[AArch64][SME2] Add LLVM IR intrinsics for multi-multi dots

Adds intrinsics for the following SME2 instructions:

* sdot (multi-multi, 2 and 4 vectors, 32-bit and 64-bit ZA)
* udot (multi-multi, 2 and 4 vectors, 32-bit and 64-bit ZA)
* usdot (multi-multi, 2 and 4 vectors)
* fdot (multi-multi, 2 and 4 vectors)
* bfdot (multi-multi, 2 and 4 vectors)

NOTE: These intrinsics are still in development and are subject to future changes.

Differential Revision: https://reviews.llvm.org/D142478

diff --git a/llvm/include/llvm/IR/IntrinsicsAArch64.td b/llvm/include/llvm/IR/IntrinsicsAArch64.td
index 996ff97..21ff48a 100644 (file)
--- a/llvm/include/llvm/IR/IntrinsicsAArch64.td
+++ b/llvm/include/llvm/IR/IntrinsicsAArch64.td
@@ -3170,6 +3170,9 @@ let TargetPrefix = "aarch64" in {
     foreach sz = ["za32", "za64"] in {
       def int_aarch64_sme_ # ty # dot_single_ # sz # _vg1x2 : SME2_Matrix_ArrayVector_VG2_Multi_Single_Intrinsic;
       def int_aarch64_sme_ # ty # dot_single_ # sz # _vg1x4 : SME2_Matrix_ArrayVector_VG4_Multi_Single_Intrinsic;
+
+      def int_aarch64_sme_ # ty # dot_ # sz # _vg1x2 : SME2_Matrix_ArrayVector_VG2_Multi_Multi_Intrinsic;
+      def int_aarch64_sme_ # ty # dot_ # sz # _vg1x4 : SME2_Matrix_ArrayVector_VG4_Multi_Multi_Intrinsic;
     }
   }
 
@@ -3178,7 +3181,13 @@ let TargetPrefix = "aarch64" in {
     def int_aarch64_sme_ # ty # dot_single_za32_vg1x4 : SME2_Matrix_ArrayVector_VG4_Multi_Single_Intrinsic;
   }
 
+  def int_aarch64_sme_usdot_za32_vg1x2 : SME2_Matrix_ArrayVector_VG2_Multi_Multi_Intrinsic;
+  def int_aarch64_sme_usdot_za32_vg1x4 : SME2_Matrix_ArrayVector_VG4_Multi_Multi_Intrinsic;
+
   // Multi-vector half-precision or bfloat floating-point dot-product
   def int_aarch64_sme_fdot_single_za32_vg1x2 : SME2_Matrix_ArrayVector_VG2_Multi_Single_Intrinsic;
   def int_aarch64_sme_fdot_single_za32_vg1x4 : SME2_Matrix_ArrayVector_VG4_Multi_Single_Intrinsic;
+
+  def int_aarch64_sme_fdot_za32_vg1x2 : SME2_Matrix_ArrayVector_VG2_Multi_Multi_Intrinsic;
+  def int_aarch64_sme_fdot_za32_vg1x4 : SME2_Matrix_ArrayVector_VG4_Multi_Multi_Intrinsic;
 }

diff --git a/llvm/lib/Target/AArch64/AArch64SMEInstrInfo.td b/llvm/lib/Target/AArch64/AArch64SMEInstrInfo.td
index 52484a1..acfdf51 100644 (file)
--- a/llvm/lib/Target/AArch64/AArch64SMEInstrInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64SMEInstrInfo.td
@@ -450,15 +450,15 @@ defm FDOT_VG2_M2ZZI_HtoS : sme2_multi_vec_array_vg2_index_32b<"fdot", 0b1001, ZZ
 defm FDOT_VG4_M4ZZI_HtoS : sme2_multi_vec_array_vg4_index_32b<"fdot", 0b1001, ZZZZ_h_mul_r, ZPR4b16, nxv8f16, null_frag>;
 defm FDOT_VG2_M2ZZ_HtoS  : sme2_dot_mla_add_sub_array_vg2_single<"fdot", 0b0010000, MatrixOp32, ZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fdot_single_za32_vg1x2>;
 defm FDOT_VG4_M4ZZ_HtoS  : sme2_dot_mla_add_sub_array_vg4_single<"fdot", 0b0110000, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8f16, int_aarch64_sme_fdot_single_za32_vg1x4>;
-defm FDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"fdot",  0b010000, MatrixOp32, ZZ_h_mul_r, nxv8f16, null_frag>;
-defm FDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"fdot",  0b010000, MatrixOp32, ZZZZ_h_mul_r, nxv8f16, null_frag>;
+defm FDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"fdot",  0b010000, MatrixOp32, ZZ_h_mul_r, nxv8f16, int_aarch64_sme_fdot_za32_vg1x2>;
+defm FDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"fdot",  0b010000, MatrixOp32, ZZZZ_h_mul_r, nxv8f16, int_aarch64_sme_fdot_za32_vg1x4>;
 
