From 06191d132771814ff04df7898141e0db81d39215 Mon Sep 17 00:00:00 2001
From: Caroline Concatto <caroline.concatto@arm.com>
Date: Mon, 23 Jan 2023 18:01:48 +0000
Subject: [PATCH] [AArch64][SME2] Add Multi-vector saturating extract narrow
 and interleave intrinsics

 Add the following intrinsic:
      SQCVTN
      SQCVTUN
      UQCVTN

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

Reviewed By: kmclaughlin

Differential Revision: https://reviews.llvm.org/D142089
---
 llvm/include/llvm/IR/IntrinsicsAArch64.td          |  10 ++
 llvm/lib/Target/AArch64/AArch64SMEInstrInfo.td     |   6 +-
 llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td     |   6 +-
 llvm/lib/Target/AArch64/SMEInstrFormats.td         |   6 +-
 llvm/lib/Target/AArch64/SVEInstrFormats.td         |   9 +-
 llvm/test/CodeGen/AArch64/sme2-intrinsics-qcvtn.ll | 140 +++++++++++++++++++++
 6 files changed, 164 insertions(+), 13 deletions(-)
 create mode 100644 llvm/test/CodeGen/AArch64/sme2-intrinsics-qcvtn.ll

diff --git a/llvm/include/llvm/IR/IntrinsicsAArch64.td b/llvm/include/llvm/IR/IntrinsicsAArch64.td
index ca14f62..e365f2a 100644
--- a/llvm/include/llvm/IR/IntrinsicsAArch64.td
+++ b/llvm/include/llvm/IR/IntrinsicsAArch64.td
@@ -2889,4 +2889,14 @@ let TargetPrefix = "aarch64" in {
   def int_aarch64_sve_sqcvt_x4  : SME2_CVT_VG4_SINGLE_Intrinsic;
   def int_aarch64_sve_uqcvt_x4  : SME2_CVT_VG4_SINGLE_Intrinsic;
   def int_aarch64_sve_sqcvtu_x4 : SME2_CVT_VG4_SINGLE_Intrinsic;
+
+  //
+  // Multi-vector saturating extract narrow and interleave
+  //
+  def int_aarch64_sve_sqcvtn_x2  : SME2_CVT_VG2_SINGLE_Intrinsic;
+  def int_aarch64_sve_uqcvtn_x2  : SME2_CVT_VG2_SINGLE_Intrinsic;
+  def int_aarch64_sve_sqcvtun_x2 : SME2_CVT_VG2_SINGLE_Intrinsic;
+  def int_aarch64_sve_sqcvtn_x4  : SME2_CVT_VG4_SINGLE_Intrinsic;
+  def int_aarch64_sve_uqcvtn_x4  : SME2_CVT_VG4_SINGLE_Intrinsic;
+  def int_aarch64_sve_sqcvtun_x4 : SME2_CVT_VG4_SINGLE_Intrinsic;
 }
diff --git a/llvm/lib/Target/AArch64/AArch64SMEInstrInfo.td b/llvm/lib/Target/AArch64/AArch64SMEInstrInfo.td
index ba9b4b9..a213d65 100644
--- a/llvm/lib/Target/AArch64/AArch64SMEInstrInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64SMEInstrInfo.td
@@ -374,9 +374,9 @@ defm SQCVTU_Z2Z_StoH : sme2_cvt_vg2_single<"sqcvtu", 0b1110, nxv8i16, nxv4i32, i
 defm SQCVT_Z4Z      : sme2_int_cvt_vg4_single<"sqcvt", 0b000, int_aarch64_sve_sqcvt_x4>;
 defm UQCVT_Z4Z      : sme2_int_cvt_vg4_single<"uqcvt", 0b001, int_aarch64_sve_uqcvt_x4>;
 defm SQCVTU_Z4Z     : sme2_int_cvt_vg4_single<"sqcvtu", 0b100, int_aarch64_sve_sqcvtu_x4>;
-defm SQCVTN_Z4Z     : sme2_int_cvt_vg4_single<"sqcvtn", 0b010, null_frag>;
-defm SQCVTUN_Z4Z    : sme2_int_cvt_vg4_single<"sqcvtun", 0b110, null_frag>;
-defm UQCVTN_Z4Z     : sme2_int_cvt_vg4_single<"uqcvtn", 0b011, null_frag>;
+defm SQCVTN_Z4Z     : sme2_int_cvt_vg4_single<"sqcvtn", 0b010, int_aarch64_sve_sqcvtn_x4>;
+defm SQCVTUN_Z4Z    : sme2_int_cvt_vg4_single<"sqcvtun", 0b110, int_aarch64_sve_sqcvtun_x4>;
+defm UQCVTN_Z4Z     : sme2_int_cvt_vg4_single<"uqcvtn", 0b011, int_aarch64_sve_uqcvtn_x4>;
 
