecl_idx<0b0, "bfmlalb", int_aarch64_sve_bfmlalb_lane>;
+    defm BFMMLA_T_ZZI : sve_bfloat_matmul_longvecl_idx<0b1, "bfmlalt", int_aarch64_sve_bfmlalt_lane>;
+    defm BFCVT_ZPmZ   : sve_bfloat_convert<0b1, "bfcvt",   int_aarch64_sve_fcvt_bf16f32>;
+    defm BFCVTNT_ZPmZ : sve_bfloat_convert<0b0, "bfcvtnt", int_aarch64_sve_fcvtnt_bf16f32>;
   }
 
   // InstAliases
   def : InstAlias<"mov $Zd, $Zn",
                   (ORR_ZZZ ZPR64:$Zd, ZPR64:$Zn, ZPR64:$Zn), 1>;
   def : InstAlias<"mov $Pd, $Pg/m, $Pn",
                   (SEL_PPPP PPR8:$Pd, PPRAny:$Pg, PPR8:$Pn, PPR8:$Pd), 1>;
   def : InstAlias<"mov $Pd, $Pn",

llvm/lib/Target/AArch64/SVEInstrFormats.td

 
 class sve_bfloat_dot<string asm>
 : sve_bfloat_dot_base<0b10, asm, "\t$Zda, $Zn, $Zm",
   (ins ZPR32:$_Zda, ZPR16:$Zn, ZPR16:$Zm)> {
   bits<5> Zm;
   let Inst{20-16} = Zm;
 }
 
+multiclass sve_bfloat_dot<string asm, SDPatternOperator op> {
+  def NAME : sve_bfloat_dot<asm>;
+  def : SVE_3_Op_Pat<nxv4f32, op, nxv4f32, nxv8bf16, nxv8bf16 ,!cast<Instruction>(NAME)>;
+}
+
 class sve_bfloat_dot_indexed<string asm>
 : sve_bfloat_dot_base<0b01, asm, "\t$Zda, $Zn, $Zm$iop",
   (ins ZPR32:$_Zda, ZPR16:$Zn, ZPR3b16:$Zm, VectorIndexS:$iop)> {
   bits<2> iop;
   bits<3> Zm;
   let Inst{20-19} = iop;
   let Inst{18-16} = Zm;
 }
 
+multiclass sve_bfloat_dot_indexed<string asm, SDPatternOperator op> {
+  def NAME : sve_bfloat_dot_indexed<asm>;
+  def : SVE_4_Op_Imm_Pat<nxv4f32, op, nxv4f32, nxv8bf16, nxv8bf16, i64, VectorIndexS_timm, !cast<Instruction>(NAME)>;
+}
+
 class sve_bfloat_matmul<string asm>
 : I<(outs ZPR32:$Zda), (ins ZPR32:$_Zda, ZPR16:$Zn, ZPR16:$Zm),
   asm, "\t$Zda, $Zn, $Zm", "", []>, Sched<[]> {
   bits<5> Zm;
   bits<5> Zda;
   bits<5> Zn;
   let Inst{31-21} = 0b01100100011;
   let Inst{20-16} = Zm;
   let Inst{15-10} = 0b111001;
   let Inst{9-5}   = Zn;
   let Inst{4-0}   = Zda;
 
   let Constraints = "$Zda = $_Zda";
   let DestructiveInstType = DestructiveOther;
   let ElementSize = ElementSizeH;
 }
 
+multiclass sve_bfloat_matmul<string asm, SDPatternOperator op> {
+  def NAME : sve_bfloat_matmul<asm>;
+  def : SVE_3_Op_Pat<nxv4f32, op, nxv4f32, nxv8bf16, nxv8bf16 ,!cast<Instruction>(NAME)>;
+}
+
 class sve_bfloat_matmul_longvecl<bit BT, string asm>
 : sve_bfloat_matmul<asm> {
   let Inst{23}    = 0b1;
   let Inst{14-13} = 0b00;
   let Inst{10}    = BT;
 }
 
+multiclass sve_bfloat_matmul_longvecl<bit BT, string asm, SDPatternOperator op> {
+  def NAME : sve_bfloat_matmul_longvecl<BT, asm>;
+  def : SVE_3_Op_Pat<nxv4f32, op, nxv4f32, nxv8bf16, nxv8bf16 ,!cast<Instruction>(NAME)>;
+}
+
 class sve_bfloat_matmul_longvecl_idx<bit BT, string asm>
 : sve_bfloat_dot_base<0b01, asm, "\t$Zda, $Zn, $Zm$iop",
   (ins ZPR32:$_Zda, ZPR16:$Zn, ZPR3b16:$Zm, VectorIndexH:$iop)> {
   bits<3> iop;
   bits<3> Zm;
   let Inst{23}    = 0b1;
   let Inst{20-19} = iop{2-1};
   let Inst{18-16} = Zm;
   let Inst{11}    = iop{0};
   let Inst{10}    = BT;
 }
 