 defm BFDOT_VG2_M2ZZI_HtoS : sme2_multi_vec_array_vg2_index_32b<"bfdot", 0b1011, ZZ_h_mul_r, ZPR4b16, nxv8bf16, null_frag>;
 defm BFDOT_VG4_M4ZZI_HtoS : sme2_multi_vec_array_vg4_index_32b<"bfdot", 0b1011, ZZZZ_h_mul_r, ZPR4b16, nxv8bf16, null_frag>;
 defm BFDOT_VG2_M2ZZ_HtoS  : sme2_dot_mla_add_sub_array_vg2_single<"bfdot", 0b0010010, MatrixOp32, ZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fdot_single_za32_vg1x2>;
 defm BFDOT_VG4_M4ZZ_HtoS  : sme2_dot_mla_add_sub_array_vg4_single<"bfdot", 0b0110010, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8bf16, int_aarch64_sme_fdot_single_za32_vg1x4>;
-defm BFDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"bfdot",  0b010010, MatrixOp32, ZZ_h_mul_r, nxv8bf16, null_frag>;
-defm BFDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"bfdot",  0b010010, MatrixOp32, ZZZZ_h_mul_r, nxv8bf16, null_frag>;
+defm BFDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"bfdot",  0b010010, MatrixOp32, ZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fdot_za32_vg1x2>;
+defm BFDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"bfdot",  0b010010, MatrixOp32, ZZZZ_h_mul_r, nxv8bf16, int_aarch64_sme_fdot_za32_vg1x4>;
 
 defm BFVDOT_VG2_M2ZZI_HtoS : sme2_multi_vec_array_vg2_index_32b<"bfvdot", 0b0011, ZZ_h_mul_r, ZPR4b16, nxv8bf16, int_aarch64_sme_fvdot_lane_za32_vg1x2>;
 
@@ -470,12 +470,12 @@ defm SDOT_VG4_M4ZZI_HToS : sme2_multi_vec_array_vg4_index_32b<"sdot", 0b1000, ZZ
 defm SDOT_VG4_M4ZZI_BToS : sme2_multi_vec_array_vg4_index_32b<"sdot", 0b1100, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, null_frag>;
 defm SDOT_VG2_M2ZZ_HtoS  : sme2_dot_mla_add_sub_array_vg2_single<"sdot", 0b1010101, MatrixOp32, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_single_za32_vg1x2>;
 defm SDOT_VG4_M4ZZ_HtoS  : sme2_dot_mla_add_sub_array_vg4_single<"sdot", 0b1110101, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_single_za32_vg1x4>;
-defm SDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"sdot", 0b110101, MatrixOp32, ZZ_h_mul_r, nxv8i16, null_frag>;
-defm SDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"sdot", 0b110101, MatrixOp32, ZZZZ_h_mul_r, nxv8i16, null_frag>;
+defm SDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"sdot", 0b110101, MatrixOp32, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_sdot_za32_vg1x2>;
+defm SDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"sdot", 0b110101, MatrixOp32, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_sdot_za32_vg1x4>;
 defm SDOT_VG2_M2ZZ_BtoS : sme2_dot_mla_add_sub_array_vg2_single<"sdot", 0b0010100, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_sdot_single_za32_vg1x2>;
 defm SDOT_VG4_M4ZZ_BtoS : sme2_dot_mla_add_sub_array_vg4_single<"sdot", 0b0110100, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_sdot_single_za32_vg1x4>;
-defm SDOT_VG2_M2Z2Z_BtoS : sme2_dot_mla_add_sub_array_vg2_multi<"sdot", 0b010100, MatrixOp32, ZZ_b_mul_r, nxv16i8, null_frag>;
-defm SDOT_VG4_M4Z4Z_BtoS : sme2_dot_mla_add_sub_array_vg4_multi<"sdot", 0b010100, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, null_frag>;
+defm SDOT_VG2_M2Z2Z_BtoS : sme2_dot_mla_add_sub_array_vg2_multi<"sdot", 0b010100, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_sdot_za32_vg1x2>;
+defm SDOT_VG4_M4Z4Z_BtoS : sme2_dot_mla_add_sub_array_vg4_multi<"sdot", 0b010100, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_sdot_za32_vg1x4>;
 
 defm SUDOT_VG2_M2ZZI_BToS : sme2_multi_vec_array_vg2_index_32b<"sudot", 0b1111, ZZ_b_mul_r, ZPR4b8, nxv16i8, null_frag>;
 defm SUDOT_VG4_M4ZZI_BToS : sme2_multi_vec_array_vg4_index_32b<"sudot", 0b1111, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, null_frag>;
@@ -493,19 +493,19 @@ defm UDOT_VG4_M4ZZI_BtoS : sme2_multi_vec_array_vg4_index_32b<"udot", 0b1110, ZZ
 defm UDOT_VG4_M4ZZI_HToS : sme2_multi_vec_array_vg4_index_32b<"udot", 0b1010, ZZZZ_h_mul_r, ZPR4b16, nxv8i16, null_frag>;
 defm UDOT_VG2_M2ZZ_HtoS  : sme2_dot_mla_add_sub_array_vg2_single<"udot", 0b1010111, MatrixOp32, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_udot_single_za32_vg1x2>;
 defm UDOT_VG4_M4ZZ_HtoS  : sme2_dot_mla_add_sub_array_vg4_single<"udot", 0b1110111, MatrixOp32, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_udot_single_za32_vg1x4>;
-defm UDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"udot", 0b110111, MatrixOp32, ZZ_h_mul_r, nxv8i16, null_frag>;
-defm UDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"udot", 0b110111, MatrixOp32, ZZZZ_h_mul_r, nxv8i16, null_frag>;
+defm UDOT_VG2_M2Z2Z_HtoS : sme2_dot_mla_add_sub_array_vg2_multi<"udot", 0b110111, MatrixOp32, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_udot_za32_vg1x2>;
+defm UDOT_VG4_M4Z4Z_HtoS : sme2_dot_mla_add_sub_array_vg4_multi<"udot", 0b110111, MatrixOp32, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_udot_za32_vg1x4>;
 defm UDOT_VG2_M2ZZ_BtoS : sme2_dot_mla_add_sub_array_vg2_single<"udot", 0b0010110, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_udot_single_za32_vg1x2>;
 defm UDOT_VG4_M4ZZ_BtoS : sme2_dot_mla_add_sub_array_vg4_single<"udot", 0b0110110, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_udot_single_za32_vg1x4>;
-defm UDOT_VG2_M2Z2Z_BtoS : sme2_dot_mla_add_sub_array_vg2_multi<"udot", 0b010110, MatrixOp32, ZZ_b_mul_r, nxv16i8, null_frag>;
-defm UDOT_VG4_M4Z4Z_BtoS : sme2_dot_mla_add_sub_array_vg4_multi<"udot", 0b010110, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, null_frag>;
+defm UDOT_VG2_M2Z2Z_BtoS : sme2_dot_mla_add_sub_array_vg2_multi<"udot", 0b010110, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_udot_za32_vg1x2>;
+defm UDOT_VG4_M4Z4Z_BtoS : sme2_dot_mla_add_sub_array_vg4_multi<"udot", 0b010110, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_udot_za32_vg1x4>;
 
