[AArch64] Add EXT patterns for 64-bit EXT of a subvector of a 128-bit vector

[john.brawn]

Summary

If we have a 64-bit EXT where one of the operands is a subvector of a 128-bit vector then in some cases we can eliminate an extract_subvector by converting to a 128-bit EXT of the 128-bit vector.

[dmgreen]

I've managed to convince myself that this looks OK. A couple of nits depending on what you think of them.

Otherwise LGTM.

lib/Target/AArch64/AArch64InstrInfo.td

4164

Perhaps call this Add8 or something like it? Up to you.

4165

It looks like the i64 from here is an i32 everywhere else

lib/Target/AArch64/AArch64InstrInfo.td

def : Pat<(concat_vectors (v2i32 V64:$Rd),
                             V128:$Rn, V128:$Rm)>;
 
 //----------------------------------------------------------------------------
 // AdvSIMD bitwise extract from vector instruction.
 //----------------------------------------------------------------------------
 
 defm EXT : SIMDBitwiseExtract<"ext">;
 
+def AdjustExtImm : SDNodeXForm<imm, [{
+  return CurDAG->getTargetConstant(8 + N->getZExtValue(), SDLoc(N), MVT::i64);
+}]>;
 multiclass ExtPat<ValueType VT64, ValueType VT128, int N> {
   def : Pat<(VT64 (AArch64ext V64:$Rn, V64:$Rm, (i32 imm:$imm))),
             (EXTv8i8 V64:$Rn, V64:$Rm, imm:$imm)>;
   def : Pat<(VT128 (AArch64ext V128:$Rn, V128:$Rm, (i32 imm:$imm))),
             (EXTv16i8 V128:$Rn, V128:$Rm, imm:$imm)>;
   // We use EXT to handle extract_subvector to copy the upper 64-bits of a
   // 128-bit vector.
   def : Pat<(VT64 (extract_subvector V128:$Rn, (i64 N))),
             (EXTRACT_SUBREG (EXTv16i8 V128:$Rn, V128:$Rn, 8), dsub)>;
+  // A 64-bit EXT of two halves of the same 128-bit register can be done as a
+  // single 128-bit EXT.
+  def : Pat<(VT64 (AArch64ext (extract_subvector V128:$Rn, (i64 0)),
+                              (extract_subvector V128:$Rn, (i64 N)),
+                              (i32 imm:$imm))),
+            (EXTRACT_SUBREG (EXTv16i8 V128:$Rn, V128:$Rn, imm:$imm), dsub)>;
+  // A 64-bit EXT of the high half of a 128-bit register can be done using a
+  // 128-bit EXT of the whole register with an adjustment to the immediate. The
+  // top half of the other operand will be unset, but that doesn't matter as it
+  // will not be used.
+  def : Pat<(VT64 (AArch64ext (extract_subvector V128:$Rn, (i64 N)),
+                              V64:$Rm,
+                              (i32 imm:$imm))),
+            (EXTRACT_SUBREG (EXTv16i8 V128:$Rn,
+                                      (SUBREG_TO_REG (i32 0), V64:$Rm, dsub),
+                                      (AdjustExtImm imm:$imm)), dsub)>;
 }
 
 defm : ExtPat<v8i8, v16i8, 8>;
 defm : ExtPat<v4i16, v8i16, 4>;
 defm : ExtPat<v4f16, v8f16, 4>;
 defm : ExtPat<v2i32, v4i32, 2>;
 defm : ExtPat<v2f32, v4f32, 2>;
 defm : ExtPat<v1i64, v2i64, 1>;

