[AArch64] support neon_sshl and neon_ushl in performIntrinsicCombine.

[fhahn]

Summary

If we have a constant vector mask with the shift values being all equal,
we can simplify aarch64_neon_sshl to VSHL.

This pattern can be generated by code using vshlq_s32(a,vdupq_n_s32(n))
instead of vshlq_n_s32(a, n), because it is used in contexts where n is
not guaranteed to be constant, before inlining.

We can do a similar combine for aarch64_neon_ushl, but we have to be
a bit more careful, because we can only match ushll/ushll2 for vector
shifts with a zero-extended first operand.

Also adds 2 tests marked with FIXME, where we can further increase
codegen.

[t.p.northover]

I think there are probably other shifts that we can include while we're in the area. Most obviously aarch64_neon_ushl, but maybe others too.

[fhahn]

Add support for neon_ushl.

[fhahn]

In D62308#1513868, @t.p.northover wrote:

I think there are probably other shifts that we can include while we're in the area. Most obviously aarch64_neon_ushl, but maybe others too.

Thanks Tim. I've added support for neon_ushl. I think w have to be a bit more careful with that one, as we only have ushll/ushll2 that take immediates

[fhahn]

ping

[fhahn]

Ping

[anemet]

Does this also apply to right shifts?

[fhahn]

In D62308#1661854, @anemet wrote:

Does this also apply to right shifts?

I think the relevant cases should be handled already. For vector right shifts, it looks like there are only immediate forms and we turn left shifts with negative immediate into the corresponding right shifts.

[fhahn]

ping

[anemet]

Some minor test questions/suggestions. Feel free to commit after addressing.

llvm/test/CodeGen/AArch64/arm64-vshift.ll

1204

I don't see any negative tests when we zero-extend not to the next one higher type.

1405

technically this is not sshll (long)

1431

Isn't it used for the extensions?

1446

I think that we should also have other shl tests (.4s non-foldable and perhaps some other sizes).

[fhahn]

Add additional test cases and limit this patch to converting cases with appropriate zext/sext.

I'll submit a separate patch for turning ushl -> shl, if the shift is all constant.

[anemet]

In D62308#1676691, @fhahn wrote:

I'll submit a separate patch for turning ushl -> shl, if the shift is all constant.

... and sshl -> shl?

[fhahn]

In D62308#1676800, @anemet wrote:

In D62308#1676691, @fhahn wrote:

I'll submit a separate patch for turning ushl -> shl, if the shift is all constant.

... and sshl -> shl?

Yep, let's look at that separately.

llvm/test/CodeGen/AArch64/arm64-vshift.ll

1431

Yeah, I initially thought there might be a long version of sshr, but looks like there is not

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
   case Intrinsic::aarch64_neon_urshl:
     Opcode = AArch64ISD::URSHR_I;
     IsRightShift = true;
     break;
   case Intrinsic::aarch64_neon_sqshlu:
     Opcode = AArch64ISD::SQSHLU_I;
     IsRightShift = false;
     break;
+  case Intrinsic::aarch64_neon_sshl:
+  case Intrinsic::aarch64_neon_ushl: {
+    // ushll/ushll2 provide unsigned shifts with immediate operands and
+    // sshll/sshll2 provide signed shifts with immediates, so we have to make
+    // sure we only match patterns here we can later match to them.
+    SDValue Op0 = N->getOperand(1);
+    if (Op0.getNode()->getOpcode() != (IID == Intrinsic::aarch64_neon_ushl
+                                           ? ISD::ZERO_EXTEND
+                                           : ISD::SIGN_EXTEND))
+      return SDValue();
+
+    EVT FromType = Op0.getOperand(0).getValueType();
+    EVT ToType = Op0.getValueType();
+    unsigned FromSize = FromType.getScalarSizeInBits();
+    if (!FromType.isVector() || !ToType.isVector() ||
+        (FromSize != 8 && FromSize != 16 && FromSize != 32) ||
+        2 * FromSize != ToType.getScalarSizeInBits())
+      return SDValue();
+
+    Opcode = AArch64ISD::VSHL;
+    IsRightShift = false;
+    break;
+  }
   }
 
