[SveEmitter] Builtins for SVE matrix multiply `mmla`.

Summary:
Guarded by __ARM_FEATURE_SVE_MATMUL_INT8:

* svmmla_u32
* svmmla_s32
* svusmmla_s32

Guarded by __ARM_FEATURE_SVE_MATMUL_FP32:

* svmmla_f32

Guarded by __ARM_FEATURE_SVE_MATMUL_FP64:

* svmmla_f64

Reviewers: sdesmalen, kmclaughlin, efriedma, rengolin

Subscribers: tschuett, kristof.beyls, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D79639

clang/include/clang/Basic/arm_sve.td

@@ -69,6 +69,7 @@
 // R: scalar of 1/2 width element type (splat to vector type)
 // r: scalar of 1/4 width element type (splat to vector type)
 // e: 1/2 width unsigned elements, 2x element count
+// b: 1/4 width unsigned elements, 4x element count
 // h: 1/2 width elements, 2x element count
 // q: 1/4 width elements, 4x element count
 // o: 4x width elements, 1/4 element count
@@ -1235,6 +1236,21 @@ def SVQINCP_N_S64 : SInst<"svqincp[_n_s64]_{d}", "llP", "PcPsPiPl", MergeNone, "
 def SVQINCP_N_U32 : SInst<"svqincp[_n_u32]_{d}", "mmP", "PcPsPiPl", MergeNone, "aarch64_sve_uqincp_n32">;
 def SVQINCP_N_U64 : SInst<"svqincp[_n_u64]_{d}", "nnP", "PcPsPiPl", MergeNone, "aarch64_sve_uqincp_n64">;
 
+let ArchGuard = "defined(__ARM_FEATURE_SVE_MATMUL_INT8)" in {
+def SVMLLA_S32   : SInst<"svmmla[_s32]",   "ddqq","i",  MergeNone, "aarch64_sve_smmla">;
+def SVMLLA_U32   : SInst<"svmmla[_u32]",   "ddqq","Ui", MergeNone, "aarch64_sve_ummla">;
+def SVUSMLLA_S32 : SInst<"svusmmla[_s32]", "ddbq","i",  MergeNone, "aarch64_sve_usmmla">;
+}
+
+let ArchGuard = "defined(__ARM_FEATURE_SVE_MATMUL_FP32)" in {
+def SVMLLA_F32 : SInst<"svmmla[_f32]", "dddd","f", MergeNone, "aarch64_sve_fmmla">;
+}
+
+let ArchGuard = "defined(__ARM_FEATURE_SVE_MATMUL_FP64)" in {
+def SVMLLA_F64 : SInst<"svmmla[_f64]", "dddd","d", MergeNone, "aarch64_sve_fmmla">;
+}
+
+
 ////////////////////////////////////////////////////////////////////////////////
 // SVE2 WhileGE/GT
 let ArchGuard = "defined(__ARM_FEATURE_SVE2)" in {

clang/test/CodeGen/aarch64-sve-intrinsics/acle_sve_matmul_fp32.c

@@ -0,0 +1,18 @@
+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_MATMUL_FP32 -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_MATMUL_FP32 -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_svmmla_f32(svfloat32_t x, svfloat32_t y, svfloat32_t z) {
+  // CHECK-LABEL: test_svmmla_f32
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x float> @llvm.aarch64.sve.fmmla.nxv4f32(<vscale x 4 x float> %x, <vscale x 4 x float> %y, <vscale x 4 x float> %z)
+  // CHECK: ret <vscale x 4 x float> %[[RET]]
+  return SVE_ACLE_FUNC(svmmla,_f32,,)(x, y, z);
+}

clang/test/CodeGen/aarch64-sve-intrinsics/acle_sve_matmul_fp64.c

@@ -0,0 +1,18 @@
+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_MATMUL_FP64 -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_MATMUL_FP64 -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
+
+svfloat64_t test_svmmla_f64(svfloat64_t x, svfloat64_t y, svfloat64_t z) {
+  // CHECK-LABEL: test_svmmla_f64
+  // CHECK: %[[RET:.*]] = call <vscale x 2 x double> @llvm.aarch64.sve.fmmla.nxv2f64(<vscale x 2 x double> %x, <vscale x 2 x double> %y, <vscale x 2 x double> %z)
+  // CHECK: ret <vscale x 2 x double> %[[RET]]
+  return SVE_ACLE_FUNC(svmmla,_f64,,)(x, y, z);
+}

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

@@ -0,0 +1,32 @@
+// RUN: %clang_cc1 -D__ARM_FEATURE_SVE -D__ARM_FEATURE_SVE_MATMUL_INT8 -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_MATMUL_INT8 -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
+
+svint32_t test_svmmla_s32(svint32_t x, svint8_t y, svint8_t z) {
+  // CHECK-LABEL: test_svmmla_s32
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x i32> @llvm.aarch64.sve.smmla.nxv4i32(<vscale x 4 x i32> %x, <vscale x 16 x i8> %y, <vscale x 16 x i8> %z)
+  // CHECK: ret <vscale x 4 x i32> %[[RET]]
+  return SVE_ACLE_FUNC(svmmla,_s32,,)(x, y, z);
+}
+
+svuint32_t test_svmmla_u32(svuint32_t x, svuint8_t y, svuint8_t z) {
+  // CHECK-LABEL: test_svmmla_u32
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x i32> @llvm.aarch64.sve.ummla.nxv4i32(<vscale x 4 x i32> %x, <vscale x 16 x i8> %y, <vscale x 16 x i8> %z)
+  // CHECK: ret <vscale x 4 x i32> %[[RET]]
+  return SVE_ACLE_FUNC(svmmla,_u32,,)(x, y, z);
+}
+
+svint32_t test_svusmmla_s32(svint32_t x, svuint8_t y, svint8_t z) {
+  // CHECK-LABEL: test_svusmmla_s32
+  // CHECK: %[[RET:.*]] = call <vscale x 4 x i32> @llvm.aarch64.sve.usmmla.nxv4i32(<vscale x 4 x i32> %x, <vscale x 16 x i8> %y, <vscale x 16 x i8> %z)
+  // CHECK: ret <vscale x 4 x i32> %[[RET]]
+  return SVE_ACLE_FUNC(svusmmla,_s32,,)(x, y, z);
+}

clang/utils/TableGen/SveEmitter.cpp

@@ -513,6 +513,11 @@ void SVEType::applyModifier(char Mod) {
   case 'q':
     ElementBitwidth /= 4;
     break;
+  case 'b':
+    Signed = false;
+    Float = false;
+    ElementBitwidth /= 4;
+    break;
   case 'o':
     ElementBitwidth *= 4;
     break;