[clang] [llvm] [AArch64] Implement FP8 SVE Intrinsics for narrowing conversions (PR #118124)

Momchil Velikov via cfe-commits cfe-commits at lists.llvm.org
Fri Nov 29 11:30:27 PST 2024


https://github.com/momchil-velikov created https://github.com/llvm/llvm-project/pull/118124

None

>From 3a9643e6c2d61eae2e23df42c19b1410d4a5fcc5 Mon Sep 17 00:00:00 2001
From: Momchil Velikov <momchil.velikov at arm.com>
Date: Thu, 21 Nov 2024 11:21:29 +0000
Subject: [PATCH 1/2] FP8 CVT/CVTL

---
 clang/include/clang/Basic/arm_sve.td          |  10 +
 .../fp8-intrinsics/acle_sve2_fp8_cvt.c        | 173 ++++++++++++++++++
 .../aarch64-sve2-intrinsics/acle_sve2_fp8.c   |  24 +++
 llvm/include/llvm/IR/IntrinsicsAArch64.td     |  17 ++
 .../lib/Target/AArch64/AArch64SVEInstrInfo.td |  16 +-
 llvm/lib/Target/AArch64/SVEInstrFormats.td    |   7 +-
 .../test/CodeGen/AArch64/fp8-sve-cvt-cvtlt.ll |  78 ++++++++
 7 files changed, 316 insertions(+), 9 deletions(-)
 create mode 100644 clang/test/CodeGen/AArch64/fp8-intrinsics/acle_sve2_fp8_cvt.c
 create mode 100644 clang/test/Sema/aarch64-sve2-intrinsics/acle_sve2_fp8.c
 create mode 100644 llvm/test/CodeGen/AArch64/fp8-sve-cvt-cvtlt.ll