 defm FCVTZS_2Z2Z_StoS : sme2_fp_cvt_vg2_multi<"fcvtzs", 0b00010>;
 defm FCVTZS_4Z4Z_StoS : sme2_fp_cvt_vg4_multi<"fcvtzs", 0b0001000>;
diff --git a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
index 8a37916..6a42c4f 100644
--- a/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td
@@ -3691,9 +3691,9 @@ defm PEXT_PCI : sve2p1_pred_as_ctr_to_mask<"pext">;
 defm PEXT_2PCI : sve2p1_pred_as_ctr_to_mask_pair<"pext">;
 defm PTRUE_C  : sve2p1_ptrue_pn<"ptrue">;
 
-defm SQCVTN_Z2Z_StoH  : sve2p1_multi_vec_extract_narrow<"sqcvtn", 0b00>;
-defm UQCVTN_Z2Z_StoH  : sve2p1_multi_vec_extract_narrow<"uqcvtn", 0b01>;
-defm SQCVTUN_Z2Z_StoH : sve2p1_multi_vec_extract_narrow<"sqcvtun", 0b10>;
+defm SQCVTN_Z2Z_StoH  : sve2p1_multi_vec_extract_narrow<"sqcvtn", 0b00, int_aarch64_sve_sqcvtn_x2>;
+defm UQCVTN_Z2Z_StoH  : sve2p1_multi_vec_extract_narrow<"uqcvtn", 0b01, int_aarch64_sve_uqcvtn_x2>;
+defm SQCVTUN_Z2Z_StoH : sve2p1_multi_vec_extract_narrow<"sqcvtun", 0b10, int_aarch64_sve_sqcvtun_x2>;
 defm SQRSHRN_Z2ZI_StoH  : sve2p1_multi_vec_shift_narrow<"sqrshrn", 0b101>;
 defm UQRSHRN_Z2ZI_StoH  : sve2p1_multi_vec_shift_narrow<"uqrshrn", 0b111>;
 defm SQRSHRUN_Z2ZI_StoH : sve2p1_multi_vec_shift_narrow<"sqrshrun", 0b001>;
diff --git a/llvm/lib/Target/AArch64/SMEInstrFormats.td b/llvm/lib/Target/AArch64/SMEInstrFormats.td
index 33ff5f0..22635a3 100644
--- a/llvm/lib/Target/AArch64/SMEInstrFormats.td
+++ b/llvm/lib/Target/AArch64/SMEInstrFormats.td
@@ -133,10 +133,6 @@ class SME2_ZA_TwoOp_VG4_Multi_Index_Pat<string name, SDPatternOperator intrinsic
                                               (REG_SEQUENCE ZPR4Mul4, vt:$Zn1, zsub0, vt:$Zn2, zsub1, vt:$Zn3, zsub2, vt:$Zn4, zsub3),
                                               zpr_ty:$Zm, imm_ty:$i)>;
 