 defm USDOT_VG2_M2ZZI_BToS: sme2_multi_vec_array_vg2_index_32b<"usdot", 0b1101, ZZ_b_mul_r, ZPR4b8, nxv16i8, null_frag>;
 defm USDOT_VG4_M4ZZI_BToS: sme2_multi_vec_array_vg4_index_32b<"usdot", 0b1101, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, null_frag>;
 defm USDOT_VG2_M2ZZ_BToS : sme2_dot_mla_add_sub_array_vg2_single<"usdot",  0b0010101, MatrixOp32, ZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_usdot_single_za32_vg1x2>;
 defm USDOT_VG4_M4ZZ_BToS  : sme2_dot_mla_add_sub_array_vg4_single<"usdot", 0b0110101, MatrixOp32, ZZZZ_b, ZPR4b8, nxv16i8, int_aarch64_sme_usdot_single_za32_vg1x4>;
-defm USDOT_VG2_M2Z2Z_BToS : sme2_dot_mla_add_sub_array_vg2_multi<"usdot", 0b010101, MatrixOp32, ZZ_b_mul_r, nxv16i8, null_frag>;
-defm USDOT_VG4_M4Z4Z_BToS : sme2_dot_mla_add_sub_array_vg4_multi<"usdot", 0b010101, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, null_frag>;
+defm USDOT_VG2_M2Z2Z_BToS : sme2_dot_mla_add_sub_array_vg2_multi<"usdot", 0b010101, MatrixOp32, ZZ_b_mul_r, nxv16i8, int_aarch64_sme_usdot_za32_vg1x2>;
+defm USDOT_VG4_M4Z4Z_BToS : sme2_dot_mla_add_sub_array_vg4_multi<"usdot", 0b010101, MatrixOp32, ZZZZ_b_mul_r, nxv16i8, int_aarch64_sme_usdot_za32_vg1x4>;
 
 defm USVDOT_VG4_M4ZZI_BToS : sme2_multi_vec_array_vg4_index_32b<"usvdot", 0b0101, ZZZZ_b_mul_r, ZPR4b8, nxv16i8, int_aarch64_sme_usvdot_lane_za32_vg1x4>;
 
@@ -725,8 +725,8 @@ defm SDOT_VG2_M2ZZI_HtoD : sme2_multi_vec_array_vg2_index_64b<"sdot", 0b01, ZZ_h
 defm SDOT_VG4_M4ZZI_HtoD : sme2_multi_vec_array_vg4_index_64b<"sdot", 0b001, ZZZZ_h_mul_r, ZPR4b16, nxv8i16, null_frag>;
 defm SDOT_VG2_M2ZZ_HtoD : sme2_dot_mla_add_sub_array_vg2_single<"sdot", 0b1010100, MatrixOp64, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_single_za64_vg1x2>;
 defm SDOT_VG4_M4ZZ_HtoD : sme2_dot_mla_add_sub_array_vg4_single<"sdot", 0b1110100, MatrixOp64, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_sdot_single_za64_vg1x4>;
-defm SDOT_VG2_M2Z2Z_HtoD : sme2_dot_mla_add_sub_array_vg2_multi<"sdot", 0b110100, MatrixOp64, ZZ_h_mul_r, nxv8i16, null_frag>;
-defm SDOT_VG4_M4Z4Z_HtoD : sme2_dot_mla_add_sub_array_vg4_multi<"sdot", 0b110100, MatrixOp64, ZZZZ_h_mul_r, nxv8i16, null_frag>;
+defm SDOT_VG2_M2Z2Z_HtoD : sme2_dot_mla_add_sub_array_vg2_multi<"sdot", 0b110100, MatrixOp64, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_sdot_za64_vg1x2>;
+defm SDOT_VG4_M4Z4Z_HtoD : sme2_dot_mla_add_sub_array_vg4_multi<"sdot", 0b110100, MatrixOp64, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_sdot_za64_vg1x4>;
 
 defm SVDOT_VG4_M4ZZI_HtoD : sme2_multi_vec_array_vg4_index_64b<"svdot", 0b101, ZZZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_svdot_lane_za64_vg1x4>;
 