diff --git a/clang/include/clang/Basic/arm_sve.td b/clang/include/clang/Basic/arm_sve.td
index b36e592042da0b..b9d8360843aa8e 100644
--- a/clang/include/clang/Basic/arm_sve.td
+++ b/clang/include/clang/Basic/arm_sve.td
@@ -2447,3 +2447,13 @@ let SVETargetGuard = "sve2,faminmax", SMETargetGuard = "sme2,faminmax" in {
   defm SVAMIN : SInstZPZZ<"svamin", "hfd", "aarch64_sve_famin", "aarch64_sve_famin_u">;
   defm SVAMAX : SInstZPZZ<"svamax", "hfd", "aarch64_sve_famax", "aarch64_sve_famax_u">;
 }
+
+let SVETargetGuard = "sve2,fp8", SMETargetGuard = "sme2,fp8" in {
+  // 8-bit floating-point convert to BFloat16/Float16
+  def SVF1CVT : SInst<"svcvt1_{d}[_mf8]_fpm", "d~>", "bh", MergeNone, "aarch64_sve_fp8_cvt1", [VerifyRuntimeMode, SetsFPMR]>;
+  def SVF2CVT : SInst<"svcvt2_{d}[_mf8]_fpm", "d~>", "bh", MergeNone, "aarch64_sve_fp8_cvt2", [VerifyRuntimeMode, SetsFPMR]>;
+
+  // 8-bit floating-point convert to BFloat16/Float16 (top)
+  def SVF1CVTLT : SInst<"svcvtlt1_{d}[_mf8]_fpm", "d~>", "bh", MergeNone, "aarch64_sve_fp8_cvtlt1", [VerifyRuntimeMode, SetsFPMR]>;
+  def SVF2CVTLT : SInst<"svcvtlt2_{d}[_mf8]_fpm", "d~>", "bh", MergeNone, "aarch64_sve_fp8_cvtlt2", [VerifyRuntimeMode, SetsFPMR]>;
+}
diff --git a/clang/test/CodeGen/AArch64/fp8-intrinsics/acle_sve2_fp8_cvt.c b/clang/test/CodeGen/AArch64/fp8-intrinsics/acle_sve2_fp8_cvt.c
new file mode 100644
index 00000000000000..c026b8aa216f32
--- /dev/null
+++ b/clang/test/CodeGen/AArch64/fp8-intrinsics/acle_sve2_fp8_cvt.c
@@ -0,0 +1,173 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5
+// RUN: %clang_cc1        -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +sve2 -target-feature +fp8 -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s
+// RUN: %clang_cc1 -x c++ -triple aarch64-none-linux-gnu -target-feature +sme -target-feature +sme2 -target-feature +fp8 -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s -check-prefix=CHECK-CXX
+
+// RUN: %clang_cc1 -DSME_OVERLOADED_FORMS        -triple aarch64-none-linux-gnu -target-feature +sme -target-feature +sme2 -target-feature +fp8 -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s
+// RUN: %clang_cc1 -DSME_OVERLOADED_FORMS -x c++ -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +sve2 -target-feature +fp8 -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s -check-prefix=CHECK-CXX
+
+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +sve2 -target-feature +fp8 -S -disable-O0-optnone -Werror -Wall -o /dev/null %s
+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sme -target-feature +sme2 -target-feature +fp8 -S -disable-O0-optnone -Werror -Wall -o /dev/null %s
+
+// REQUIRES: aarch64-registered-target
+
+#ifdef __ARM_FEATURE_SME
+#include <arm_sme.h>
+#else
+#include <arm_sve.h>
+#endif
+
+#ifdef SVE_OVERLOADED_FORMS
+#define SVE_ACLE_FUNC(A1,A2_UNUSED,A3) A1##A3
+#else
+#define SVE_ACLE_FUNC(A1,A2,A3) A1##A2##A3
+#endif
+
+#ifdef __ARM_FEATURE_SME
+#define STREAMING __arm_streaming
+#else
+#define STREAMING
+#endif
+
+// CHECK-LABEL: define dso_local <vscale x 8 x bfloat> @test_svcvt1_bf16_mf8(
+// CHECK-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvt1.nxv8bf16(<vscale x 16 x i8> [[ZN]])
+// CHECK-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 8 x bfloat> @_Z20test_svcvt1_bf16_mf8u13__SVMfloat8_tm(
+// CHECK-CXX-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvt1.nxv8bf16(<vscale x 16 x i8> [[ZN]])
+// CHECK-CXX-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
+//
+svbfloat16_t test_svcvt1_bf16_mf8(svmfloat8_t zn, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvt1_bf16,_mf8,_fpm)(zn, fpm);
+}
+
+// CHECK-LABEL: define dso_local <vscale x 8 x bfloat> @test_svcvt2_bf16_mf8(
+// CHECK-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvt2.nxv8bf16(<vscale x 16 x i8> [[ZN]])
+// CHECK-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 8 x bfloat> @_Z20test_svcvt2_bf16_mf8u13__SVMfloat8_tm(
+// CHECK-CXX-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvt2.nxv8bf16(<vscale x 16 x i8> [[ZN]])
+// CHECK-CXX-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
+//
+svbfloat16_t test_svcvt2_bf16_mf8(svmfloat8_t zn, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvt2_bf16,_mf8,_fpm)(zn, fpm);
+}
+
+// CHECK-LABEL: define dso_local <vscale x 8 x bfloat> @test_svcvtlt1_bf16_mf8(
+// CHECK-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvtlt1.nxv8bf16(<vscale x 16 x i8> [[ZN]])
+// CHECK-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 8 x bfloat> @_Z22test_svcvtlt1_bf16_mf8u13__SVMfloat8_tm(
+// CHECK-CXX-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvtlt1.nxv8bf16(<vscale x 16 x i8> [[ZN]])
+// CHECK-CXX-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
+//
+svbfloat16_t test_svcvtlt1_bf16_mf8(svmfloat8_t zn, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvtlt1_bf16,_mf8,_fpm)(zn, fpm);
+}
+
+// CHECK-LABEL: define dso_local <vscale x 8 x bfloat> @test_svcvtlt2_bf16_mf8(
+// CHECK-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvtlt2.nxv8bf16(<vscale x 16 x i8> [[ZN]])
+// CHECK-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 8 x bfloat> @_Z22test_svcvtlt2_bf16_mf8u13__SVMfloat8_tm(
+// CHECK-CXX-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvtlt2.nxv8bf16(<vscale x 16 x i8> [[ZN]])
+// CHECK-CXX-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
+//
+svbfloat16_t test_svcvtlt2_bf16_mf8(svmfloat8_t zn, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvtlt2_bf16,_mf8,_fpm)(zn, fpm);
+}
+
+// CHECK-LABEL: define dso_local <vscale x 8 x half> @test_svcvt1_f16_mf8(
+// CHECK-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvt1.nxv8f16(<vscale x 16 x i8> [[ZN]])
+// CHECK-NEXT:    ret <vscale x 8 x half> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 8 x half> @_Z19test_svcvt1_f16_mf8u13__SVMfloat8_tm(
+// CHECK-CXX-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvt1.nxv8f16(<vscale x 16 x i8> [[ZN]])
+// CHECK-CXX-NEXT:    ret <vscale x 8 x half> [[TMP0]]
+//
+svfloat16_t test_svcvt1_f16_mf8(svmfloat8_t zn, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvt1_f16,_mf8,_fpm)(zn, fpm);
+}
+