-class SME2_Cvt_VG2_Pat<string name, SDPatternOperator intrinsic, ValueType out_vt, ValueType in_vt>
-    : Pat<(out_vt (intrinsic in_vt:$Zn1, in_vt:$Zn2)),
-                  (!cast<Instruction>(name) (REG_SEQUENCE ZPR2Mul2, in_vt:$Zn1, zsub0, in_vt:$Zn2, zsub1))>;
-
 class SME2_Cvt_VG4_Pat<string name, SDPatternOperator intrinsic, ValueType out_vt, ValueType in_vt>
     : Pat<(out_vt (intrinsic in_vt:$Zn1, in_vt:$Zn2, in_vt:$Zn3, in_vt:$Zn4)),
                   (!cast<Instruction>(name) (REG_SEQUENCE ZPR4Mul4, in_vt:$Zn1, zsub0, in_vt:$Zn2, zsub1, in_vt:$Zn3, zsub2, in_vt:$Zn4, zsub3))>;
@@ -2087,7 +2083,7 @@ class sme2_cvt_vg2_single<string mnemonic, bits<4> op>
 multiclass sme2_cvt_vg2_single<string mnemonic, bits<4> op, ValueType out_vt,
                                ValueType in_vt, SDPatternOperator intrinsic> {
   def NAME :  sme2_cvt_vg2_single<mnemonic, op>;
-  def : SME2_Cvt_VG2_Pat<NAME, intrinsic, out_vt, in_vt>;
+  def : SVE2p1_Cvt_VG2_Pat<NAME, intrinsic, out_vt, in_vt>;
 }
 
 class sme2_cvt_unpk_vector_vg2<bits<2>sz, bits<3> op, bit u, RegisterOperand first_ty,
diff --git a/llvm/lib/Target/AArch64/SVEInstrFormats.td b/llvm/lib/Target/AArch64/SVEInstrFormats.td
index dcaada7..cef8d41 100644
--- a/llvm/lib/Target/AArch64/SVEInstrFormats.td
+++ b/llvm/lib/Target/AArch64/SVEInstrFormats.td
@@ -549,6 +549,10 @@ class SVE_Shift_Add_All_Active_Pat<ValueType vtd, SDPatternOperator op, ValueTyp
 : Pat<(vtd (add vt1:$Op1, (op (pt (SVEAllActive)), vt2:$Op2, vt3:$Op3))),
       (inst $Op1, $Op2, $Op3)>;
 
+class SVE2p1_Cvt_VG2_Pat<string name, SDPatternOperator intrinsic, ValueType out_vt, ValueType in_vt>
+    : Pat<(out_vt (intrinsic in_vt:$Zn1, in_vt:$Zn2)),
+                  (!cast<Instruction>(name) (REG_SEQUENCE ZPR2Mul2, in_vt:$Zn1, zsub0, in_vt:$Zn2, zsub1))>;
+
 //===----------------------------------------------------------------------===//
 // SVE pattern match helpers.
 //===----------------------------------------------------------------------===//
@@ -8861,8 +8865,9 @@ class sve2p1_multi_vec_extract_narrow<string mnemonic, bits<2> opc, bits<3> tsz>
   let Inst{4-0}   = Zd;
 }
 
-multiclass sve2p1_multi_vec_extract_narrow<string mnemonic, bits<2> opc> {
-  def : sve2p1_multi_vec_extract_narrow<mnemonic, opc, 0b010>;
+multiclass sve2p1_multi_vec_extract_narrow<string mnemonic, bits<2> opc, SDPatternOperator intrinsic> {
+  def NAME : sve2p1_multi_vec_extract_narrow<mnemonic, opc, 0b010>;
+  def : SVE2p1_Cvt_VG2_Pat<NAME, intrinsic, nxv8i16, nxv4i32>;
 }
 