   if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits) {
     SDLoc dl(N);
     return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1),
                        DAG.getConstant(-ShiftAmount, dl, MVT::i32));
   } else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits) {
     SDLoc dl(N);
static SDValue performIntrinsicCombine(SDNode *N,
   case Intrinsic::aarch64_neon_pmull:
   case Intrinsic::aarch64_neon_sqdmull:
     return tryCombineLongOpWithDup(IID, N, DCI, DAG);
   case Intrinsic::aarch64_neon_sqshl:
   case Intrinsic::aarch64_neon_uqshl:
   case Intrinsic::aarch64_neon_sqshlu:
   case Intrinsic::aarch64_neon_srshl:
   case Intrinsic::aarch64_neon_urshl:
+  case Intrinsic::aarch64_neon_sshl:
+  case Intrinsic::aarch64_neon_ushl:
     return tryCombineShiftImm(IID, N, DAG);
   case Intrinsic::aarch64_crc32b:
   case Intrinsic::aarch64_crc32cb:
     return tryCombineCRC32(0xff, N, DAG);
   case Intrinsic::aarch64_crc32h:
   case Intrinsic::aarch64_crc32ch:
     return tryCombineCRC32(0xffff, N, DAG);
   }

llvm/test/CodeGen/AArch64/arm64-vshift.ll

 ;CHECK: ushll.2d v0, {{v[0-9]+}}, #1
         %load1 = load <4 x i32>, <4 x i32>* %A
         %tmp1 = shufflevector <4 x i32> %load1, <4 x i32> undef, <2 x i32> <i32 2, i32 3>
         %tmp2 = zext <2 x i32> %tmp1 to <2 x i64>
         %tmp3 = shl <2 x i64> %tmp2, <i64 1, i64 1>
         ret <2 x i64> %tmp3
 }
 
