summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--llvm/include/llvm/IR/IntrinsicsAArch64.td9
-rw-r--r--llvm/lib/Target/AArch64/AArch64InstrFormats.td40
-rw-r--r--llvm/lib/Target/AArch64/AArch64InstrInfo.td12
-rw-r--r--llvm/test/CodeGen/AArch64/neon-dot-product.ll126
4 files changed, 174 insertions, 13 deletions
diff --git a/llvm/include/llvm/IR/IntrinsicsAArch64.td b/llvm/include/llvm/IR/IntrinsicsAArch64.td
index 50341338c39..994671419cd 100644
--- a/llvm/include/llvm/IR/IntrinsicsAArch64.td
+++ b/llvm/include/llvm/IR/IntrinsicsAArch64.td
@@ -149,6 +149,11 @@ let TargetPrefix = "aarch64" in { // All intrinsics start with "llvm.aarch64.".
class AdvSIMD_1Arg_Intrinsic
: Intrinsic<[llvm_any_ty], [LLVMMatchType<0>], [IntrNoMem]>;
+
+ class AdvSIMD_Dot_Intrinsic
+ : Intrinsic<[llvm_anyvector_ty],
+ [LLVMMatchType<0>, llvm_anyvector_ty, LLVMMatchType<1>],
+ [IntrNoMem]>;
}
// Arithmetic ops
@@ -415,6 +420,10 @@ let TargetPrefix = "aarch64", IntrProperties = [IntrNoMem] in {
// Scalar FP Inexact Narrowing
def int_aarch64_sisd_fcvtxn : Intrinsic<[llvm_float_ty], [llvm_double_ty],
[IntrNoMem]>;
+
+ // v8.2-A Dot Product
+ def int_aarch64_neon_udot : AdvSIMD_Dot_Intrinsic;
+ def int_aarch64_neon_sdot : AdvSIMD_Dot_Intrinsic;
}
let TargetPrefix = "aarch64" in { // All intrinsics start with "llvm.aarch64.".
diff --git a/llvm/lib/Target/AArch64/AArch64InstrFormats.td b/llvm/lib/Target/AArch64/AArch64InstrFormats.td
index a1b34379b5b..f1fd6c958e9 100644
--- a/llvm/lib/Target/AArch64/AArch64InstrFormats.td
+++ b/llvm/lib/Target/AArch64/AArch64InstrFormats.td
@@ -4595,11 +4595,24 @@ class BaseSIMDThreeSameVectorTied<bit Q, bit U, bits<3> size, bits<5> opcode,
}
class BaseSIMDThreeSameVectorDot<bit Q, bit U, string asm, string kind1,
- string kind2> :
- BaseSIMDThreeSameVector<Q, U, 0b100, 0b10010, V128, asm, kind1, [] > {
+ string kind2, RegisterOperand RegType,
+ ValueType AccumType, ValueType InputType,
+ SDPatternOperator OpNode> :
+ BaseSIMDThreeSameVectorTied<Q, U, 0b100, 0b10010, RegType, asm, kind1,
+ [(set (AccumType RegType:$dst),
+ (OpNode (AccumType RegType:$Rd),
+ (InputType RegType:$Rn),
+ (InputType RegType:$Rm)))]> {
let AsmString = !strconcat(asm, "{\t$Rd" # kind1 # ", $Rn" # kind2 # ", $Rm" # kind2 # "}");
}
+multiclass SIMDThreeSameVectorDot<bit U, string asm, SDPatternOperator OpNode> {
+ def v8i8 : BaseSIMDThreeSameVectorDot<0, U, asm, ".2s", ".8b", V64,
+ v2i32, v8i8, OpNode>;
+ def v16i8 : BaseSIMDThreeSameVectorDot<1, U, asm, ".4s", ".16b", V128,
+ v4i32, v16i8, OpNode>;
+}
+
// All operand sizes distinguished in the encoding.