 // SVE2 multi-vec shift narrow
diff --git a/llvm/test/CodeGen/AArch64/sme2-intrinsics-qcvtn.ll b/llvm/test/CodeGen/AArch64/sme2-intrinsics-qcvtn.ll
new file mode 100644
index 0000000..072282c
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/sme2-intrinsics-qcvtn.ll
@@ -0,0 +1,140 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py$
+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve,+sme2,+bf16 -verify-machineinstrs < %s | FileCheck %s
+
+;
+; SQCVTN
+;
+
+; x2
+define <vscale x 8 x i16 > @multi_vector_qcvtn_x2_s16_s32(<vscale x 4 x i32> %unused, <vscale x 4 x i32> %zn1, <vscale x 4 x i32> %zn2) {
+; CHECK-LABEL: multi_vector_qcvtn_x2_s16_s32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z3.d, z2.d
+; CHECK-NEXT:    mov z2.d, z1.d
+; CHECK-NEXT:    sqcvtn z0.h, { z2.s, z3.s }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 8 x i16> @llvm.aarch64.sve.sqcvtn.x2.nxv4i32(<vscale x 4 x i32> %zn1, <vscale x 4 x i32> %zn2)
+  ret <vscale x 8 x i16> %res
+}
+
+; x4
+define <vscale x 16 x i8 > @multi_vector_qcvtn_x4_s8_s32(<vscale x 4 x i32> %unused, <vscale x 4 x i32> %zn1, <vscale x 4 x i32> %zn2, <vscale x 4 x i32> %zn3, <vscale x 4 x i32> %zn4) {
+; CHECK-LABEL: multi_vector_qcvtn_x4_s8_s32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    sqcvtn z0.b, { z4.s - z7.s }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 16 x i8> @llvm.aarch64.sve.sqcvtn.x4.nxv4i32(<vscale x 4 x i32> %zn1, <vscale x 4 x i32> %zn2, <vscale x 4 x i32> %zn3, <vscale x 4 x i32> %zn4)
+  ret <vscale x 16 x i8> %res
+}
+
+define <vscale x 8 x i16> @multi_vector_qcvtn_x4_s16_s64(<vscale x 2 x i64> %unused, <vscale x 2 x i64> %zn1, <vscale x 2 x i64> %zn2, <vscale x 2 x i64> %zn3, <vscale x 2 x i64> %zn4) {
+; CHECK-LABEL: multi_vector_qcvtn_x4_s16_s64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    sqcvtn z0.h, { z4.d - z7.d }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 8 x i16> @llvm.aarch64.sve.sqcvtn.x4.nxv2i64(<vscale x 2 x i64> %zn1, <vscale x 2 x i64> %zn2, <vscale x 2 x i64> %zn3, <vscale x 2 x i64> %zn4)
+  ret <vscale x 8 x i16> %res
+}
+
+;
+; UQCVTN
+;
+
+; x2
+define <vscale x 8 x i16> @multi_vector_qcvtn_x2_u16_u32(<vscale x 4 x i32> %unused, <vscale x 4 x i32> %zn0, <vscale x 4 x i32> %zn1) {
+; CHECK-LABEL: multi_vector_qcvtn_x2_u16_u32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z3.d, z2.d
+; CHECK-NEXT:    mov z2.d, z1.d
+; CHECK-NEXT:    uqcvtn z0.h, { z2.s, z3.s }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 8 x i16> @llvm.aarch64.sve.uqcvtn.x2.nxv4i32(<vscale x 4 x i32> %zn0, <vscale x 4 x i32> %zn1)
+  ret<vscale x 8 x i16> %res
+}
+
+; x4
+define <vscale x 16 x i8> @multi_vector_qcvtn_x4_u8_u32(<vscale x 4 x i32> %unused, <vscale x 4 x i32> %zn1, <vscale x 4 x i32> %zn2, <vscale x 4 x i32> %zn3, <vscale x 4 x i32> %zn4) {
+; CHECK-LABEL: multi_vector_qcvtn_x4_u8_u32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    uqcvtn z0.b, { z4.s - z7.s }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 16 x i8> @llvm.aarch64.sve.uqcvtn.x4.nxv4i32(<vscale x 4 x i32> %zn1, <vscale x 4 x i32> %zn2, <vscale x 4 x i32> %zn3, <vscale x 4 x i32> %zn4)
+  ret <vscale x 16 x i8> %res
+}
+