+// CHECK-LABEL: define dso_local <vscale x 8 x half> @test_svcvt2_f16_mf8(
+// CHECK-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvt2.nxv8f16(<vscale x 16 x i8> [[ZN]])
+// CHECK-NEXT:    ret <vscale x 8 x half> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 8 x half> @_Z19test_svcvt2_f16_mf8u13__SVMfloat8_tm(
+// CHECK-CXX-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvt2.nxv8f16(<vscale x 16 x i8> [[ZN]])
+// CHECK-CXX-NEXT:    ret <vscale x 8 x half> [[TMP0]]
+//
+svfloat16_t test_svcvt2_f16_mf8(svmfloat8_t zn, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvt2_f16,_mf8,_fpm)(zn, fpm);
+}
+
+// CHECK-LABEL: define dso_local <vscale x 8 x half> @test_svcvtlt1_f16_mf8(
+// CHECK-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvtlt1.nxv8f16(<vscale x 16 x i8> [[ZN]])
+// CHECK-NEXT:    ret <vscale x 8 x half> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 8 x half> @_Z21test_svcvtlt1_f16_mf8u13__SVMfloat8_tm(
+// CHECK-CXX-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvtlt1.nxv8f16(<vscale x 16 x i8> [[ZN]])
+// CHECK-CXX-NEXT:    ret <vscale x 8 x half> [[TMP0]]
+//
+svfloat16_t test_svcvtlt1_f16_mf8(svmfloat8_t zn, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvtlt1_f16,_mf8,_fpm)(zn, fpm);
+}
+
+// CHECK-LABEL: define dso_local <vscale x 8 x half> @test_svcvtlt2_f16_mf8(
+// CHECK-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvtlt2.nxv8f16(<vscale x 16 x i8> [[ZN]])
+// CHECK-NEXT:    ret <vscale x 8 x half> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 8 x half> @_Z21test_svcvtlt2_f16_mf8u13__SVMfloat8_tm(
+// CHECK-CXX-SAME: <vscale x 16 x i8> [[ZN:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvtlt2.nxv8f16(<vscale x 16 x i8> [[ZN]])
+// CHECK-CXX-NEXT:    ret <vscale x 8 x half> [[TMP0]]
+//
+svfloat16_t test_svcvtlt2_f16_mf8(svmfloat8_t zn, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvtlt2_f16,_mf8,_fpm)(zn, fpm);
+}
diff --git a/clang/test/Sema/aarch64-sve2-intrinsics/acle_sve2_fp8.c b/clang/test/Sema/aarch64-sve2-intrinsics/acle_sve2_fp8.c
new file mode 100644
index 00000000000000..aafd42f798d935
--- /dev/null
+++ b/clang/test/Sema/aarch64-sve2-intrinsics/acle_sve2_fp8.c
@@ -0,0 +1,24 @@
+// REQUIRES: aarch64-registered-target
+
+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sve -verify -emit-llvm %s
+
+#include <arm_sve.h>
+
+void test_features(svmfloat8_t zn, fpm_t fpm) {
+  svcvt1_bf16_mf8_fpm(zn, fpm);
+  // expected-error at -1 {{'svcvt1_bf16_mf8_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+  svcvt2_bf16_mf8_fpm(zn, fpm);
+  // expected-error at -1 {{'svcvt2_bf16_mf8_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+  svcvtlt1_bf16_mf8_fpm(zn, fpm);
+  // expected-error at -1 {{'svcvtlt1_bf16_mf8_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+  svcvtlt2_bf16_mf8_fpm(zn, fpm);
+  // expected-error at -1 {{'svcvtlt2_bf16_mf8_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+  svcvt1_f16_mf8_fpm(zn, fpm);
+  // expected-error at -1 {{'svcvt1_f16_mf8_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+  svcvt2_f16_mf8_fpm(zn, fpm);
+  // expected-error at -1 {{'svcvt2_f16_mf8_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+  svcvtlt1_f16_mf8_fpm(zn, fpm);
+  // expected-error at -1 {{'svcvtlt1_f16_mf8_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+  svcvtlt2_f16_mf8_fpm(zn, fpm);
+  // expected-error at -1 {{'svcvtlt2_f16_mf8_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+}
diff --git a/llvm/include/llvm/IR/IntrinsicsAArch64.td b/llvm/include/llvm/IR/IntrinsicsAArch64.td
index a91616b9556828..13bc5e08d2756f 100644
--- a/llvm/include/llvm/IR/IntrinsicsAArch64.td
+++ b/llvm/include/llvm/IR/IntrinsicsAArch64.td
@@ -3864,3 +3864,20 @@ def int_aarch64_sve_famin_u : AdvSIMD_Pred2VectorArg_Intrinsic;
 // Neon absolute maximum and minimum
 def int_aarch64_neon_famax :  AdvSIMD_2VectorArg_Intrinsic;
 def int_aarch64_neon_famin :  AdvSIMD_2VectorArg_Intrinsic;
+
+//
+// FP8 intrinsics
+//
+let TargetPrefix = "aarch64" in {
+
+// Conversions
+class SVE2_FP8_Cvt
+  : DefaultAttrsIntrinsic<[llvm_anyvector_ty],
+                          [llvm_nxv16i8_ty],
+                          [IntrReadMem, IntrInaccessibleMemOnly]>;
+
+def int_aarch64_sve_fp8_cvt1   : SVE2_FP8_Cvt;
+def int_aarch64_sve_fp8_cvt2   : SVE2_FP8_Cvt;
+def int_aarch64_sve_fp8_cvtlt1 : SVE2_FP8_Cvt;
+def int_aarch64_sve_fp8_cvtlt2 : SVE2_FP8_Cvt;
+}
diff --git a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
index fb0eb7a80c6d72..5365a00f3f42ee 100644
--- a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
@@ -4369,14 +4369,14 @@ let Predicates = [HasNonStreamingSVE2p2orSME2p2] in {
 //===----------------------------------------------------------------------===//
 let Predicates = [HasSVE2orSME2, HasFP8] in {
 // FP8 upconvert
-defm F1CVT_ZZ     : sve2_fp8_cvt_single<0b0, 0b00, "f1cvt">;
-defm F2CVT_ZZ     : sve2_fp8_cvt_single<0b0, 0b01, "f2cvt">;
-defm BF1CVT_ZZ    : sve2_fp8_cvt_single<0b0, 0b10, "bf1cvt">;
-defm BF2CVT_ZZ    : sve2_fp8_cvt_single<0b0, 0b11, "bf2cvt">;
-defm F1CVTLT_ZZ   : sve2_fp8_cvt_single<0b1, 0b00, "f1cvtlt">;
-defm F2CVTLT_ZZ   : sve2_fp8_cvt_single<0b1, 0b01, "f2cvtlt">;
-defm BF1CVTLT_ZZ  : sve2_fp8_cvt_single<0b1, 0b10, "bf1cvtlt">;
-defm BF2CVTLT_ZZ  : sve2_fp8_cvt_single<0b1, 0b11, "bf2cvtlt">;
+defm F1CVT_ZZ     : sve2_fp8_cvt_single<0b0, 0b00, "f1cvt",    nxv8f16,  int_aarch64_sve_fp8_cvt1>;
+defm F2CVT_ZZ     : sve2_fp8_cvt_single<0b0, 0b01, "f2cvt",    nxv8f16,  int_aarch64_sve_fp8_cvt2>;
+defm BF1CVT_ZZ    : sve2_fp8_cvt_single<0b0, 0b10, "bf1cvt",   nxv8bf16, int_aarch64_sve_fp8_cvt1>;
+defm BF2CVT_ZZ    : sve2_fp8_cvt_single<0b0, 0b11, "bf2cvt",   nxv8bf16, int_aarch64_sve_fp8_cvt2>;
+defm F1CVTLT_ZZ   : sve2_fp8_cvt_single<0b1, 0b00, "f1cvtlt",  nxv8f16,  int_aarch64_sve_fp8_cvtlt1>;
+defm F2CVTLT_ZZ   : sve2_fp8_cvt_single<0b1, 0b01, "f2cvtlt",  nxv8f16,  int_aarch64_sve_fp8_cvtlt2>;
+defm BF1CVTLT_ZZ  : sve2_fp8_cvt_single<0b1, 0b10, "bf1cvtlt", nxv8bf16, int_aarch64_sve_fp8_cvtlt1>;
+defm BF2CVTLT_ZZ  : sve2_fp8_cvt_single<0b1, 0b11, "bf2cvtlt", nxv8bf16, int_aarch64_sve_fp8_cvtlt2>;
 