@@ -734,8 +734,8 @@ defm UDOT_VG2_M2ZZI_HtoD : sme2_multi_vec_array_vg2_index_64b<"udot", 0b11, ZZ_h
 defm UDOT_VG4_M4ZZI_HtoD : sme2_multi_vec_array_vg4_index_64b<"udot", 0b011, ZZZZ_h_mul_r, ZPR4b16, nxv8i16, null_frag>;
 defm UDOT_VG2_M2ZZ_HtoD : sme2_dot_mla_add_sub_array_vg2_single<"udot", 0b1010110, MatrixOp64, ZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_udot_single_za64_vg1x2>;
 defm UDOT_VG4_M4ZZ_HtoD : sme2_dot_mla_add_sub_array_vg4_single<"udot", 0b1110110, MatrixOp64, ZZZZ_h, ZPR4b16, nxv8i16, int_aarch64_sme_udot_single_za64_vg1x4>;
-defm UDOT_VG2_M2Z2Z_HtoD : sme2_dot_mla_add_sub_array_vg2_multi<"udot", 0b110110, MatrixOp64, ZZ_h_mul_r, nxv8i16, null_frag>;
-defm UDOT_VG4_M4Z4Z_HtoD : sme2_dot_mla_add_sub_array_vg4_multi<"udot", 0b110110, MatrixOp64, ZZZZ_h_mul_r, nxv8i16, null_frag>;
+defm UDOT_VG2_M2Z2Z_HtoD : sme2_dot_mla_add_sub_array_vg2_multi<"udot", 0b110110, MatrixOp64, ZZ_h_mul_r, nxv8i16, int_aarch64_sme_udot_za64_vg1x2>;
+defm UDOT_VG4_M4Z4Z_HtoD : sme2_dot_mla_add_sub_array_vg4_multi<"udot", 0b110110, MatrixOp64, ZZZZ_h_mul_r, nxv8i16, int_aarch64_sme_udot_za64_vg1x4>;
 
 defm UVDOT_VG4_M4ZZI_HtoD : sme2_multi_vec_array_vg4_index_64b<"uvdot", 0b111, ZZZZ_h_mul_r, ZPR4b16, nxv8i16, int_aarch64_sme_uvdot_lane_za64_vg1x4>;
 

diff --git a/llvm/test/CodeGen/AArch64/sme2-intrinsics-fp-dots.ll b/llvm/test/CodeGen/AArch64/sme2-intrinsics-fp-dots.ll
index 00d560c..50b6e09 100644 (file)
--- a/llvm/test/CodeGen/AArch64/sme2-intrinsics-fp-dots.ll
+++ b/llvm/test/CodeGen/AArch64/sme2-intrinsics-fp-dots.ll
@@ -4,6 +4,96 @@
 target triple="aarch64-linux-gnu"
 
 
+; == Multi, multi (16-bit float) ==
+
+define void @fdot_multi_za32_f16_vg1x2(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x half> %zn0, <vscale x 8 x half> %zn1, <vscale x 8 x half> %zn2, <vscale x 8 x half> %zn3) #0 {
+; CHECK-LABEL: fdot_multi_za32_f16_vg1x2:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z5.d, z4.d
+; CHECK-NEXT:    mov z7.d, z2.d
+; CHECK-NEXT:    mov z4.d, z3.d
+; CHECK-NEXT:    mov z6.d, z1.d
+; CHECK-NEXT:    fdot za.s[w8, 0, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    fdot za.s[w8, 7, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    ret
+  call void @llvm.aarch64.sme.fdot.za32.vg1x2.nxv8f16(i32 %slice, <vscale x 8 x half> %zn0, <vscale x 8 x half> %zn1, <vscale x 8 x half> %zn2, <vscale x 8 x half> %zn3)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.fdot.za32.vg1x2.nxv8f16(i32 %slice2, <vscale x 8 x half> %zn0, <vscale x 8 x half> %zn1, <vscale x 8 x half> %zn2, <vscale x 8 x half> %zn3)
+  ret void
+}
+
+define void @fdot_multi_za32_f16_vg1x4(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x half> %zn0, <vscale x 8 x half> %zn1, <vscale x 8 x half> %zn2, <vscale x 8 x half> %zn3,
+; CHECK-LABEL: fdot_multi_za32_f16_vg1x4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z26.d, z7.d
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    mov z25.d, z6.d
+; CHECK-NEXT:    mov z31.d, z4.d
+; CHECK-NEXT:    mov z24.d, z5.d
+; CHECK-NEXT:    mov z30.d, z3.d
+; CHECK-NEXT:    ld1h { z27.h }, p0/z, [x1]
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z29.d, z2.d
+; CHECK-NEXT:    mov z28.d, z1.d
+; CHECK-NEXT:    fdot za.s[w8, 0, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    fdot za.s[w8, 7, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    ret
+                                        <vscale x 8 x half> %zn4, <vscale x 8 x half> %zn5, <vscale x 8 x half> %zn6, <vscale x 8 x half> %zn7) #0 {
+  call void @llvm.aarch64.sme.fdot.za32.vg1x4.nxv8f16(i32 %slice, <vscale x 8 x half> %zn0, <vscale x 8 x half> %zn1, <vscale x 8 x half> %zn2, <vscale x 8 x half> %zn3,
+                                                      <vscale x 8 x half> %zn4, <vscale x 8 x half> %zn5, <vscale x 8 x half> %zn6, <vscale x 8 x half> %zn7)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.fdot.za32.vg1x4.nxv8f16(i32 %slice2, <vscale x 8 x half> %zn0, <vscale x 8 x half> %zn1, <vscale x 8 x half> %zn2, <vscale x 8 x half> %zn3,
+                                                      <vscale x 8 x half> %zn4, <vscale x 8 x half> %zn5, <vscale x 8 x half> %zn6, <vscale x 8 x half> %zn7)
+  ret void
+}
+
+
+; == Multi, multi (16-bit bfloat) ==
+
+define void @bfdot_multi_za32_bf16_vg1x2(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x bfloat> %zn0, <vscale x 8 x bfloat> %zn1, <vscale x 8 x bfloat> %zn2, <vscale x 8 x bfloat> %zn3) #0 {
+; CHECK-LABEL: bfdot_multi_za32_bf16_vg1x2:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z5.d, z4.d
+; CHECK-NEXT:    mov z7.d, z2.d
+; CHECK-NEXT:    mov z4.d, z3.d
+; CHECK-NEXT:    mov z6.d, z1.d
+; CHECK-NEXT:    bfdot za.s[w8, 0, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    bfdot za.s[w8, 7, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    ret
+  call void @llvm.aarch64.sme.fdot.za32.vg1x2.nxv8bf16(i32 %slice, <vscale x 8 x bfloat> %zn0, <vscale x 8 x bfloat> %zn1, <vscale x 8 x bfloat> %zn2, <vscale x 8 x bfloat> %zn3)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.fdot.za32.vg1x2.nxv8bf16(i32 %slice2, <vscale x 8 x bfloat> %zn0, <vscale x 8 x bfloat> %zn1, <vscale x 8 x bfloat> %zn2, <vscale x 8 x bfloat> %zn3)
+  ret void
+}
+
+define void @fdot_multi_za32_bf16_vg1x4(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x bfloat> %zn0, <vscale x 8 x bfloat> %zn1, <vscale x 8 x bfloat> %zn2, <vscale x 8 x bfloat> %zn3,
+; CHECK-LABEL: fdot_multi_za32_bf16_vg1x4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z26.d, z7.d
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    mov z25.d, z6.d
+; CHECK-NEXT:    mov z31.d, z4.d
+; CHECK-NEXT:    mov z24.d, z5.d
+; CHECK-NEXT:    mov z30.d, z3.d
+; CHECK-NEXT:    ld1h { z27.h }, p0/z, [x1]
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z29.d, z2.d
+; CHECK-NEXT:    mov z28.d, z1.d
+; CHECK-NEXT:    bfdot za.s[w8, 0, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    bfdot za.s[w8, 7, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    ret
+                                        <vscale x 8 x bfloat> %zn4, <vscale x 8 x bfloat> %zn5, <vscale x 8 x bfloat> %zn6, <vscale x 8 x bfloat> %zn7) #0 {
+  call void @llvm.aarch64.sme.fdot.za32.vg1x4.nxv8bf16(i32 %slice, <vscale x 8 x bfloat> %zn0, <vscale x 8 x bfloat> %zn1, <vscale x 8 x bfloat> %zn2, <vscale x 8 x bfloat> %zn3,
+                                                       <vscale x 8 x bfloat> %zn4, <vscale x 8 x bfloat> %zn5, <vscale x 8 x bfloat> %zn6, <vscale x 8 x bfloat> %zn7)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.fdot.za32.vg1x4.nxv8bf16(i32 %slice2, <vscale x 8 x bfloat> %zn0, <vscale x 8 x bfloat> %zn1, <vscale x 8 x bfloat> %zn2, <vscale x 8 x bfloat> %zn3,
+                                                        <vscale x 8 x bfloat> %zn4, <vscale x 8 x bfloat> %zn5, <vscale x 8 x bfloat> %zn6, <vscale x 8 x bfloat> %zn7)
+  ret void
+}
+
+
 ; == Multi, single (16-bit float) ==
 