+declare <16 x i8> @llvm.aarch64.neon.ushl.v16i8(<16 x i8>, <16 x i8>)
+declare <8 x i16> @llvm.aarch64.neon.ushl.v8i16(<8 x i16>, <8 x i16>)
+declare <4 x i32> @llvm.aarch64.neon.ushl.v4i32(<4 x i32>, <4 x i32>)
+declare <2 x i64> @llvm.aarch64.neon.ushl.v2i64(<2 x i64>, <2 x i64>)
+
+define <8 x i16> @neon.ushll8h_constant_shift(<8 x i8>* %A) nounwind {
+;CHECK-LABEL: neon.ushll8h_constant_shift
+;CHECK: ushll.8h v0, {{v[0-9]+}}, #1
+  %tmp1 = load <8 x i8>, <8 x i8>* %A
+  %tmp2 = zext <8 x i8> %tmp1 to <8 x i16>
+  %tmp3 = call <8 x i16> @llvm.aarch64.neon.ushl.v8i16(<8 x i16> %tmp2, <8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>)
+  ret <8 x i16> %tmp3
+}
+
+define <8 x i16> @neon.ushl8h_no_constant_shift(<8 x i8>* %A) nounwind {
+;CHECK-LABEL: neon.ushl8h_no_constant_shift
+;CHECK: ushl.8h v0, v0, v0
+  %tmp1 = load <8 x i8>, <8 x i8>* %A
+  %tmp2 = zext <8 x i8> %tmp1 to <8 x i16>
+  %tmp3 = call <8 x i16> @llvm.aarch64.neon.ushl.v8i16(<8 x i16> %tmp2, <8 x i16> %tmp2)
+  ret <8 x i16> %tmp3
+}
+
+; Here we do not extend to the double the bitwidth, so we cannot fold to ushll.
+define <4 x i32> @neon.ushll8h_constant_shift_extend_not_2x(<4 x i8>* %A) nounwind {
+;CHECK-LABEL: @neon.ushll8h_constant_shift_extend_not_2x
+;CHECK-NOT: ushll.8h v0,
+;CHECK: ldrb  w8, [x0]
+;CHECK: movi.4s v1, #1
+;CHECK: fmov  s0, w8
+;CHECK: ldrb  w8, [x0, #1]
+;CHECK: mov.s v0[1], w8
+;CHECK: ldrb  w8, [x0, #2]
+;CHECK: mov.s v0[2], w8
+;CHECK: ldrb  w8, [x0, #3]
+;CHECK: mov.s v0[3], w8
+;CHECK: ushl.4s v0, v0, v1
+  %tmp1 = load <4 x i8>, <4 x i8>* %A
+  %tmp2 = zext <4 x i8> %tmp1 to <4 x i32>
+  %tmp3 = call <4 x i32> @llvm.aarch64.neon.ushl.v4i32(<4 x i32> %tmp2, <4 x i32> <i32 1, i32 1, i32 1, i32 1>)
+  ret <4 x i32> %tmp3
+}
+
+define <8 x i16> @neon.ushl8_noext_constant_shift(<8 x i16>* %A) nounwind {
+; CHECK-LABEL: neon.ushl8_noext_constant_shift
+; CHECK:      ldr       q0, [x0]
+; CHECK-NEXT: movi.8h   v1, #1
+; CHECK-NEXT: ushl.8h   v0, v0, v1
+; CHECK-NEXT: ret
+  %tmp1 = load <8 x i16>, <8 x i16>* %A
+  %tmp3 = call <8 x i16> @llvm.aarch64.neon.ushl.v8i16(<8 x i16> %tmp1, <8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>)
+  ret <8 x i16> %tmp3
+}
+
+define <4 x i32> @neon.ushll4s_constant_shift(<4 x i16>* %A) nounwind {
+;CHECK-LABEL: neon.ushll4s_constant_shift
+;CHECK: ushll.4s v0, {{v[0-9]+}}, #1
+  %tmp1 = load <4 x i16>, <4 x i16>* %A
+  %tmp2 = zext <4 x i16> %tmp1 to <4 x i32>
+  %tmp3 = call <4 x i32> @llvm.aarch64.neon.ushl.v4i32(<4 x i32> %tmp2, <4 x i32> <i32 1, i32 1, i32 1, i32 1>)
+  ret <4 x i32> %tmp3
+}
+
+; FIXME: unnecessary ushll.4s v0, v0, #0?
+define <4 x i32> @neon.ushll4s_neg_constant_shift(<4 x i16>* %A) nounwind {
+; CHECK-LABEL: neon.ushll4s_neg_constant_shift
+; CHECK: movi.2d v1, #0xffffffffffffffff
+; CHECK: ushll.4s v0, v0, #0
+; CHECK: ushl.4s v0, v0, v1
+  %tmp1 = load <4 x i16>, <4 x i16>* %A
+  %tmp2 = zext <4 x i16> %tmp1 to <4 x i32>
+  %tmp3 = call <4 x i32> @llvm.aarch64.neon.ushl.v4i32(<4 x i32> %tmp2, <4 x i32> <i32 -1, i32 -1, i32 -1, i32 -1>)
+  ret <4 x i32> %tmp3
+}
+
+; FIXME: should be constant folded.
+define <4 x i32> @neon.ushll4s_constant_fold() nounwind {
+; CHECK-LABEL: neon.ushll4s_constant_fold
+; CHECK:      movi.4s v1, #1
+; CHECK-NEXT: ushl.4s v0, v0, v1
+;
+  %tmp3 = call <4 x i32> @llvm.aarch64.neon.ushl.v4i32(<4 x i32> <i32 0, i32 1, i32 2, i32 3>, <4 x i32> <i32 1, i32 1, i32 1, i32 1>)
+  ret <4 x i32> %tmp3
+}
+
+define <2 x i64> @neon.ushll2d_constant_shift(<2 x i32>* %A) nounwind {
+;CHECK-LABEL: neon.ushll2d_constant_shift
+;CHECK: ushll.2d v0, {{v[0-9]+}}, #1
+  %tmp1 = load <2 x i32>, <2 x i32>* %A
+  %tmp2 = zext <2 x i32> %tmp1 to <2 x i64>
+  %tmp3 = call <2 x i64> @llvm.aarch64.neon.ushl.v2i64(<2 x i64> %tmp2, <2 x i64> <i64 1, i64 1>)
+  ret <2 x i64> %tmp3
+}
+
 define <8 x i16> @sshll8h(<8 x i8>* %A) nounwind {
 ;CHECK-LABEL: sshll8h:
 ;CHECK: sshll.8h v0, {{v[0-9]+}}, #1
         %tmp1 = load <8 x i8>, <8 x i8>* %A
         %tmp2 = sext <8 x i8> %tmp1 to <8 x i16>
         %tmp3 = shl <8 x i16> %tmp2, <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
         ret <8 x i16> %tmp3
 }
 