multiclass SIMDThreeSameVector<bit U, bits<5> opc, string asm,
SDPatternOperator OpNode> {
@@ -7029,14 +7042,31 @@ class BaseSIMDIndexedTied<bit Q, bit U, bit Scalar, bits<2> size, bits<4> opc,
// ARMv8.2 Index Dot product instructions
class BaseSIMDThreeSameVectorDotIndex<bit Q, bit U, string asm, string dst_kind,
- string lhs_kind, string rhs_kind> :
- BaseSIMDIndexedTied<Q, U, 0b0, 0b10, 0b1110, V128, V128, V128, VectorIndexS,
- asm, "", dst_kind, lhs_kind, rhs_kind, []> {
+ string lhs_kind, string rhs_kind,
+ RegisterOperand RegType,
+ ValueType AccumType, ValueType InputType,
+ SDPatternOperator OpNode> :
+ BaseSIMDIndexedTied<Q, U, 0b0, 0b10, 0b1110, RegType, RegType, V128,
+ VectorIndexS, asm, "", dst_kind, lhs_kind, rhs_kind,
+ [(set (AccumType RegType:$dst),
+ (AccumType (OpNode (AccumType RegType:$Rd),
+ (InputType RegType:$Rn),
+ (InputType (bitconvert (AccumType
+ (AArch64duplane32 (v4i32 V128:$Rm),
+ VectorIndexS:$idx)))))))]> {
bits<2> idx;
let Inst{21} = idx{0}; // L
let Inst{11} = idx{1}; // H
}
+multiclass SIMDThreeSameVectorDotIndex<bit U, string asm,
+ SDPatternOperator OpNode> {
+ def v8i8 : BaseSIMDThreeSameVectorDotIndex<0, U, asm, ".2s", ".8b", ".4b", V64,
+ v2i32, v8i8, OpNode>;
+ def v16i8 : BaseSIMDThreeSameVectorDotIndex<1, U, asm, ".4s", ".16b", ".4b", V128,
+ v4i32, v16i8, OpNode>;
+}
+
multiclass SIMDFPIndexed<bit U, bits<4> opc, string asm,
SDPatternOperator OpNode> {
let Predicates = [HasNEON, HasFullFP16] in {
diff --git a/llvm/lib/Target/AArch64/AArch64InstrInfo.td b/llvm/lib/Target/AArch64/AArch64InstrInfo.td
index 8eba86aa1b9..d7b9a804247 100644
--- a/llvm/lib/Target/AArch64/AArch64InstrInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64InstrInfo.td
@@ -453,14 +453,10 @@ def ISB : CRmSystemI<barrier_op, 0b110, "isb",
// ARMv8.2 Dot Product
let Predicates = [HasDotProd] in {
-def UDOT2S : BaseSIMDThreeSameVectorDot<0, 1, "udot", ".2s", ".8b">;
-def SDOT2S : BaseSIMDThreeSameVectorDot<0, 0, "sdot", ".2s", ".8b">;
-def UDOT4S : BaseSIMDThreeSameVectorDot<1, 1, "udot", ".4s", ".16b">;
-def SDOT4S : BaseSIMDThreeSameVectorDot<1, 0, "sdot", ".4s", ".16b">;
-def UDOTIDX2S : BaseSIMDThreeSameVectorDotIndex<0, 1, "udot", ".2s", ".8b", ".4b">;
-def SDOTIDX2S : BaseSIMDThreeSameVectorDotIndex<0, 0, "sdot", ".2s", ".8b", ".4b">;
-def UDOTIDX4S : BaseSIMDThreeSameVectorDotIndex<1, 1, "udot", ".4s", ".16b", ".4b">;
-def SDOTIDX4S : BaseSIMDThreeSameVectorDotIndex<1, 0, "sdot", ".4s", ".16b", ".4b">;
+defm SDOT : SIMDThreeSameVectorDot<0, "sdot", int_aarch64_neon_sdot>;
+defm UDOT : SIMDThreeSameVectorDot<1, "udot", int_aarch64_neon_udot>;
+defm SDOTlane : SIMDThreeSameVectorDotIndex<0, "sdot", int_aarch64_neon_sdot>;
+defm UDOTlane : SIMDThreeSameVectorDotIndex<1, "udot", int_aarch64_neon_udot>;
}
let Predicates = [HasRCPC] in {
diff --git a/llvm/test/CodeGen/AArch64/neon-dot-product.ll b/llvm/test/CodeGen/AArch64/neon-dot-product.ll
new file mode 100644
index 00000000000..5ad75a7d24c