test/CodeGen/AArch64/ext-narrow-index.ll

+; RUN: llc < %s -mtriple=aarch64 | FileCheck %s
+
+; Tests of shufflevector where the index operand is half the width of the vector
+; operands. We should get one ext instruction and not two.
+
+; i8 tests
+define <8 x i8> @i8_off0(<16 x i8> %arg1, <16 x i8> %arg2) {
+; CHECK-LABEL: i8_off0:
+; CHECK-NOT: mov
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <16 x i8> %arg1, <16 x i8> %arg2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  ret <8 x i8> %shuffle
+}
+
+define <8 x i8> @i8_off1(<16 x i8> %arg1, <16 x i8> %arg2) {
+; CHECK-LABEL: i8_off1:
+; CHECK-NOT: mov
+; CHECK: ext v0.16b, v0.16b, v0.16b, #1
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <16 x i8> %arg1, <16 x i8> %arg2, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>
+  ret <8 x i8> %shuffle
+}
+
+define <8 x i8> @i8_off8(<16 x i8> %arg1, <16 x i8> %arg2) {
+; CHECK-LABEL: i8_off8:
+; CHECK-NOT: mov
+; CHECK: ext v0.16b, v0.16b, v0.16b, #8
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <16 x i8> %arg1, <16 x i8> %arg2, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  ret <8 x i8> %shuffle
+}
+
+define <8 x i8> @i8_off15(<16 x i8> %arg1, <16 x i8> %arg2) {
+; CHECK-LABEL: i8_off15:
+; CHECK: ext v0.16b, v0.16b, v1.16b, #15
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <16 x i8> %arg1, <16 x i8> %arg2, <8 x i32> <i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22>
+  ret <8 x i8> %shuffle
+}
+
+define <8 x i8> @i8_off22(<16 x i8> %arg1, <16 x i8> %arg2) {
+; CHECK-LABEL: i8_off22:
+; CHECK: ext v0.16b, v1.16b, v1.16b, #6
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <16 x i8> %arg1, <16 x i8> %arg2, <8 x i32> <i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29>
+  ret <8 x i8> %shuffle
+}
+
+; i16 tests
+define <4 x i16> @i16_off0(<8 x i16> %arg1, <8 x i16> %arg2) {
+; CHECK-LABEL: i16_off0:
+; CHECK-NOT: mov
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <8 x i16> %arg1, <8 x i16> %arg2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  ret <4 x i16> %shuffle
+}
+
+define <4 x i16> @i16_off1(<8 x i16> %arg1, <8 x i16> %arg2) {
+; CHECK-LABEL: i16_off1:
+; CHECK-NOT: mov
+; CHECK: ext v0.16b, v0.16b, v0.16b, #2
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <8 x i16> %arg1, <8 x i16> %arg2, <4 x i32> <i32 1, i32 2, i32 3, i32 4>
+  ret <4 x i16> %shuffle
+}
+
+define <4 x i16> @i16_off7(<8 x i16> %arg1, <8 x i16> %arg2) {
+; CHECK-LABEL: i16_off7:
+; CHECK: ext v0.16b, v0.16b, v1.16b, #14
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <8 x i16> %arg1, <8 x i16> %arg2, <4 x i32> <i32 7, i32 8, i32 9, i32 10>
+  ret <4 x i16> %shuffle
+}
+
+define <4 x i16> @i16_off8(<8 x i16> %arg1, <8 x i16> %arg2) {
+; CHECK-LABEL: i16_off8:
+; CHECK: mov v0.16b, v1.16b
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <8 x i16> %arg1, <8 x i16> %arg2, <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+  ret <4 x i16> %shuffle
+}
+
+; i32 tests
+define <2 x i32> @i32_off0(<4 x i32> %arg1, <4 x i32> %arg2) {
+; CHECK-LABEL: i32_off0:
+; CHECK-NOT: mov
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <4 x i32> %arg1, <4 x i32> %arg2, <2 x i32> <i32 0, i32 1>
+  ret <2 x i32> %shuffle
+}
+