 define void @fdot_single_za32_f16_vg1x2(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x half> %zn0, <vscale x 8 x half> %zn1, <vscale x 8 x half> %zn2) #0 {
@@ -77,6 +167,18 @@ define void @bfdot_single_za32_bf16_vg1x4(i32 %slice, <vscale x 16 x i8> %unused
 attributes #0 = { nounwind "target-features"="+sme2" }
 
 
+; == Multi, multi (16-bit float)
+
+declare void @llvm.aarch64.sme.fdot.za32.vg1x2.nxv8f16(i32, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+declare void @llvm.aarch64.sme.fdot.za32.vg1x4.nxv8f16(i32, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>,
+                                                       <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+
+; == Multi, multi (16-bit bfloat)
+
+declare void @llvm.aarch64.sme.fdot.za32.vg1x2.nxv8bf16(i32, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>)
+declare void @llvm.aarch64.sme.fdot.za32.vg1x4.nxv8bf16(i32, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>,
+                                                         <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>)
+
 ; == Multi, single (16-bit float)
 
 declare void @llvm.aarch64.sme.fdot.single.za32.vg1x2.nxv8f16(i32, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
@@ -86,3 +188,4 @@ declare void @llvm.aarch64.sme.fdot.single.za32.vg1x4.nxv8f16(i32, <vscale x 8 x
 
 declare void @llvm.aarch64.sme.fdot.single.za32.vg1x2.nxv8bf16(i32, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>)
 declare void @llvm.aarch64.sme.fdot.single.za32.vg1x4.nxv8bf16(i32, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>)
+

diff --git a/llvm/test/CodeGen/AArch64/sme2-intrinsics-int-dots.ll b/llvm/test/CodeGen/AArch64/sme2-intrinsics-int-dots.ll
index 6928ced..5173e18 100644 (file)
--- a/llvm/test/CodeGen/AArch64/sme2-intrinsics-int-dots.ll
+++ b/llvm/test/CodeGen/AArch64/sme2-intrinsics-int-dots.ll
@@ -4,6 +4,306 @@
 target triple="aarch64-linux-gnu"
 