 // FP8 downconvert
 defm FCVTN_Z2Z_HtoB  : sve2_fp8_down_cvt_single<0b00, "fcvtn", ZZ_h_mul_r>;
diff --git a/llvm/lib/Target/AArch64/SVEInstrFormats.td b/llvm/lib/Target/AArch64/SVEInstrFormats.td
index faaaca3f28d758..58770bf6e274d3 100644
--- a/llvm/lib/Target/AArch64/SVEInstrFormats.td
+++ b/llvm/lib/Target/AArch64/SVEInstrFormats.td
@@ -10733,10 +10733,15 @@ class sve2_fp8_cvt_single<bit L, bits<2> opc, string mnemonic,
   let Inst{9-5}   = Zn;
   let Inst{4-0}   = Zd;
   let Uses = [FPMR, FPCR];
+
+  let mayLoad  = 1;
+  let mayStore = 0;
 }
 
-multiclass sve2_fp8_cvt_single<bit L, bits<2> opc, string mnemonic> {
+multiclass sve2_fp8_cvt_single<bit L, bits<2> opc, string mnemonic, ValueType vtd, SDPatternOperator op> {
   def _BtoH : sve2_fp8_cvt_single<L, opc, mnemonic, ZPR16, ZPR8>;
+  
+  def : SVE_1_Op_Pat<vtd, op, nxv16i8, !cast<Instruction>(NAME # _BtoH)>;
 }
 
 // FP8 downconvert
diff --git a/llvm/test/CodeGen/AArch64/fp8-sve-cvt-cvtlt.ll b/llvm/test/CodeGen/AArch64/fp8-sve-cvt-cvtlt.ll
new file mode 100644
index 00000000000000..bf0030462db98f
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/fp8-sve-cvt-cvtlt.ll
@@ -0,0 +1,78 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mattr=+bf16,+sve2,+fp8 < %s | FileCheck %s
+; RUN: llc -mattr=+bf16,+sme2,+fp8 --force-streaming < %s | FileCheck %s
+
+target triple = "aarch64-linux"
+
+define <vscale x 8 x bfloat> @cvt1_bf16(<vscale x 16 x i8> %s) {
+; CHECK-LABEL: cvt1_bf16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    bf1cvt z0.h, z0.b
+; CHECK-NEXT:    ret
+    %r = call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvt1.nxv8bf16(<vscale x 16 x i8> %s)
+    ret <vscale x 8 x bfloat> %r
+}
+
+define <vscale x 8 x bfloat> @cvt2_bf16(<vscale x 16 x i8> %s) {
+; CHECK-LABEL: cvt2_bf16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    bf2cvt z0.h, z0.b
+; CHECK-NEXT:    ret
+    %r = call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvt2.nxv8bf16(<vscale x 16 x i8> %s)
+    ret <vscale x 8 x bfloat> %r
+}
+
+define <vscale x 8 x bfloat> @cvtlt1_bf16(<vscale x 16 x i8> %s) {
+; CHECK-LABEL: cvtlt1_bf16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    bf1cvtlt z0.h, z0.b
+; CHECK-NEXT:    ret
+    %r = call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvtlt1.nxv8bf16(<vscale x 16 x i8> %s)
+    ret <vscale x 8 x bfloat> %r
+}
+
+define <vscale x 8 x bfloat> @cvtlt2_bf16(<vscale x 16 x i8> %s) {
+; CHECK-LABEL: cvtlt2_bf16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    bf2cvtlt z0.h, z0.b
+; CHECK-NEXT:    ret
+    %r = call <vscale x 8 x bfloat> @llvm.aarch64.sve.fp8.cvtlt2.nxv8bf16(<vscale x 16 x i8> %s)
+    ret <vscale x 8 x bfloat> %r
+}
+
+define <vscale x 8 x half> @cvt1_f16(<vscale x 16 x i8> %s) {
+; CHECK-LABEL: cvt1_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    f1cvt z0.h, z0.b
+; CHECK-NEXT:    ret
+    %r = call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvt1.nxv8f16(<vscale x 16 x i8> %s)
+    ret <vscale x 8 x half> %r
+}
+
+define <vscale x 8 x half> @cvt2_f16(<vscale x 16 x i8> %s) {
+; CHECK-LABEL: cvt2_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    f2cvt z0.h, z0.b
+; CHECK-NEXT:    ret
+    %r = call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvt2.nxv8f16(<vscale x 16 x i8> %s)
+    ret <vscale x 8 x half> %r
+}
+
+
+define <vscale x 8 x half> @cvtlt1_f16(<vscale x 16 x i8> %s) {
+; CHECK-LABEL: cvtlt1_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    f1cvtlt z0.h, z0.b
+; CHECK-NEXT:    ret
+    %r = call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvtlt1.nxv8f16(<vscale x 16 x i8> %s)
+    ret <vscale x 8 x half> %r
+}
+
+define <vscale x 8 x half> @cvtlt2_f16(<vscale x 16 x i8> %s) {
+; CHECK-LABEL: cvtlt2_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    f2cvtlt z0.h, z0.b
+; CHECK-NEXT:    ret
+    %r = call <vscale x 8 x half> @llvm.aarch64.sve.fp8.cvtlt2.nxv8f16(<vscale x 16 x i8> %s)
+    ret <vscale x 8 x half> %r
+}