+define <2 x i32> @i32_off1(<4 x i32> %arg1, <4 x i32> %arg2) {
+; CHECK-LABEL: i32_off1:
+; CHECK-NOT: mov
+; CHECK: ext v0.16b, v0.16b, v0.16b, #4
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <4 x i32> %arg1, <4 x i32> %arg2, <2 x i32> <i32 1, i32 2>
+  ret <2 x i32> %shuffle
+}
+
+define <2 x i32> @i32_off3(<4 x i32> %arg1, <4 x i32> %arg2) {
+; CHECK-LABEL: i32_off3:
+; CHECK: ext v0.16b, v0.16b, v1.16b, #12
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <4 x i32> %arg1, <4 x i32> %arg2, <2 x i32> <i32 3, i32 4>
+  ret <2 x i32> %shuffle
+}
+
+define <2 x i32> @i32_off4(<4 x i32> %arg1, <4 x i32> %arg2) {
+; CHECK-LABEL: i32_off4:
+; CHECK: mov v0.16b, v1.16b
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <4 x i32> %arg1, <4 x i32> %arg2, <2 x i32> <i32 4, i32 5>
+  ret <2 x i32> %shuffle
+}
+
+; i64 tests
+define <1 x i64> @i64_off0(<2 x i64> %arg1, <2 x i64> %arg2) {
+; CHECK-LABEL: i64_off0:
+; CHECK-NOT: mov
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <2 x i64> %arg1, <2 x i64> %arg2, <1 x i32> <i32 0>
+  ret <1 x i64> %shuffle
+}
+
+define <1 x i64> @i64_off1(<2 x i64> %arg1, <2 x i64> %arg2) {
+; CHECK-LABEL: i64_off1:
+; CHECK-NOT: mov
+; CHECK: ext v0.16b, v0.16b, v0.16b, #8
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <2 x i64> %arg1, <2 x i64> %arg2, <1 x i32> <i32 1>
+  ret <1 x i64> %shuffle
+}
+
+define <1 x i64> @i64_off2(<2 x i64> %arg1, <2 x i64> %arg2) {
+; CHECK-LABEL: i64_off2:
+; CHECK: mov v0.16b, v1.16b
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <2 x i64> %arg1, <2 x i64> %arg2, <1 x i32> <i32 2>
+  ret <1 x i64> %shuffle
+}
+
+; i8 tests with second operand zero
+define <8 x i8> @i8_zero_off0(<16 x i8> %arg1) {
+; CHECK-LABEL: i8_zero_off0:
+; CHECK-NOT: mov
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <16 x i8> %arg1, <16 x i8> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  ret <8 x i8> %shuffle
+}
+
+define <8 x i8> @i8_zero_off1(<16 x i8> %arg1) {
+; CHECK-LABEL: i8_zero_off1:
+; CHECK-NOT: mov
+; CHECK: ext v0.16b, v0.16b, v0.16b, #1
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <16 x i8> %arg1, <16 x i8> zeroinitializer, <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>
+  ret <8 x i8> %shuffle
+}
+
+define <8 x i8> @i8_zero_off8(<16 x i8> %arg1) {
+; CHECK-LABEL: i8_zero_off8:
+; CHECK-NOT: mov
+; CHECK: ext v0.16b, v0.16b, v0.16b, #8
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <16 x i8> %arg1, <16 x i8> zeroinitializer, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  ret <8 x i8> %shuffle
+}
+
+define <8 x i8> @i8_zero_off15(<16 x i8> %arg1) {
+; CHECK-LABEL: i8_zero_off15:
+; CHECK: movi [[REG:v[0-9]+]].2d, #0
+; CHECK: ext v0.16b, v0.16b, [[REG]].16b, #15
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <16 x i8> %arg1, <16 x i8> zeroinitializer, <8 x i32> <i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22>
+  ret <8 x i8> %shuffle
+}
+
+define <8 x i8> @i8_zero_off22(<16 x i8> %arg1) {
+; CHECK-LABEL: i8_zero_off22:
+; CHECK: movi v0.2d, #0
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <16 x i8> %arg1, <16 x i8> zeroinitializer, <8 x i32> <i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29>
+  ret <8 x i8> %shuffle
+}
+