+multiclass sve_bfloat_matmul_longvecl_idx<bit BT, string asm, SDPatternOperator op> {
+  def NAME : sve_bfloat_matmul_longvecl_idx<BT, asm>;
+  def : SVE_4_Op_Imm_Pat<nxv4f32, op, nxv4f32, nxv8bf16, nxv8bf16, i64, VectorIndexH_timm, !cast<Instruction>(NAME)>;
+}
+
 class sve_bfloat_convert<bit N, string asm>
 : I<(outs ZPR16:$Zd), (ins ZPR16:$_Zd, PPR3bAny:$Pg, ZPR32:$Zn),
   asm, "\t$Zd, $Pg/m, $Zn", "", []>, Sched<[]> {
   bits<5> Zd;
   bits<3> Pg;
   bits<5> Zn;
   let Inst{31-25} = 0b0110010;
   let Inst{24}    = N;
   let Inst{23-13} = 0b10001010101;
   let Inst{12-10} = Pg;
   let Inst{9-5}   = Zn;
   let Inst{4-0}   = Zd;
 
   let Constraints = "$Zd = $_Zd";
   let DestructiveInstType = DestructiveOther;
   let hasSideEffects = 1;
   let ElementSize = ElementSizeS;
 }
 
+multiclass sve_bfloat_convert<bit N, string asm, SDPatternOperator op> {
+  def NAME : sve_bfloat_convert<N, asm>;
+  def : SVE_3_Op_Pat<nxv8bf16, op, nxv8bf16, nxv8i1, nxv4f32, !cast<Instruction>(NAME)>;
+}
+
 //===----------------------------------------------------------------------===//
 // SVE Integer Matrix Multiply Group
 //===----------------------------------------------------------------------===//
 
 class sve_int_matmul<bits<2> uns, string asm>
 : I<(outs ZPR32:$Zda), (ins ZPR32:$_Zda, ZPR8:$Zn, ZPR8:$Zm), asm,
   "\t$Zda, $Zn, $Zm", "", []>, Sched<[]> {
   bits<5> Zda;