-define <4 x i32> @sshll4s(<4 x i16>* %A) nounwind {
-;CHECK-LABEL: sshll4s:
+define <2 x i64> @sshll2d(<2 x i32>* %A) nounwind {
+;CHECK-LABEL: sshll2d:
+;CHECK: sshll.2d v0, {{v[0-9]+}}, #1
+        %tmp1 = load <2 x i32>, <2 x i32>* %A
+        %tmp2 = sext <2 x i32> %tmp1 to <2 x i64>
+        %tmp3 = shl <2 x i64> %tmp2, <i64 1, i64 1>
+        ret <2 x i64> %tmp3
+}
+
+declare <16 x i8> @llvm.aarch64.neon.sshl.v16i8(<16 x i8>, <16 x i8>)
+declare <8 x i16> @llvm.aarch64.neon.sshl.v8i16(<8 x i16>, <8 x i16>)
+declare <4 x i32> @llvm.aarch64.neon.sshl.v4i32(<4 x i32>, <4 x i32>)
+declare <2 x i64> @llvm.aarch64.neon.sshl.v2i64(<2 x i64>, <2 x i64>)
+
+define <16 x i8> @neon.sshl16b_constant_shift(<16 x i8>* %A) nounwind {
+;CHECK-LABEL: neon.sshl16b_constant_shift
+;CHECK: sshl.16b {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+        %tmp1 = load <16 x i8>, <16 x i8>* %A
+        %tmp2 = call <16 x i8> @llvm.aarch64.neon.sshl.v16i8(<16 x i8> %tmp1, <16 x i8> <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>)
+        ret <16 x i8> %tmp2
+}
+
+define <8 x i16> @neon.sshll8h_constant_shift(<8 x i8>* %A) nounwind {
+;CHECK-LABEL: neon.sshll8h_constant_shift
+;CHECK: sshll.8h v0, {{v[0-9]+}}, #1
+        %tmp1 = load <8 x i8>, <8 x i8>* %A
+        %tmp2 = sext <8 x i8> %tmp1 to <8 x i16>
+        %tmp3 = call <8 x i16> @llvm.aarch64.neon.sshl.v8i16(<8 x i16> %tmp2, <8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>)
+        ret <8 x i16> %tmp3
+}
+
+define <4 x i32> @neon.sshl4s_wrong_ext_constant_shift(<4 x i8>* %A) nounwind {
+;CHECK-LABEL: neon.sshl4s_wrong_ext_constant_shift
+;CHECK: sshl.4s {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+        %tmp1 = load <4 x i8>, <4 x i8>* %A
+        %tmp2 = sext <4 x i8> %tmp1 to <4 x i32>
+        %tmp3 = call <4 x i32> @llvm.aarch64.neon.sshl.v4i32(<4 x i32> %tmp2, <4 x i32> <i32 1, i32 1, i32 1, i32 1>)
+        ret <4 x i32> %tmp3
+}
+
+
+define <4 x i32> @neon.sshll4s_constant_shift(<4 x i16>* %A) nounwind {
+;CHECK-LABEL: neon.sshll4s_constant_shift
 ;CHECK: sshll.4s v0, {{v[0-9]+}}, #1
         %tmp1 = load <4 x i16>, <4 x i16>* %A
         %tmp2 = sext <4 x i16> %tmp1 to <4 x i32>
-        %tmp3 = shl <4 x i32> %tmp2, <i32 1, i32 1, i32 1, i32 1>
+        %tmp3 = call <4 x i32> @llvm.aarch64.neon.sshl.v4i32(<4 x i32> %tmp2, <4 x i32> <i32 1, i32 1, i32 1, i32 1>)
         ret <4 x i32> %tmp3
 }
 