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/neon-dot-product.ll
@@ -0,0 +1,126 @@
+; RUN: llc -mtriple aarch64-none-linux-gnu -mattr=+dotprod < %s | FileCheck %s
+
+declare <2 x i32> @llvm.aarch64.neon.udot.v2i32.v8i8(<2 x i32>, <8 x i8>, <8 x i8>)
+declare <4 x i32> @llvm.aarch64.neon.udot.v4i32.v16i8(<4 x i32>, <16 x i8>, <16 x i8>)
+declare <2 x i32> @llvm.aarch64.neon.sdot.v2i32.v8i8(<2 x i32>, <8 x i8>, <8 x i8>)
+declare <4 x i32> @llvm.aarch64.neon.sdot.v4i32.v16i8(<4 x i32>, <16 x i8>, <16 x i8>)
+
+define <2 x i32> @test_vdot_u32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c) #0 {
+entry:
+; CHECK-LABEL: test_vdot_u32:
+; CHECK: udot v0.2s, v1.8b, v2.8b
+ %vdot1.i = call <2 x i32> @llvm.aarch64.neon.udot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c) #2
+ ret <2 x i32> %vdot1.i
+}
+
+define <4 x i32> @test_vdotq_u32(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c) #0 {
+entry:
+; CHECK-LABEL: test_vdotq_u32:
+; CHECK: udot v0.4s, v1.16b, v2.16b
+ %vdot1.i = call <4 x i32> @llvm.aarch64.neon.udot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c) #2
+ ret <4 x i32> %vdot1.i
+}
+
+define <2 x i32> @test_vdot_s32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c) #0 {
+entry:
+; CHECK-LABEL: test_vdot_s32:
+; CHECK: sdot v0.2s, v1.8b, v2.8b
+ %vdot1.i = call <2 x i32> @llvm.aarch64.neon.sdot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c) #2
+ ret <2 x i32> %vdot1.i
+}
+
+define <4 x i32> @test_vdotq_s32(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c) #0 {
+entry:
+; CHECK-LABEL: test_vdotq_s32:
+; CHECK: sdot v0.4s, v1.16b, v2.16b
+ %vdot1.i = call <4 x i32> @llvm.aarch64.neon.sdot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c) #2
+ ret <4 x i32> %vdot1.i
+}
+
+define <2 x i32> @test_vdot_lane_u32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c) {
+entry:
+; CHECK-LABEL: test_vdot_lane_u32:
+; CHECK: udot v0.2s, v1.8b, v2.4b[1]
+ %.cast = bitcast <8 x i8> %c to <2 x i32>
+ %shuffle = shufflevector <2 x i32> %.cast, <2 x i32> undef, <2 x i32> <i32 1, i32 1>
+ %.cast5 = bitcast <2 x i32> %shuffle to <8 x i8>
+ %vdot1.i = call <2 x i32> @llvm.aarch64.neon.udot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> %.cast5) #2
+ ret <2 x i32> %vdot1.i
+}
+
+define <4 x i32> @test_vdotq_lane_u32(<4 x i32> %a, <16 x i8> %b, <8 x i8> %c) {
+entry:
+; CHECK-LABEL: test_vdotq_lane_u32:
+; CHECK: udot v0.4s, v1.16b, v2.4b[1]
+ %.cast = bitcast <8 x i8> %c to <2 x i32>
+ %shuffle = shufflevector <2 x i32> %.cast, <2 x i32> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
+ %.cast3 = bitcast <4 x i32> %shuffle to <16 x i8>
+ %vdot1.i = call <4 x i32> @llvm.aarch64.neon.udot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> %.cast3) #2
+ ret <4 x i32> %vdot1.i
+}
+
+define <2 x i32> @test_vdot_laneq_u32(<2 x i32> %a, <8 x i8> %b, <16 x i8> %c) {
+entry:
+; CHECK-LABEL: test_vdot_laneq_u32:
+; CHECK: udot v0.2s, v1.8b, v2.4b[1]