 
+; == Multi, multi (unsigned) ==
+
+define void @udot_multi_za32_u16_vg1x2(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3) #0 {
+; CHECK-LABEL: udot_multi_za32_u16_vg1x2:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z5.d, z4.d
+; CHECK-NEXT:    mov z7.d, z2.d
+; CHECK-NEXT:    mov z4.d, z3.d
+; CHECK-NEXT:    mov z6.d, z1.d
+; CHECK-NEXT:    udot za.s[w8, 0, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    udot za.s[w8, 7, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    ret
+  call void @llvm.aarch64.sme.udot.za32.vg1x2.nxv8i16(i32 %slice, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.udot.za32.vg1x2.nxv8i16(i32 %slice2, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3)
+  ret void
+}
+
+define void @udot_multi_za32_u16_vg1x4(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+; CHECK-LABEL: udot_multi_za32_u16_vg1x4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z26.d, z7.d
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    mov z25.d, z6.d
+; CHECK-NEXT:    mov z31.d, z4.d
+; CHECK-NEXT:    mov z24.d, z5.d
+; CHECK-NEXT:    mov z30.d, z3.d
+; CHECK-NEXT:    ld1h { z27.h }, p0/z, [x1]
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z29.d, z2.d
+; CHECK-NEXT:    mov z28.d, z1.d
+; CHECK-NEXT:    udot za.s[w8, 0, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    udot za.s[w8, 7, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    ret
+                                        <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7) #0 {
+  call void @llvm.aarch64.sme.udot.za32.vg1x4.nxv8i16(i32 %slice, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+                                                      <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.udot.za32.vg1x4.nxv8i16(i32 %slice2, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+                                                      <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7)
+  ret void
+}
+
+define void @udot_multi_za32_u8_vg1x2(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3) #0 {
+; CHECK-LABEL: udot_multi_za32_u8_vg1x2:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z5.d, z4.d
+; CHECK-NEXT:    mov z7.d, z2.d
+; CHECK-NEXT:    mov z4.d, z3.d
+; CHECK-NEXT:    mov z6.d, z1.d
+; CHECK-NEXT:    udot za.s[w8, 0, vgx2], { z6.b, z7.b }, { z4.b, z5.b }
+; CHECK-NEXT:    udot za.s[w8, 7, vgx2], { z6.b, z7.b }, { z4.b, z5.b }
+; CHECK-NEXT:    ret
+  call void @llvm.aarch64.sme.udot.za32.vg1x2.nxv16i8(i32 %slice, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.udot.za32.vg1x2.nxv16i8(i32 %slice2, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3)
+  ret void
+}
+
+define void @udot_multi_za32_u8_vg1x4(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3,
+; CHECK-LABEL: udot_multi_za32_u8_vg1x4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z26.d, z7.d
+; CHECK-NEXT:    ptrue p0.b
+; CHECK-NEXT:    mov z25.d, z6.d
+; CHECK-NEXT:    mov z31.d, z4.d
+; CHECK-NEXT:    mov z24.d, z5.d
+; CHECK-NEXT:    mov z30.d, z3.d
+; CHECK-NEXT:    ld1b { z27.b }, p0/z, [x1]
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z29.d, z2.d
+; CHECK-NEXT:    mov z28.d, z1.d
+; CHECK-NEXT:    udot za.s[w8, 0, vgx4], { z28.b - z31.b }, { z24.b - z27.b }
+; CHECK-NEXT:    udot za.s[w8, 7, vgx4], { z28.b - z31.b }, { z24.b - z27.b }
+; CHECK-NEXT:    ret
+                                      <vscale x 16 x i8> %zn4, <vscale x 16 x i8> %zn5, <vscale x 16 x i8> %zn6, <vscale x 16 x i8> %zn7) #0 {
+  call void @llvm.aarch64.sme.udot.za32.vg1x4.nxv16i8(i32 %slice, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3,
+                                                      <vscale x 16 x i8> %zn4, <vscale x 16 x i8> %zn5, <vscale x 16 x i8> %zn6, <vscale x 16 x i8> %zn7)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.udot.za32.vg1x4.nxv16i8(i32 %slice2, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3,
+                                                      <vscale x 16 x i8> %zn4, <vscale x 16 x i8> %zn5, <vscale x 16 x i8> %zn6, <vscale x 16 x i8> %zn7)
+  ret void
+}
+
+define void @udot_multi_za64_u16_vg1x2(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3) #1 {
+; CHECK-LABEL: udot_multi_za64_u16_vg1x2:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z5.d, z4.d
+; CHECK-NEXT:    mov z7.d, z2.d
+; CHECK-NEXT:    mov z4.d, z3.d
+; CHECK-NEXT:    mov z6.d, z1.d
+; CHECK-NEXT:    udot za.d[w8, 0, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    udot za.d[w8, 7, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    ret
+  call void @llvm.aarch64.sme.udot.za64.vg1x2.nxv8i16(i32 %slice, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.udot.za64.vg1x2.nxv8i16(i32 %slice2, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3)
+  ret void
+}
+
+define void @udot_multi_za64_u16_vg1x4(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+; CHECK-LABEL: udot_multi_za64_u16_vg1x4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z26.d, z7.d
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    mov z25.d, z6.d
+; CHECK-NEXT:    mov z31.d, z4.d
+; CHECK-NEXT:    mov z24.d, z5.d
+; CHECK-NEXT:    mov z30.d, z3.d
+; CHECK-NEXT:    ld1h { z27.h }, p0/z, [x1]
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z29.d, z2.d
+; CHECK-NEXT:    mov z28.d, z1.d
+; CHECK-NEXT:    udot za.d[w8, 0, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    udot za.d[w8, 7, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    ret
+                                       <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7) #1 {
+  call void @llvm.aarch64.sme.udot.za64.vg1x4.nxv8i16(i32 %slice, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+                                                      <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.udot.za64.vg1x4.nxv8i16(i32 %slice2, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+                                                      <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7)
+  ret void
+}
+
+define void @usdot_multi_za32_u8_vg1x2(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3) #0 {
+; CHECK-LABEL: usdot_multi_za32_u8_vg1x2:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z5.d, z4.d
+; CHECK-NEXT:    mov z7.d, z2.d
+; CHECK-NEXT:    mov z4.d, z3.d
+; CHECK-NEXT:    mov z6.d, z1.d
+; CHECK-NEXT:    usdot za.s[w8, 0, vgx2], { z6.b, z7.b }, { z4.b, z5.b }
+; CHECK-NEXT:    usdot za.s[w8, 7, vgx2], { z6.b, z7.b }, { z4.b, z5.b }
+; CHECK-NEXT:    ret
+  call void @llvm.aarch64.sme.usdot.za32.vg1x2.nxv16i8(i32 %slice, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.usdot.za32.vg1x2.nxv16i8(i32 %slice2, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3)
+  ret void
+}
+
+define void @usdot_multi_za32_u8_vg1x4(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3,
+; CHECK-LABEL: usdot_multi_za32_u8_vg1x4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z26.d, z7.d
+; CHECK-NEXT:    ptrue p0.b
+; CHECK-NEXT:    mov z25.d, z6.d
+; CHECK-NEXT:    mov z31.d, z4.d
+; CHECK-NEXT:    mov z24.d, z5.d
+; CHECK-NEXT:    mov z30.d, z3.d
+; CHECK-NEXT:    ld1b { z27.b }, p0/z, [x1]
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z29.d, z2.d