llvm/test/CodeGen/AArch64/sve-intrinsics-bfloat.ll

+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve,+bf16 -asm-verbose=0 < %s | FileCheck %s
+
+;
+; BFDOT
+;
+
+define <vscale x 4 x float> @bfdot_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfdot_f32:
+; CHECK-NEXT:  bfdot z0.s, z1.h, z2.h
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfdot(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfdot_lane_0_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfdot_lane_0_f32:
+; CHECK-NEXT:  bfdot z0.s, z1.h, z2.h[0]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfdot.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 0)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfdot_lane_1_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfdot_lane_1_f32:
+; CHECK-NEXT:  bfdot z0.s, z1.h, z2.h[1]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfdot.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 1)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfdot_lane_2_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfdot_lane_2_f32:
+; CHECK-NEXT:  bfdot z0.s, z1.h, z2.h[2]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfdot.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 2)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfdot_lane_3_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfdot_lane_3_f32:
+; CHECK-NEXT:  bfdot z0.s, z1.h, z2.h[3]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfdot.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 3)
+  ret <vscale x 4 x float> %out
+}
+
+;
+; BFMLALB
+;
+
+define <vscale x 4 x float> @bfmlalb_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalb_f32:
+; CHECK-NEXT:  bfmlalb z0.s, z1.h, z2.h
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalb_lane_0_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalb_lane_0_f32:
+; CHECK-NEXT:  bfmlalb z0.s, z1.h, z2.h[0]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 0)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalb_lane_1_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalb_lane_1_f32:
+; CHECK-NEXT:  bfmlalb z0.s, z1.h, z2.h[1]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 1)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalb_lane_2_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalb_lane_2_f32:
+; CHECK-NEXT:  bfmlalb z0.s, z1.h, z2.h[2]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 2)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalb_lane_3_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalb_lane_3_f32:
+; CHECK-NEXT:  bfmlalb z0.s, z1.h, z2.h[3]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 3)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalb_lane_4_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalb_lane_4_f32:
+; CHECK-NEXT:  bfmlalb z0.s, z1.h, z2.h[4]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 4)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalb_lane_5_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalb_lane_5_f32:
+; CHECK-NEXT:  bfmlalb z0.s, z1.h, z2.h[5]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 5)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalb_lane_6_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalb_lane_6_f32:
+; CHECK-NEXT:  bfmlalb z0.s, z1.h, z2.h[6]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 6)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalb_lane_7_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalb_lane_7_f32:
+; CHECK-NEXT:  bfmlalb z0.s, z1.h, z2.h[7]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 7)
+  ret <vscale x 4 x float> %out
+}
+
+;
+; BFMLALT
+;
+
+define <vscale x 4 x float> @bfmlalt_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalt_f32:
+; CHECK-NEXT:  bfmlalt z0.s, z1.h, z2.h
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalt_lane_0_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalt_lane_0_f32:
+; CHECK-NEXT:  bfmlalt z0.s, z1.h, z2.h[0]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 0)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalt_lane_1_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalt_lane_1_f32:
+; CHECK-NEXT:  bfmlalt z0.s, z1.h, z2.h[1]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 1)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalt_lane_2_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalt_lane_2_f32:
+; CHECK-NEXT:  bfmlalt z0.s, z1.h, z2.h[2]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 2)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalt_lane_3_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalt_lane_3_f32:
+; CHECK-NEXT:  bfmlalt z0.s, z1.h, z2.h[3]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 3)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalt_lane_4_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalt_lane_4_f32:
+; CHECK-NEXT:  bfmlalt z0.s, z1.h, z2.h[4]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 4)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalt_lane_5_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalt_lane_5_f32:
+; CHECK-NEXT:  bfmlalt z0.s, z1.h, z2.h[5]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 5)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalt_lane_6_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalt_lane_6_f32:
+; CHECK-NEXT:  bfmlalt z0.s, z1.h, z2.h[6]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 6)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 4 x float> @bfmlalt_lane_7_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmlalt_lane_7_f32:
+; CHECK-NEXT:  bfmlalt z0.s, z1.h, z2.h[7]
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt.lane(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c, i64 7)
+  ret <vscale x 4 x float> %out
+}
+
+;
+; BFMMLA
+;
+
+define <vscale x 4 x float> @bfmmla_f32(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c) nounwind {
+; CHECK-LABEL: bfmmla_f32:
+; CHECK-NEXT:  bfmmla z0.s, z1.h, z2.h
+; CHECK-NEXT:  ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.bfmmla(<vscale x 4 x float> %a, <vscale x 8 x bfloat> %b, <vscale x 8 x bfloat> %c)
+  ret <vscale x 4 x float> %out
+}
+
+;
+; BFCVT
+;
+
+define <vscale x 8 x bfloat> @fcvt_bf16_f32(<vscale x 8 x bfloat> %a, <vscale x 8 x i1> %pg, <vscale x 4 x float> %b) nounwind {
+; CHECK-LABEL: fcvt_bf16_f32:
+; CHECK-NEXT: bfcvt z0.h, p0/m, z1.s
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x bfloat> @llvm.aarch64.sve.fcvt.bf16f32(<vscale x 8 x bfloat> %a, <vscale x 8 x i1> %pg, <vscale x 4 x float> %b)
+  ret <vscale x 8 x bfloat> %out
+}
+
+;
+; BFCVTNT
+;
+
+define <vscale x 8 x bfloat> @fcvtnt_bf16_f32(<vscale x 8 x bfloat> %a, <vscale x 8 x i1> %pg, <vscale x 4 x float> %b) nounwind {
+; CHECK-LABEL: fcvtnt_bf16_f32:
+; CHECK-NEXT: bfcvtnt z0.h, p0/m, z1.s
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x bfloat> @llvm.aarch64.sve.fcvtnt.bf16f32(<vscale x 8 x bfloat> %a, <vscale x 8 x i1> %pg, <vscale x 4 x float> %b)
+  ret <vscale x 8 x bfloat> %out
+}
+
+declare <vscale x 4 x float> @llvm.aarch64.sve.bfdot(<vscale x 4 x float>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>)
+declare <vscale x 4 x float> @llvm.aarch64.sve.bfdot.lane(<vscale x 4 x float>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, i64)
+declare <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb(<vscale x 4 x float>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>)
+declare <vscale x 4 x float> @llvm.aarch64.sve.bfmlalb.lane(<vscale x 4 x float>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, i64)
+declare <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt(<vscale x 4 x float>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>)
+declare <vscale x 4 x float> @llvm.aarch64.sve.bfmlalt.lane(<vscale x 4 x float>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, i64)
+declare <vscale x 4 x float> @llvm.aarch64.sve.bfmmla(<vscale x 4 x float>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>)
+declare <vscale x 8 x bfloat> @llvm.aarch64.sve.fcvt.bf16f32(<vscale x 8 x bfloat>, <vscale x 8 x i1>, <vscale x 4 x float>)
+declare <vscale x 8 x bfloat> @llvm.aarch64.sve.fcvtnt.bf16f32(<vscale x 8 x bfloat>, <vscale x 8 x i1>, <vscale x 4 x float>)