>From 12582745c36d948bb37f717d6967f2d71ba372e8 Mon Sep 17 00:00:00 2001
From: Momchil Velikov <momchil.velikov at arm.com>
Date: Mon, 25 Nov 2024 09:47:41 +0000
Subject: [PATCH 2/2] FP8 CVTN

---
 clang/include/clang/Basic/arm_sve.td          |   7 ++
 .../fp8-intrinsics/acle_sve2_fp8_cvtn.c       | 101 ++++++++++++++++++
 .../aarch64-sve2-intrinsics/acle_sve2_fp8.c   |  11 +-
 llvm/include/llvm/IR/IntrinsicsAArch64.td     |  13 +++
 .../lib/Target/AArch64/AArch64SVEInstrInfo.td |   9 +-
 llvm/lib/Target/AArch64/SVEInstrFormats.td    |  37 ++++++-
 llvm/test/CodeGen/AArch64/fp8-sve-cvtn.ll     |  49 +++++++++
 7 files changed, 221 insertions(+), 6 deletions(-)
 create mode 100644 clang/test/CodeGen/AArch64/fp8-intrinsics/acle_sve2_fp8_cvtn.c
 create mode 100644 llvm/test/CodeGen/AArch64/fp8-sve-cvtn.ll

diff --git a/clang/include/clang/Basic/arm_sve.td b/clang/include/clang/Basic/arm_sve.td
index b9d8360843aa8e..efba2d4671d819 100644
--- a/clang/include/clang/Basic/arm_sve.td
+++ b/clang/include/clang/Basic/arm_sve.td
@@ -2456,4 +2456,11 @@ let SVETargetGuard = "sve2,fp8", SMETargetGuard = "sme2,fp8" in {
   // 8-bit floating-point convert to BFloat16/Float16 (top)
   def SVF1CVTLT : SInst<"svcvtlt1_{d}[_mf8]_fpm", "d~>", "bh", MergeNone, "aarch64_sve_fp8_cvtlt1", [VerifyRuntimeMode, SetsFPMR]>;
   def SVF2CVTLT : SInst<"svcvtlt2_{d}[_mf8]_fpm", "d~>", "bh", MergeNone, "aarch64_sve_fp8_cvtlt2", [VerifyRuntimeMode, SetsFPMR]>;
+
+  // BFloat16/Float16 convert, narrow and interleave to 8-bit floating-point
+  def SVFCVTN : SInst<"svcvtn_mf8[_{d}_x2]_fpm", "~2>", "bh", MergeNone, "aarch64_sve_fp8_cvtn", [VerifyRuntimeMode, SetsFPMR]>;
+
+  // Single-precision convert, narrow and interleave to 8-bit floating-point (top and bottom)
+  def SVFCVTNB : SInst<"svcvtnb_mf8[_f32_x2]_fpm", "~2>",  "f", MergeNone, "aarch64_sve_fp8_cvtnb", [VerifyRuntimeMode, SetsFPMR]>;
+  def SVFCVTNT : SInst<"svcvtnt_mf8[_f32_x2]_fpm", "~~2>", "f", MergeNone, "aarch64_sve_fp8_cvtnt", [VerifyRuntimeMode, SetsFPMR]>;
 }
diff --git a/clang/test/CodeGen/AArch64/fp8-intrinsics/acle_sve2_fp8_cvtn.c b/clang/test/CodeGen/AArch64/fp8-intrinsics/acle_sve2_fp8_cvtn.c
new file mode 100644
index 00000000000000..ed5b0ce02af4bd
--- /dev/null
+++ b/clang/test/CodeGen/AArch64/fp8-intrinsics/acle_sve2_fp8_cvtn.c
@@ -0,0 +1,101 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5
+// RUN: %clang_cc1        -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +sve2 -target-feature +fp8 -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s
+// RUN: %clang_cc1 -x c++ -triple aarch64-none-linux-gnu -target-feature +sme -target-feature +sme2 -target-feature +fp8 -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s -check-prefix=CHECK-CXX
+
+// RUN: %clang_cc1 -DSVE_OVERLOADED_FORMS        -triple aarch64-none-linux-gnu -target-feature +sme -target-feature +sme2 -target-feature +fp8 -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s
+// RUN: %clang_cc1 -DSVE_OVERLOADED_FORMS -x c++ -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +sve2 -target-feature +fp8 -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s -check-prefix=CHECK-CXX
+
+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +sve2 -target-feature +fp8 -S -disable-O0-optnone -Werror -Wall -o /dev/null %s
+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sme -target-feature +sme2 -target-feature +fp8 -S -disable-O0-optnone -Werror -Wall -o /dev/null %s
+
+// REQUIRES: aarch64-registered-target
+
+#ifdef __ARM_FEATURE_SME
+#include <arm_sme.h>
+#else
+#include <arm_sve.h>
+#endif
+
+#ifdef SVE_OVERLOADED_FORMS
+#define SVE_ACLE_FUNC(A1,A2_UNUSED,A3) A1##A3
+#else
+#define SVE_ACLE_FUNC(A1,A2,A3) A1##A2##A3
+#endif
+
+#ifdef __ARM_FEATURE_SME
+#define STREAMING __arm_streaming
+#else
+#define STREAMING
+#endif
+
+// CHECK-LABEL: define dso_local <vscale x 16 x i8> @test_svcvtn_f8_bf16(
+// CHECK-SAME: <vscale x 8 x bfloat> [[ZN_ZM_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_ZM_COERCE1:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtn.nxv8bf16(<vscale x 8 x bfloat> [[ZN_ZM_COERCE0]], <vscale x 8 x bfloat> [[ZN_ZM_COERCE1]])
+// CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 16 x i8> @_Z19test_svcvtn_f8_bf1614svbfloat16x2_tm(
+// CHECK-CXX-SAME: <vscale x 8 x bfloat> [[ZN_ZM_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_ZM_COERCE1:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtn.nxv8bf16(<vscale x 8 x bfloat> [[ZN_ZM_COERCE0]], <vscale x 8 x bfloat> [[ZN_ZM_COERCE1]])
+// CHECK-CXX-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+svmfloat8_t test_svcvtn_f8_bf16(svbfloat16x2_t zn_zm, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvtn_mf8,_bf16_x2,_fpm)(zn_zm, fpm);
+}
+
+// CHECK-LABEL: define dso_local <vscale x 16 x i8> @test_svcvtn_f8_f16(
+// CHECK-SAME: <vscale x 8 x half> [[ZN_ZM_COERCE0:%.*]], <vscale x 8 x half> [[ZN_ZM_COERCE1:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtn.nxv8f16(<vscale x 8 x half> [[ZN_ZM_COERCE0]], <vscale x 8 x half> [[ZN_ZM_COERCE1]])
+// CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 16 x i8> @_Z18test_svcvtn_f8_f1613svfloat16x2_tm(
+// CHECK-CXX-SAME: <vscale x 8 x half> [[ZN_ZM_COERCE0:%.*]], <vscale x 8 x half> [[ZN_ZM_COERCE1:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtn.nxv8f16(<vscale x 8 x half> [[ZN_ZM_COERCE0]], <vscale x 8 x half> [[ZN_ZM_COERCE1]])
+// CHECK-CXX-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+svmfloat8_t test_svcvtn_f8_f16(svfloat16x2_t zn_zm, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvtn_mf8,_f16_x2,_fpm)(zn_zm, fpm);
+}
+
+// CHECK-LABEL: define dso_local <vscale x 16 x i8> @test_svcvtnb_f8_f32(
+// CHECK-SAME: <vscale x 4 x float> [[ZN_ZM_COERCE0:%.*]], <vscale x 4 x float> [[ZN_ZM_COERCE1:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtnb.nxv4f32(<vscale x 4 x float> [[ZN_ZM_COERCE0]], <vscale x 4 x float> [[ZN_ZM_COERCE1]])
+// CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 16 x i8> @_Z19test_svcvtnb_f8_f3213svfloat32x2_tm(
+// CHECK-CXX-SAME: <vscale x 4 x float> [[ZN_ZM_COERCE0:%.*]], <vscale x 4 x float> [[ZN_ZM_COERCE1:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtnb.nxv4f32(<vscale x 4 x float> [[ZN_ZM_COERCE0]], <vscale x 4 x float> [[ZN_ZM_COERCE1]])
+// CHECK-CXX-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+svmfloat8_t test_svcvtnb_f8_f32(svfloat32x2_t zn_zm, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvtnb_mf8,_f32_x2,_fpm)(zn_zm, fpm);
+}
+
+// CHECK-LABEL: define dso_local <vscale x 16 x i8> @test_svcvtnt_f8_f32(
+// CHECK-SAME: <vscale x 16 x i8> [[ZD:%.*]], <vscale x 4 x float> [[ZN_ZM_COERCE0:%.*]], <vscale x 4 x float> [[ZN_ZM_COERCE1:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtnt.nxv4f32(<vscale x 16 x i8> [[ZD]], <vscale x 4 x float> [[ZN_ZM_COERCE0]], <vscale x 4 x float> [[ZN_ZM_COERCE1]])
+// CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+// CHECK-CXX-LABEL: define dso_local <vscale x 16 x i8> @_Z19test_svcvtnt_f8_f32u13__SVMfloat8_t13svfloat32x2_tm(
+// CHECK-CXX-SAME: <vscale x 16 x i8> [[ZD:%.*]], <vscale x 4 x float> [[ZN_ZM_COERCE0:%.*]], <vscale x 4 x float> [[ZN_ZM_COERCE1:%.*]], i64 noundef [[FPM:%.*]]) #[[ATTR0]] {
+// CHECK-CXX-NEXT:  [[ENTRY:.*:]]
+// CHECK-CXX-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPM]])
+// CHECK-CXX-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtnt.nxv4f32(<vscale x 16 x i8> [[ZD]], <vscale x 4 x float> [[ZN_ZM_COERCE0]], <vscale x 4 x float> [[ZN_ZM_COERCE1]])
+// CHECK-CXX-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+svmfloat8_t test_svcvtnt_f8_f32(svmfloat8_t zd, svfloat32x2_t zn_zm, fpm_t fpm) STREAMING {
+  return SVE_ACLE_FUNC(svcvtnt_mf8,_f32_x2,_fpm)(zd, zn_zm, fpm);
+}
diff --git a/clang/test/Sema/aarch64-sve2-intrinsics/acle_sve2_fp8.c b/clang/test/Sema/aarch64-sve2-intrinsics/acle_sve2_fp8.c
index aafd42f798d935..803bdd5f40183a 100644
--- a/clang/test/Sema/aarch64-sve2-intrinsics/acle_sve2_fp8.c
+++ b/clang/test/Sema/aarch64-sve2-intrinsics/acle_sve2_fp8.c
@@ -1,6 +1,6 @@
 // REQUIRES: aarch64-registered-target
 
-// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sve -verify -emit-llvm %s
+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sve -target-feature +bf16 -verify -emit-llvm %s
 
 #include <arm_sve.h>
 
@@ -21,4 +21,13 @@ void test_features(svmfloat8_t zn, fpm_t fpm) {
   // expected-error at -1 {{'svcvtlt1_f16_mf8_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
   svcvtlt2_f16_mf8_fpm(zn, fpm);
   // expected-error at -1 {{'svcvtlt2_f16_mf8_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+
+  svcvtn_mf8_bf16_x2_fpm(svcreate2(svundef_bf16(), svundef_bf16()), fpm);
+  // expected-error at -1 {{'svcvtn_mf8_bf16_x2_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+  svcvtn_mf8_f16_x2_fpm(svcreate2(svundef_f16(), svundef_f16()), fpm);
+  // expected-error at -1 {{'svcvtn_mf8_f16_x2_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+  svcvtnb_mf8_f32_x2_fpm(svcreate2(svundef_f32(), svundef_f32()), fpm);
+  // expected-error at -1 {{'svcvtnb_mf8_f32_x2_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
+  svcvtnt_mf8_f32_x2_fpm(zn, svcreate2(svundef_f32(), svundef_f32()), fpm);
+  // expected-error at -1 {{'svcvtnt_mf8_f32_x2_fpm' needs target feature (sve,sve2,fp8)|(sme,sme2,fp8)}}
 }
diff --git a/llvm/include/llvm/IR/IntrinsicsAArch64.td b/llvm/include/llvm/IR/IntrinsicsAArch64.td
index 13bc5e08d2756f..622dbd9e5c97db 100644
--- a/llvm/include/llvm/IR/IntrinsicsAArch64.td
+++ b/llvm/include/llvm/IR/IntrinsicsAArch64.td
@@ -3880,4 +3880,17 @@ def int_aarch64_sve_fp8_cvt1   : SVE2_FP8_Cvt;
 def int_aarch64_sve_fp8_cvt2   : SVE2_FP8_Cvt;
 def int_aarch64_sve_fp8_cvtlt1 : SVE2_FP8_Cvt;
 def int_aarch64_sve_fp8_cvtlt2 : SVE2_FP8_Cvt;
+
+class SVE2_FP8_Narrow_Cvt
+  : DefaultAttrsIntrinsic<[llvm_nxv16i8_ty],
+                          [llvm_anyvector_ty, LLVMMatchType<0>],
+                          [IntrReadMem, IntrInaccessibleMemOnly]>;
+
+def int_aarch64_sve_fp8_cvtn  : SVE2_FP8_Narrow_Cvt;
+def int_aarch64_sve_fp8_cvtnb : SVE2_FP8_Narrow_Cvt;
+
+def int_aarch64_sve_fp8_cvtnt
+  : DefaultAttrsIntrinsic<[llvm_nxv16i8_ty],
+                          [llvm_nxv16i8_ty, llvm_anyvector_ty, LLVMMatchType<0>],
+                          [IntrReadMem, IntrInaccessibleMemOnly]>;
 }
diff --git a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
index 5365a00f3f42ee..47d6418db549d7 100644
--- a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
@@ -4379,10 +4379,11 @@ defm BF1CVTLT_ZZ  : sve2_fp8_cvt_single<0b1, 0b10, "bf1cvtlt", nxv8bf16, int_aar
 defm BF2CVTLT_ZZ  : sve2_fp8_cvt_single<0b1, 0b11, "bf2cvtlt", nxv8bf16, int_aarch64_sve_fp8_cvtlt2>;
 
 // FP8 downconvert
-defm FCVTN_Z2Z_HtoB  : sve2_fp8_down_cvt_single<0b00, "fcvtn", ZZ_h_mul_r>;
-defm FCVTNB_Z2Z_StoB : sve2_fp8_down_cvt_single<0b01, "fcvtnb", ZZ_s_mul_r>;
-defm BFCVTN_Z2Z_HtoB : sve2_fp8_down_cvt_single<0b10, "bfcvtn", ZZ_h_mul_r>;
-defm FCVTNT_Z2Z_StoB : sve2_fp8_down_cvt_single<0b11, "fcvtnt", ZZ_s_mul_r>;
+defm FCVTN_Z2Z_HtoB  : sve2_fp8_down_cvt_single<0b00, "fcvtn",  ZZ_h_mul_r, nxv8f16,  int_aarch64_sve_fp8_cvtn>;
+defm FCVTNB_Z2Z_StoB : sve2_fp8_down_cvt_single<0b01, "fcvtnb", ZZ_s_mul_r, nxv4f32,  int_aarch64_sve_fp8_cvtnb>;
+defm BFCVTN_Z2Z_HtoB : sve2_fp8_down_cvt_single<0b10, "bfcvtn", ZZ_h_mul_r, nxv8bf16, int_aarch64_sve_fp8_cvtn>;
+
+defm FCVTNT_Z2Z_StoB : sve2_fp8_down_cvt_single_top<0b11, "fcvtnt", ZZ_s_mul_r, nxv4f32,  int_aarch64_sve_fp8_cvtnt>;
 } // End HasSVE2orSME2, HasFP8
 
 let Predicates = [HasSVE2orSME2, HasFAMINMAX] in {
diff --git a/llvm/lib/Target/AArch64/SVEInstrFormats.td b/llvm/lib/Target/AArch64/SVEInstrFormats.td
index 58770bf6e274d3..95cb5ff8c4a03b 100644
--- a/llvm/lib/Target/AArch64/SVEInstrFormats.td
+++ b/llvm/lib/Target/AArch64/SVEInstrFormats.td
@@ -10758,10 +10758,45 @@ class sve2_fp8_down_cvt_single<bits<2> opc, string mnemonic,
   let Inst{5} = 0b0;
   let Inst{4-0} = Zd;
   let Uses = [FPMR, FPCR];
+
+  let mayLoad  = 1;
+  let mayStore = 0;
 }
 