+; CHECK-NEXT:    mov z28.d, z1.d
+; CHECK-NEXT:    usdot za.s[w8, 0, vgx4], { z28.b - z31.b }, { z24.b - z27.b }
+; CHECK-NEXT:    usdot za.s[w8, 7, vgx4], { z28.b - z31.b }, { z24.b - z27.b }
+; CHECK-NEXT:    ret
+                                      <vscale x 16 x i8> %zn4, <vscale x 16 x i8> %zn5, <vscale x 16 x i8> %zn6, <vscale x 16 x i8> %zn7) #0 {
+  call void @llvm.aarch64.sme.usdot.za32.vg1x4.nxv16i8(i32 %slice, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3,
+                                                      <vscale x 16 x i8> %zn4, <vscale x 16 x i8> %zn5, <vscale x 16 x i8> %zn6, <vscale x 16 x i8> %zn7)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.usdot.za32.vg1x4.nxv16i8(i32 %slice2, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3,
+                                                      <vscale x 16 x i8> %zn4, <vscale x 16 x i8> %zn5, <vscale x 16 x i8> %zn6, <vscale x 16 x i8> %zn7)
+  ret void
+}
+
+
+; == Multi, multi (signed) ==
+
+define void @sdot_multi_za32_u16_vg1x2(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3) #0 {
+; CHECK-LABEL: sdot_multi_za32_u16_vg1x2:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z5.d, z4.d
+; CHECK-NEXT:    mov z7.d, z2.d
+; CHECK-NEXT:    mov z4.d, z3.d
+; CHECK-NEXT:    mov z6.d, z1.d
+; CHECK-NEXT:    sdot za.s[w8, 0, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    sdot za.s[w8, 7, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    ret
+  call void @llvm.aarch64.sme.sdot.za32.vg1x2.nxv8i16(i32 %slice, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.sdot.za32.vg1x2.nxv8i16(i32 %slice2, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3)
+  ret void
+}
+
+define void @sdot_multi_za32_u16_vg1x4(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+; CHECK-LABEL: sdot_multi_za32_u16_vg1x4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z26.d, z7.d
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    mov z25.d, z6.d
+; CHECK-NEXT:    mov z31.d, z4.d
+; CHECK-NEXT:    mov z24.d, z5.d
+; CHECK-NEXT:    mov z30.d, z3.d
+; CHECK-NEXT:    ld1h { z27.h }, p0/z, [x1]
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z29.d, z2.d
+; CHECK-NEXT:    mov z28.d, z1.d
+; CHECK-NEXT:    sdot za.s[w8, 0, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    sdot za.s[w8, 7, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    ret
+                                        <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7) #0 {
+  call void @llvm.aarch64.sme.sdot.za32.vg1x4.nxv8i16(i32 %slice, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+                                                      <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.sdot.za32.vg1x4.nxv8i16(i32 %slice2, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+                                                      <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7)
+  ret void
+}
+
+define void @sdot_multi_za32_u8_vg1x2(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3) #0 {
+; CHECK-LABEL: sdot_multi_za32_u8_vg1x2:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z5.d, z4.d
+; CHECK-NEXT:    mov z7.d, z2.d
+; CHECK-NEXT:    mov z4.d, z3.d
+; CHECK-NEXT:    mov z6.d, z1.d
+; CHECK-NEXT:    sdot za.s[w8, 0, vgx2], { z6.b, z7.b }, { z4.b, z5.b }
+; CHECK-NEXT:    sdot za.s[w8, 7, vgx2], { z6.b, z7.b }, { z4.b, z5.b }
+; CHECK-NEXT:    ret
+  call void @llvm.aarch64.sme.sdot.za32.vg1x2.nxv16i8(i32 %slice, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.sdot.za32.vg1x2.nxv16i8(i32 %slice2, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3)
+  ret void
+}
+
+define void @sdot_multi_za32_u8_vg1x4(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3,
+; CHECK-LABEL: sdot_multi_za32_u8_vg1x4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z26.d, z7.d
+; CHECK-NEXT:    ptrue p0.b
+; CHECK-NEXT:    mov z25.d, z6.d
+; CHECK-NEXT:    mov z31.d, z4.d
+; CHECK-NEXT:    mov z24.d, z5.d
+; CHECK-NEXT:    mov z30.d, z3.d
+; CHECK-NEXT:    ld1b { z27.b }, p0/z, [x1]
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z29.d, z2.d
+; CHECK-NEXT:    mov z28.d, z1.d
+; CHECK-NEXT:    sdot za.s[w8, 0, vgx4], { z28.b - z31.b }, { z24.b - z27.b }
+; CHECK-NEXT:    sdot za.s[w8, 7, vgx4], { z28.b - z31.b }, { z24.b - z27.b }
+; CHECK-NEXT:    ret
+                                      <vscale x 16 x i8> %zn4, <vscale x 16 x i8> %zn5, <vscale x 16 x i8> %zn6, <vscale x 16 x i8> %zn7) #0 {
+  call void @llvm.aarch64.sme.sdot.za32.vg1x4.nxv16i8(i32 %slice, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3,
+                                                      <vscale x 16 x i8> %zn4, <vscale x 16 x i8> %zn5, <vscale x 16 x i8> %zn6, <vscale x 16 x i8> %zn7)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.sdot.za32.vg1x4.nxv16i8(i32 %slice2, <vscale x 16 x i8> %zn0, <vscale x 16 x i8> %zn1, <vscale x 16 x i8> %zn2, <vscale x 16 x i8> %zn3,
+                                                      <vscale x 16 x i8> %zn4, <vscale x 16 x i8> %zn5, <vscale x 16 x i8> %zn6, <vscale x 16 x i8> %zn7)
+  ret void
+}
+
+define void @sdot_multi_za64_u16_vg1x2(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3) #1 {
+; CHECK-LABEL: sdot_multi_za64_u16_vg1x2:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z5.d, z4.d
+; CHECK-NEXT:    mov z7.d, z2.d
+; CHECK-NEXT:    mov z4.d, z3.d
+; CHECK-NEXT:    mov z6.d, z1.d
+; CHECK-NEXT:    sdot za.d[w8, 0, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    sdot za.d[w8, 7, vgx2], { z6.h, z7.h }, { z4.h, z5.h }
+; CHECK-NEXT:    ret
+  call void @llvm.aarch64.sme.sdot.za64.vg1x2.nxv8i16(i32 %slice, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.sdot.za64.vg1x2.nxv8i16(i32 %slice2, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3)
+  ret void
+}
+
+define void @sdot_multi_za64_u16_vg1x4(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+; CHECK-LABEL: sdot_multi_za64_u16_vg1x4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z26.d, z7.d
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    mov z25.d, z6.d
+; CHECK-NEXT:    mov z31.d, z4.d
+; CHECK-NEXT:    mov z24.d, z5.d
+; CHECK-NEXT:    mov z30.d, z3.d
+; CHECK-NEXT:    ld1h { z27.h }, p0/z, [x1]
+; CHECK-NEXT:    mov w8, w0
+; CHECK-NEXT:    mov z29.d, z2.d
+; CHECK-NEXT:    mov z28.d, z1.d
+; CHECK-NEXT:    sdot za.d[w8, 0, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    sdot za.d[w8, 7, vgx4], { z28.h - z31.h }, { z24.h - z27.h }
+; CHECK-NEXT:    ret
+                                       <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7) #1 {
+  call void @llvm.aarch64.sme.sdot.za64.vg1x4.nxv8i16(i32 %slice, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+                                                      <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7)
+  %slice2 = add i32 %slice, 7
+  call void @llvm.aarch64.sme.sdot.za64.vg1x4.nxv8i16(i32 %slice2, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2, <vscale x 8 x i16> %zn3,
+                                                      <vscale x 8 x i16> %zn4, <vscale x 8 x i16> %zn5, <vscale x 8 x i16> %zn6, <vscale x 8 x i16> %zn7)
+  ret void
+}
+
+
 ; == Multi, single (unsigned) ==
 
