[llvm][CodeGen] IR intrinsics for SVE2 contiguous conflict detection instructions.

[fpetrogalli]

Summary

The IR intrinsics are mapped to the following SVE2 instructions:

• WHILERW <Pd>.<T>, <Xn>, <Xm>
• WHILEWR <Pd>.<T>, <Xn>, <Xm>

The intrinsics introduced in this patch are the IR counterpart of the
SVE ACLE functions svwhilerw and svwhilewr (all data type
variants).

Patch by Maciej Gąbka <[e-mail]>.

[kmclaughlin]

LGTM!

llvm/include/llvm/IR/IntrinsicsAArch64.td

class SVE2_1VectorArg_Imm_Narrowing_Intrinsic
                   [IntrNoMem, ImmArg<1>]>;
 
   class SVE2_2VectorArg_Imm_Narrowing_Intrinsic
       : Intrinsic<[LLVMSubdivide2VectorType<0>],
                   [LLVMSubdivide2VectorType<0>, llvm_anyvector_ty,
                    llvm_i32_ty],
                   [IntrNoMem, ImmArg<2>]>;
 
+  class SVE2_CONFLICT_DETECT_Intrinsic
+    : Intrinsic<[llvm_anyvector_ty],
+                [LLVMAnyPointerType<llvm_any_ty>,
+                 LLVMMatchType<1>]>;
+
   class SVE2_3VectorArg_Indexed_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
                 [LLVMMatchType<0>,
                  LLVMSubdivide2VectorType<0>,
                  LLVMSubdivide2VectorType<0>,
                  llvm_i32_ty],
                 [IntrNoMem, ImmArg<3>]>;
 

 // SVE2 - Optional bit permutation
 //
 
 def int_aarch64_sve_bdep_x : AdvSIMD_2VectorArg_Intrinsic;
 def int_aarch64_sve_bext_x : AdvSIMD_2VectorArg_Intrinsic;
 def int_aarch64_sve_bgrp_x : AdvSIMD_2VectorArg_Intrinsic;
 
 }
+
+//
+// SVE2 - Contiguous conflict detection
+//
+
+def int_aarch64_sve_whilerw_b : SVE2_CONFLICT_DETECT_Intrinsic;
+def int_aarch64_sve_whilerw_h : SVE2_CONFLICT_DETECT_Intrinsic;
+def int_aarch64_sve_whilerw_s : SVE2_CONFLICT_DETECT_Intrinsic;
+def int_aarch64_sve_whilerw_d : SVE2_CONFLICT_DETECT_Intrinsic;
+def int_aarch64_sve_whilewr_b : SVE2_CONFLICT_DETECT_Intrinsic;
+def int_aarch64_sve_whilewr_h : SVE2_CONFLICT_DETECT_Intrinsic;
+def int_aarch64_sve_whilewr_s : SVE2_CONFLICT_DETECT_Intrinsic;
+def int_aarch64_sve_whilewr_d : SVE2_CONFLICT_DETECT_Intrinsic;

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

let Predicates = [HasSVE2] in {
   defm WHILEHI_PWW : sve_int_while4_rr<0b101, "whilehi", int_aarch64_sve_whilehi>;
 
   defm WHILEGE_PXX : sve_int_while8_rr<0b000, "whilege", int_aarch64_sve_whilege>;
   defm WHILEGT_PXX : sve_int_while8_rr<0b001, "whilegt", int_aarch64_sve_whilegt>;
   defm WHILEHS_PXX : sve_int_while8_rr<0b100, "whilehs", int_aarch64_sve_whilehs>;
   defm WHILEHI_PXX : sve_int_while8_rr<0b101, "whilehi", int_aarch64_sve_whilehi>;
 
   // SVE2 pointer conflict compare
-  defm WHILEWR_PXX : sve2_int_while_rr<0b0, "whilewr">;
-  defm WHILERW_PXX : sve2_int_while_rr<0b1, "whilerw">;
+  defm WHILEWR_PXX : sve2_int_while_rr<0b0, "whilewr", "int_aarch64_sve_whilewr">;
+  defm WHILERW_PXX : sve2_int_while_rr<0b1, "whilerw", "int_aarch64_sve_whilerw">;
 }
 
 let Predicates = [HasSVE2AES] in {
   // SVE2 crypto destructive binary operations
   defm AESE_ZZZ_B : sve2_crypto_des_bin_op<0b00, "aese", ZPR8, int_aarch64_sve_aese, nxv16i8>;
   defm AESD_ZZZ_B : sve2_crypto_des_bin_op<0b01, "aesd", ZPR8, int_aarch64_sve_aesd, nxv16i8>;
 