-multiclass sve2_fp8_down_cvt_single<bits<2> opc, string mnemonic, RegisterOperand src> {
+multiclass sve2_fp8_down_cvt_single<bits<2> opc, string mnemonic, RegisterOperand src,
+                                    ValueType ty, SDPatternOperator op> {
   def NAME : sve2_fp8_down_cvt_single<opc, mnemonic, ZPR8, src>;
+
+  def : Pat<(nxv16i8 (op ty:$Zn1, ty:$Zn2)),
+            (!cast<Instruction>(NAME) (REG_SEQUENCE ZPR2Mul2, $Zn1, zsub0, $Zn2, zsub1))>;
+}
+
+class sve2_fp8_down_cvt_single_top<bits<2> opc, string mnemonic, RegisterOperand src_ty>
+  : I<(outs ZPR8:$Zd), (ins ZPR8:$_Zd, src_ty:$Zn), mnemonic, "\t$Zd, $Zn","", []>, Sched<[]> {
+  bits<5> Zd;
+  bits<4> Zn;
+
+  let Inst{31-12} = 0b01100101000010100011;
+  let Inst{11-10} = opc;
+  let Inst{9-6}   = Zn;
+  let Inst{5}     = 0b0;
+  let Inst{4-0}   = Zd;
+
+  let Constraints = "$Zd = $_Zd";
+  let DestructiveInstType = DestructiveOther;
+  let ElementSize         = ZPR8.ElementSize;
+  
+  let Uses     = [FPMR, FPCR];
+  let mayLoad  = 1;
+  let mayStore = 0;
+}
+
+multiclass sve2_fp8_down_cvt_single_top<bits<2> opc, string mnemonic, RegisterOperand src_ty,
+                                        ValueType ty, SDPatternOperator op> {
+  def NAME : sve2_fp8_down_cvt_single_top<opc, mnemonic, src_ty>;
+
+  def : Pat<(nxv16i8 (op nxv16i8:$Zd, ty:$Zn1, ty:$Zn2)),
+            (!cast<Instruction>(NAME) $Zd, (REG_SEQUENCE ZPR2Mul2, $Zn1, zsub0, $Zn2, zsub1))>;
 }
 
 // FP8 Widening Multiply-Add Long - Indexed Group
diff --git a/llvm/test/CodeGen/AArch64/fp8-sve-cvtn.ll b/llvm/test/CodeGen/AArch64/fp8-sve-cvtn.ll
new file mode 100644
index 00000000000000..7daeafc7916043
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/fp8-sve-cvtn.ll
@@ -0,0 +1,49 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mattr=+bf16,+sve2,+fp8 < %s | FileCheck %s
+; RUN: llc -mattr=+bf16,+sme2,+fp8 --force-streaming < %s | FileCheck %s
+
+target triple = "aarch64-linux"
+
+define <vscale x 16 x i8> @cvtn_bf16(<vscale x 8 x bfloat> %s1, <vscale x 8 x bfloat> %s2) {
+; CHECK-LABEL: cvtn_bf16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z1
+; CHECK-NEXT:    // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z1
+; CHECK-NEXT:    bfcvtn z0.b, { z0.h, z1.h }
+; CHECK-NEXT:    ret
+    %r = call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtn.nxv8bf16(<vscale x 8 x bfloat> %s1, <vscale x 8 x bfloat> %s2)
+    ret <vscale x 16 x i8> %r
+}
+
+define <vscale x 16 x i8> @cvtn_f16(<vscale x 8 x half> %s1, <vscale x 8 x half> %s2) {
+; CHECK-LABEL: cvtn_f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z1
+; CHECK-NEXT:    // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z1
+; CHECK-NEXT:    fcvtn z0.b, { z0.h, z1.h }
+; CHECK-NEXT:    ret
+    %r = call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtn.nxv8f16(<vscale x 8 x half> %s1, <vscale x 8 x half> %s2)
+    ret <vscale x 16 x i8> %r
+}
+
+define <vscale x 16 x i8> @cvtnb_f32(<vscale x 4 x float> %s1, <vscale x 4 x float> %s2) {
+; CHECK-LABEL: cvtnb_f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z1
+; CHECK-NEXT:    // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z1
+; CHECK-NEXT:    fcvtnb z0.b, { z0.s, z1.s }
+; CHECK-NEXT:    ret
+    %r = call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtnb.nxv4f32(<vscale x 4 x float> %s1, <vscale x 4 x float> %s2)
+    ret <vscale x 16 x i8> %r
+}
+
+define <vscale x 16 x i8> @cvtnt_f32(<vscale x 16 x i8> %d, <vscale x 4 x float> %s1, <vscale x 4 x float> %s2) {
+; CHECK-LABEL: cvtnt_f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z3.d, z2.d
+; CHECK-NEXT:    mov z2.d, z1.d
+; CHECK-NEXT:    fcvtnt z0.b, { z2.s, z3.s }
+; CHECK-NEXT:    ret
+    %r = call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtnt.nxv4f32(<vscale x 16 x i8> %d, <vscale x 4 x float> %s1, <vscale x 4 x float> %s2)
+    ret <vscale x 16 x i8> %r
+}



More information about the cfe-commits mailing list