 define void @udot_single_za32_u16_vg1x2(i32 %slice, <vscale x 16 x i8> %unused, <vscale x 8 x i16> %zn0, <vscale x 8 x i16> %zn1, <vscale x 8 x i16> %zn2) #0 {
@@ -270,6 +570,33 @@ attributes #0 = { nounwind "target-features"="+sme2" }
 attributes #1 = { nounwind "target-features"="+sme2,+sme-i16i64" }
 
 
+; == Multi, multi (unsigned)
+
+declare void @llvm.aarch64.sme.udot.za32.vg1x2.nxv8i16(i32, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare void @llvm.aarch64.sme.udot.za32.vg1x4.nxv8i16(i32, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>,
+                                                       <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare void @llvm.aarch64.sme.udot.za32.vg1x2.nxv16i8(i32, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+declare void @llvm.aarch64.sme.udot.za32.vg1x4.nxv16i8(i32, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>,
+                                                       <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+declare void @llvm.aarch64.sme.udot.za64.vg1x2.nxv8i16(i32, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare void @llvm.aarch64.sme.udot.za64.vg1x4.nxv8i16(i32, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>,
+                                                       <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare void @llvm.aarch64.sme.usdot.za32.vg1x2.nxv16i8(i32, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+declare void @llvm.aarch64.sme.usdot.za32.vg1x4.nxv16i8(i32, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>,
+                                                        <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+
+; == Multi, multi (signed)
+
+declare void @llvm.aarch64.sme.sdot.za32.vg1x2.nxv8i16(i32, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare void @llvm.aarch64.sme.sdot.za32.vg1x4.nxv8i16(i32, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>,
+                                                       <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare void @llvm.aarch64.sme.sdot.za32.vg1x2.nxv16i8(i32, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+declare void @llvm.aarch64.sme.sdot.za32.vg1x4.nxv16i8(i32, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>,
+                                                       <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+declare void @llvm.aarch64.sme.sdot.za64.vg1x2.nxv8i16(i32, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare void @llvm.aarch64.sme.sdot.za64.vg1x4.nxv8i16(i32, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>,
+                                                       <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+
 ; == Multi, single (unsigned)
 
 declare void @llvm.aarch64.sme.udot.single.za32.vg1x2.nxv8i16(i32, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)