+define <vscale x 8 x i16> @multi_vector_qcvtn_x4_u16_u64(<vscale x 2 x i64> %unused, <vscale x 2 x i64> %zn1, <vscale x 2 x i64> %zn2, <vscale x 2 x i64> %zn3, <vscale x 2 x i64> %zn4) {
+; CHECK-LABEL: multi_vector_qcvtn_x4_u16_u64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    uqcvtn z0.h, { z4.d - z7.d }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 8 x i16> @llvm.aarch64.sve.uqcvtn.x4.nxv2i64(<vscale x 2 x i64> %zn1, <vscale x 2 x i64> %zn2, <vscale x 2 x i64> %zn3, <vscale x 2 x i64> %zn4)
+  ret <vscale x 8 x i16> %res
+}
+
+;
+; SQCVTUN
+;
+
+; x2
+define <vscale x 8 x i16 > @multi_vector_qcvtn_x2_s16_u32(<vscale x 4 x i32> %unused, <vscale x 4 x i32> %zn1, <vscale x 4 x i32> %zn2) {
+; CHECK-LABEL: multi_vector_qcvtn_x2_s16_u32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z3.d, z2.d
+; CHECK-NEXT:    mov z2.d, z1.d
+; CHECK-NEXT:    sqcvtun z0.h, { z2.s, z3.s }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 8 x i16> @llvm.aarch64.sve.sqcvtun.x2.nxv4i322(<vscale x 4 x i32> %zn1, <vscale x 4 x i32> %zn2)
+  ret <vscale x 8 x i16> %res
+}
+; x4
+define <vscale x 16 x i8> @multi_vector_qcvtn_x4_u8_s32(<vscale x 4 x i32> %unused, <vscale x 4 x i32> %zn1, <vscale x 4 x i32> %zn2, <vscale x 4 x i32> %zn3, <vscale x 4 x i32> %zn4) {
+; CHECK-LABEL: multi_vector_qcvtn_x4_u8_s32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    sqcvtun z0.b, { z4.s - z7.s }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 16 x i8> @llvm.aarch64.sve.sqcvtun.x4.nxv4i32(<vscale x 4 x i32> %zn1, <vscale x 4 x i32> %zn2, <vscale x 4 x i32> %zn3, <vscale x 4 x i32> %zn4)
+  ret <vscale x 16 x i8> %res
+}
+
+define <vscale x 8 x i16> @multi_vector_qcvtn_x4_u16_s64(<vscale x 2 x i64> %unused, <vscale x 2 x i64> %zn1, <vscale x 2 x i64> %zn2, <vscale x 2 x i64> %zn3, <vscale x 2 x i64> %zn4) {
+; CHECK-LABEL: multi_vector_qcvtn_x4_u16_s64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov z7.d, z4.d
+; CHECK-NEXT:    mov z6.d, z3.d
+; CHECK-NEXT:    mov z5.d, z2.d
+; CHECK-NEXT:    mov z4.d, z1.d
+; CHECK-NEXT:    sqcvtun z0.h, { z4.d - z7.d }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 8 x i16> @llvm.aarch64.sve.sqcvtun.x4.nxv2i64(<vscale x 2 x i64> %zn1, <vscale x 2 x i64> %zn2, <vscale x 2 x i64> %zn3, <vscale x 2 x i64> %zn4)
+  ret <vscale x 8 x i16> %res
+}
+
+declare <vscale x 8 x i16> @llvm.aarch64.sve.uqcvtn.x2.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 8 x i16> @llvm.aarch64.sve.sqcvtn.x2.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 8 x i16> @llvm.aarch64.sve.sqcvtun.x2.nxv4i322(<vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 16 x i8> @llvm.aarch64.sve.sqcvtn.x4.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 8 x i16> @llvm.aarch64.sve.sqcvtn.x4.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
+declare <vscale x 16 x i8> @llvm.aarch64.sve.uqcvtn.x4.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 8 x i16> @llvm.aarch64.sve.uqcvtn.x4.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
+declare <vscale x 16 x i8> @llvm.aarch64.sve.sqcvtun.x4.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 8 x i16> @llvm.aarch64.sve.sqcvtun.x4.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
-- 
2.7.4