   // SVE2 crypto unary operations

llvm/lib/Target/AArch64/SVEInstrFormats.td

: I<(outs pprty:$Pd), (ins GPR64:$Rn, GPR64:$Rm),
   let Inst{15-10} = 0b001100;
   let Inst{9-5}   = Rn;
   let Inst{4}     = rw;
   let Inst{3-0}   = Pd;
 
   let Defs = [NZCV];
 }
 
-multiclass sve2_int_while_rr<bits<1> rw, string asm> {
+multiclass sve2_int_while_rr<bits<1> rw, string asm, string op> {
   def _B : sve2_int_while_rr<0b00, rw, asm, PPR8>;
   def _H : sve2_int_while_rr<0b01, rw, asm, PPR16>;
   def _S : sve2_int_while_rr<0b10, rw, asm, PPR32>;
   def _D : sve2_int_while_rr<0b11, rw, asm, PPR64>;
+
+  def : SVE_2_Op_Pat<nxv16i1, !cast<SDPatternOperator>(op # _b), i64, i64, !cast<Instruction>(NAME # _B)>;
+  def : SVE_2_Op_Pat<nxv8i1,  !cast<SDPatternOperator>(op # _h), i64, i64, !cast<Instruction>(NAME # _H)>;
+  def : SVE_2_Op_Pat<nxv4i1,  !cast<SDPatternOperator>(op # _s), i64, i64, !cast<Instruction>(NAME # _S)>;
+  def : SVE_2_Op_Pat<nxv2i1,  !cast<SDPatternOperator>(op # _d), i64, i64, !cast<Instruction>(NAME # _D)>;
+
 }
 
 //===----------------------------------------------------------------------===//
 // SVE Floating Point Fast Reduction Group
 //===----------------------------------------------------------------------===//
 
 class sve_fp_fast_red<bits<2> sz, bits<3> opc, string asm,
                       ZPRRegOp zprty, RegisterClass dstRegClass>