+; i16 tests with second operand zero
+define <4 x i16> @i16_zero_off0(<8 x i16> %arg1) {
+; CHECK-LABEL: i16_zero_off0:
+; CHECK-NOT: mov
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <8 x i16> %arg1, <8 x i16> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  ret <4 x i16> %shuffle
+}
+
+define <4 x i16> @i16_zero_off1(<8 x i16> %arg1) {
+; CHECK-LABEL: i16_zero_off1:
+; CHECK-NOT: mov
+; CHECK: ext v0.16b, v0.16b, v0.16b, #2
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <8 x i16> %arg1, <8 x i16> zeroinitializer, <4 x i32> <i32 1, i32 2, i32 3, i32 4>
+  ret <4 x i16> %shuffle
+}
+
+define <4 x i16> @i16_zero_off7(<8 x i16> %arg1) {
+; CHECK-LABEL: i16_zero_off7:
+; CHECK: movi [[REG:v[0-9]+]].2d, #0
+; CHECK: ext v0.16b, v0.16b, [[REG]].16b, #14
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <8 x i16> %arg1, <8 x i16> zeroinitializer, <4 x i32> <i32 7, i32 8, i32 9, i32 10>
+  ret <4 x i16> %shuffle
+}
+
+define <4 x i16> @i16_zero_off8(<8 x i16> %arg1) {
+; CHECK-LABEL: i16_zero_off8:
+; CHECK: movi v0.2d, #0
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <8 x i16> %arg1, <8 x i16> zeroinitializer, <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+  ret <4 x i16> %shuffle
+}
+
+; i32 tests with second operand zero
+define <2 x i32> @i32_zero_off0(<4 x i32> %arg1) {
+; CHECK-LABEL: i32_zero_off0:
+; CHECK-NOT: mov
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <4 x i32> %arg1, <4 x i32> zeroinitializer, <2 x i32> <i32 0, i32 1>
+  ret <2 x i32> %shuffle
+}
+
+define <2 x i32> @i32_zero_off1(<4 x i32> %arg1) {
+; CHECK-LABEL: i32_zero_off1:
+; CHECK-NOT: mov
+; CHECK: ext v0.16b, v0.16b, v0.16b, #4
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <4 x i32> %arg1, <4 x i32> zeroinitializer, <2 x i32> <i32 1, i32 2>
+  ret <2 x i32> %shuffle
+}
+
+define <2 x i32> @i32_zero_off3(<4 x i32> %arg1) {
+; CHECK-LABEL: i32_zero_off3:
+; CHECK: movi [[REG:v[0-9]+]].2d, #0
+; CHECK: ext v0.16b, v0.16b, [[REG]].16b, #12
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <4 x i32> %arg1, <4 x i32> zeroinitializer, <2 x i32> <i32 3, i32 4>
+  ret <2 x i32> %shuffle
+}
+
+define <2 x i32> @i32_zero_off4(<4 x i32> %arg1) {
+; CHECK-LABEL: i32_zero_off4:
+; CHECK: movi v0.2d, #0
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <4 x i32> %arg1, <4 x i32> zeroinitializer, <2 x i32> <i32 4, i32 5>
+  ret <2 x i32> %shuffle
+}
+
+; i64 tests with second operand zero
+define <1 x i64> @i64_zero_off0(<2 x i64> %arg1) {
+; CHECK-LABEL: i64_zero_off0:
+; CHECK-NOT: mov
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <2 x i64> %arg1, <2 x i64> zeroinitializer, <1 x i32> <i32 0>
+  ret <1 x i64> %shuffle
+}
+
+define <1 x i64> @i64_zero_off1(<2 x i64> %arg1) {
+; CHECK-LABEL: i64_zero_off1:
+; CHECK-NOT: mov
+; CHECK: ext v0.16b, v0.16b, v0.16b, #8
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <2 x i64> %arg1, <2 x i64> zeroinitializer, <1 x i32> <i32 1>
+  ret <1 x i64> %shuffle
+}
+
+define <1 x i64> @i64_zero_off2(<2 x i64> %arg1) {
+; CHECK-LABEL: i64_zero_off2:
+; CHECK: fmov d0, xzr
+; CHECK-NOT: ext
+; CHECK: ret
+entry:
+  %shuffle = shufflevector <2 x i64> %arg1, <2 x i64> zeroinitializer, <1 x i32> <i32 2>
+  ret <1 x i64> %shuffle
+}