+ %.cast = bitcast <16 x i8> %c to <4 x i32>
+ %shuffle = shufflevector <4 x i32> %.cast, <4 x i32> undef, <2 x i32> <i32 1, i32 1>
+ %.cast5 = bitcast <2 x i32> %shuffle to <8 x i8>
+ %vdot1.i = call <2 x i32> @llvm.aarch64.neon.udot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> %.cast5) #2
+ ret <2 x i32> %vdot1.i
+}
+
+define <4 x i32> @test_vdotq_laneq_u32(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c) {
+entry:
+; CHECK-LABEL: test_vdotq_laneq_u32:
+; CHECK: udot v0.4s, v1.16b, v2.4b[1]
+ %.cast = bitcast <16 x i8> %c to <4 x i32>
+ %shuffle = shufflevector <4 x i32> %.cast, <4 x i32> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
+ %.cast3 = bitcast <4 x i32> %shuffle to <16 x i8>
+ %vdot1.i = call <4 x i32> @llvm.aarch64.neon.udot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> %.cast3) #2
+ ret <4 x i32> %vdot1.i
+}
+
+define <2 x i32> @test_vdot_lane_s32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c) {
+entry:
+; CHECK-LABEL: test_vdot_lane_s32:
+; CHECK: sdot v0.2s, v1.8b, v2.4b[1]
+ %.cast = bitcast <8 x i8> %c to <2 x i32>
+ %shuffle = shufflevector <2 x i32> %.cast, <2 x i32> undef, <2 x i32> <i32 1, i32 1>
+ %.cast5 = bitcast <2 x i32> %shuffle to <8 x i8>
+ %vdot1.i = call <2 x i32> @llvm.aarch64.neon.sdot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> %.cast5) #2
+ ret <2 x i32> %vdot1.i
+}
+
+define <4 x i32> @test_vdotq_lane_s32(<4 x i32> %a, <16 x i8> %b, <8 x i8> %c) {
+entry:
+; CHECK-LABEL: test_vdotq_lane_s32:
+; CHECK: sdot v0.4s, v1.16b, v2.4b[1]
+ %.cast = bitcast <8 x i8> %c to <2 x i32>
+ %shuffle = shufflevector <2 x i32> %.cast, <2 x i32> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
+ %.cast3 = bitcast <4 x i32> %shuffle to <16 x i8>
+ %vdot1.i = call <4 x i32> @llvm.aarch64.neon.sdot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> %.cast3) #2
+ ret <4 x i32> %vdot1.i
+}
+
+define <2 x i32> @test_vdot_laneq_s32(<2 x i32> %a, <8 x i8> %b, <16 x i8> %c) {
+entry:
+; CHECK-LABEL: test_vdot_laneq_s32:
+; CHECK: sdot v0.2s, v1.8b, v2.4b[1]
+ %.cast = bitcast <16 x i8> %c to <4 x i32>
+ %shuffle = shufflevector <4 x i32> %.cast, <4 x i32> undef, <2 x i32> <i32 1, i32 1>
+ %.cast5 = bitcast <2 x i32> %shuffle to <8 x i8>
+ %vdot1.i = call <2 x i32> @llvm.aarch64.neon.sdot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> %.cast5) #2
+ ret <2 x i32> %vdot1.i
+}
+
+define <4 x i32> @test_vdotq_laneq_s32(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c) {
+entry:
+; CHECK-LABEL: test_vdotq_laneq_s32:
+; CHECK: sdot v0.4s, v1.16b, v2.4b[1]
+ %.cast = bitcast <16 x i8> %c to <4 x i32>
+ %shuffle = shufflevector <4 x i32> %.cast, <4 x i32> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
+ %.cast3 = bitcast <4 x i32> %shuffle to <16 x i8>
+ %vdot1.i = call <4 x i32> @llvm.aarch64.neon.sdot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> %.cast3) #2
+ ret <4 x i32> %vdot1.i
+}
OpenPOWER on IntegriCloud