llvm/test/CodeGen/AArch64/sve2-intrinsics-contiguous-conflict-detection.ll

+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve2 -asm-verbose=0 < %s | FileCheck %s
+
+;
+; WHILERW
+;
+
+define <vscale x 16 x i1> @whilerw_i8(i8* %a, i8* %b) {
+; CHECK-LABEL: whilerw_i8:
+; CHECK: whilerw  p0.b, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 16 x i1> @llvm.aarch64.sve.whilerw.b.nx16i1(i8* %a, i8* %b)
+  ret <vscale x 16 x i1> %out
+}
+
+define <vscale x 8 x i1> @whilerw_i16(i16* %a, i16* %b) {
+; CHECK-LABEL: whilerw_i16:
+; CHECK: whilerw  p0.h, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x i1> @llvm.aarch64.sve.whilerw.h.nx8i1(i16* %a, i16* %b)
+  ret <vscale x 8 x i1> %out
+}
+
+define <vscale x 4 x i1> @whilerw_i32(i32* %a, i32* %b) {
+; CHECK-LABEL: whilerw_i32:
+; CHECK: whilerw  p0.s, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i1> @llvm.aarch64.sve.whilerw.s.nx4i1(i32* %a, i32* %b)
+  ret <vscale x 4 x i1> %out
+}
+
+define <vscale x 2 x i1> @whilerw_i64(i64* %a, i64* %b) {
+; CHECK-LABEL: whilerw_i64:
+; CHECK: whilerw  p0.d, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i1> @llvm.aarch64.sve.whilerw.d.nx2i1(i64* %a, i64* %b)
+  ret <vscale x 2 x i1> %out
+}
+
+define <vscale x 8 x i1> @whilerw_half(half* %a, half* %b) {
+; CHECK-LABEL: whilerw_half:
+; CHECK: whilerw  p0.h, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x i1> @llvm.aarch64.sve.whilerw.h.nx8i1.f16.f16(half* %a, half* %b)
+  ret <vscale x 8 x i1> %out
+}
+
+define <vscale x 4 x i1> @whilerw_float(float* %a, float* %b) {
+; CHECK-LABEL: whilerw_float:
+; CHECK: whilerw  p0.s, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i1> @llvm.aarch64.sve.whilerw.s.nx4i1.f32.f32(float* %a, float* %b)
+  ret <vscale x 4 x i1> %out
+}
+
+define <vscale x 2 x i1> @whilerw_double(double* %a, double* %b) {
+; CHECK-LABEL: whilerw_double:
+; CHECK: whilerw  p0.d, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i1> @llvm.aarch64.sve.whilerw.d.nx2i1.f64.f64(double* %a, double* %b)
+  ret <vscale x 2 x i1> %out
+}
+
+;
+; WHILEWR
+;
+
+define <vscale x 16 x i1> @whilewr_i8(i8* %a, i8* %b) {
+; CHECK-LABEL: whilewr_i8:
+; CHECK: whilewr  p0.b, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 16 x i1> @llvm.aarch64.sve.whilewr.b.nx16i1(i8* %a, i8* %b)
+  ret <vscale x 16 x i1> %out
+}
+
+define <vscale x 8 x i1> @whilewr_i16(i16* %a, i16* %b) {
+; CHECK-LABEL: whilewr_i16:
+; CHECK: whilewr  p0.h, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x i1> @llvm.aarch64.sve.whilewr.h.nx8i1(i16* %a, i16* %b)
+  ret <vscale x 8 x i1> %out
+}
+
+define <vscale x 4 x i1> @whilewr_i32(i32* %a, i32* %b) {
+; CHECK-LABEL: whilewr_i32:
+; CHECK: whilewr  p0.s, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i1> @llvm.aarch64.sve.whilewr.s.nx4i1(i32* %a, i32* %b)
+  ret <vscale x 4 x i1> %out
+}
+
+define <vscale x 2 x i1> @whilewr_i64(i64* %a, i64* %b) {
+; CHECK-LABEL: whilewr_i64:
+; CHECK: whilewr  p0.d, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i1> @llvm.aarch64.sve.whilewr.d.nx2i1(i64* %a, i64* %b)
+  ret <vscale x 2 x i1> %out
+}
+
+define <vscale x 8 x i1> @whilewr_half(half* %a, half* %b) {
+; CHECK-LABEL: whilewr_half:
+; CHECK: whilewr  p0.h, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x i1> @llvm.aarch64.sve.whilewr.h.nx8i1.f16.f16(half* %a, half* %b)
+  ret <vscale x 8 x i1> %out
+}
+
+define <vscale x 4 x i1> @whilewr_float(float* %a, float* %b) {
+; CHECK-LABEL: whilewr_float:
+; CHECK: whilewr  p0.s, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i1> @llvm.aarch64.sve.whilewr.s.nx4i1.f32.f32(float* %a, float* %b)
+  ret <vscale x 4 x i1> %out
+}
+
+define <vscale x 2 x i1> @whilewr_double(double* %a, double* %b) {
+; CHECK-LABEL: whilewr_double:
+; CHECK: whilewr  p0.d, x0, x1
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i1> @llvm.aarch64.sve.whilewr.d.nx2i1.f64.f64(double* %a, double* %b)
+  ret <vscale x 2 x i1> %out
+}
+
+declare <vscale x 16 x i1> @llvm.aarch64.sve.whilerw.b.nx16i1(i8* %a, i8* %b)
+declare <vscale x 8 x i1> @llvm.aarch64.sve.whilerw.h.nx8i1(i16* %a, i16* %b)
+declare <vscale x 4 x i1> @llvm.aarch64.sve.whilerw.s.nx4i1(i32* %a, i32* %b)
+declare <vscale x 2 x i1> @llvm.aarch64.sve.whilerw.d.nx2i1(i64* %a, i64* %b)
+
+declare <vscale x 8 x i1> @llvm.aarch64.sve.whilerw.h.nx8i1.f16.f16(half* %a, half* %b)
+declare <vscale x 4 x i1> @llvm.aarch64.sve.whilerw.s.nx4i1.f32.f32(float* %a, float* %b)
+declare <vscale x 2 x i1> @llvm.aarch64.sve.whilerw.d.nx2i1.f64.f64(double* %a, double* %b)
+
+declare <vscale x 16 x i1> @llvm.aarch64.sve.whilewr.b.nx16i1(i8* %a, i8* %b)
+declare <vscale x 8 x i1> @llvm.aarch64.sve.whilewr.h.nx8i1(i16* %a, i16* %b)
+declare <vscale x 4 x i1> @llvm.aarch64.sve.whilewr.s.nx4i1(i32* %a, i32* %b)
+declare <vscale x 2 x i1> @llvm.aarch64.sve.whilewr.d.nx2i1(i64* %a, i64* %b)
+
+declare <vscale x 8 x i1> @llvm.aarch64.sve.whilewr.h.nx8i1.f16.f16(half* %a, half* %b)
+declare <vscale x 4 x i1> @llvm.aarch64.sve.whilewr.s.nx4i1.f32.f32(float* %a, float* %b)
+declare <vscale x 2 x i1> @llvm.aarch64.sve.whilewr.d.nx2i1.f64.f64(double* %a, double* %b)