-define <2 x i64> @sshll2d(<2 x i32>* %A) nounwind {
-;CHECK-LABEL: sshll2d:
+define <4 x i32> @neon.sshll4s_neg_constant_shift(<4 x i16>* %A) nounwind {
+;CHECK-LABEL: neon.sshll4s_neg_constant_shift
+;CHECK: movi.2d v1, #0xffffffffffffffff
+;CHECK: sshll.4s v0, v0, #0
+;CHECK: sshl.4s v0, v0, v1
+        %tmp1 = load <4 x i16>, <4 x i16>* %A
+        %tmp2 = sext <4 x i16> %tmp1 to <4 x i32>
+        %tmp3 = call <4 x i32> @llvm.aarch64.neon.sshl.v4i32(<4 x i32> %tmp2, <4 x i32> <i32 -1, i32 -1, i32 -1, i32 -1>)
+        ret <4 x i32> %tmp3
+}
+
+; FIXME: should be constant folded.
+define <4 x i32> @neon.sshl4s_constant_fold() nounwind {
+;CHECK-LABEL: neon.sshl4s_constant_fold
+;CHECK: sshl.4s {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+        %tmp3 = call <4 x i32> @llvm.aarch64.neon.sshl.v4i32(<4 x i32> <i32 0, i32 1, i32 2, i32 3>, <4 x i32> <i32 1, i32 1, i32 1, i32 1>)
+        ret <4 x i32> %tmp3
+}
+
+define <4 x i32> @neon.sshl4s_no_fold(<4 x i32>* %A) nounwind {
+;CHECK-LABEL: neon.sshl4s_no_fold
+;CHECK: sshl.4s {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+        %tmp1 = load <4 x i32>, <4 x i32>* %A
+        %tmp3 = call <4 x i32> @llvm.aarch64.neon.sshl.v4i32(<4 x i32> %tmp1, <4 x i32> <i32 1, i32 1, i32 1, i32 1>)
+        ret <4 x i32> %tmp3
+}
+
+define <2 x i64> @neon.sshll2d_constant_shift(<2 x i32>* %A) nounwind {
+;CHECK-LABEL: neon.sshll2d_constant_shift
 ;CHECK: sshll.2d v0, {{v[0-9]+}}, #1
         %tmp1 = load <2 x i32>, <2 x i32>* %A
         %tmp2 = sext <2 x i32> %tmp1 to <2 x i64>
-        %tmp3 = shl <2 x i64> %tmp2, <i64 1, i64 1>
+        %tmp3 = call <2 x i64> @llvm.aarch64.neon.sshl.v2i64(<2 x i64> %tmp2, <2 x i64> <i64 1, i64 1>)
+        ret <2 x i64> %tmp3
+}
+
+; FIXME: should be constant folded.
+define <2 x i64> @neon.sshl2d_constant_fold() nounwind {
+;CHECK-LABEL: neon.sshl2d_constant_fold
+;CHECK: sshl.2d {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+        %tmp3 = call <2 x i64> @llvm.aarch64.neon.sshl.v2i64(<2 x i64> <i64 99, i64 1000>, <2 x i64> <i64 1, i64 1>)
+        ret <2 x i64> %tmp3
+}
+
+define <2 x i64> @neon.sshl2d_no_fold(<2 x i64>* %A) nounwind {
+;CHECK-LABEL: neon.sshl2d_no_fold
+;CHECK: sshl.2d {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+        %tmp2 = load <2 x i64>, <2 x i64>* %A
+        %tmp3 = call <2 x i64> @llvm.aarch64.neon.sshl.v2i64(<2 x i64> %tmp2, <2 x i64> <i64 2, i64 2>)
         ret <2 x i64> %tmp3
 }
 
 define <8 x i16> @sshll2_8h(<16 x i8>* %A) nounwind {
 ;CHECK-LABEL: sshll2_8h:
 ;CHECK: sshll.8h v0, {{v[0-9]+}}, #1
         %load1 = load <16 x i8>, <16 x i8>* %A
         %tmp1 = shufflevector <16 x i8> %load1, <16 x i8> undef, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
         %tmp2 = sext <8 x i8> %tmp1 to <8 x i16>
         %tmp3 = shl <8 x i16> %tmp2, <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>
         ret <8 x i16> %tmp3
 }
 
 define <4 x i32> @sshll2_4s(<8 x i16>* %A) nounwind {
 ;CHECK-LABEL: sshll2_4s:
 ;CHECK: sshll.4s v0, {{v[0-9]+}}, #1
         %load1 = load <8 x i16>, <8 x i16>* %A
;CHECK: sshll.2d v0, {{v[0-9]+}}, #1
         %load1 = load <4 x i32>, <4 x i32>* %A
         %tmp1 = shufflevector <4 x i32> %load1, <4 x i32> undef, <2 x i32> <i32 2, i32 3>
         %tmp2 = sext <2 x i32> %tmp1 to <2 x i64>
         %tmp3 = shl <2 x i64> %tmp2, <i64 1, i64 1>
         ret <2 x i64> %tmp3
 }
 
 define <8 x i8> @sqshli8b(<8 x i8>* %A) nounwind {
 ;CHECK-LABEL: sqshli8b:
 ;CHECK: sqshl.8b v0, {{v[0-9]+}}, #1
         %tmp1 = load <8 x i8>, <8 x i8>* %A
         %tmp3 = call <8 x i8> @llvm.aarch64.neon.sqshl.v8i8(<8 x i8> %tmp1, <8 x i8> <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>)
         ret <8 x i8> %tmp3
 }
 
 define <4 x i16> @sqshli4h(<4 x i16>* %A) nounwind {
 ;CHECK-LABEL: sqshli4h:
 ;CHECK: sqshl.4h v0, {{v[0-9]+}}, #1
         %tmp1 = load <4 x i16>, <4 x i16>* %A
         %tmp3 = call <4 x i16> @llvm.aarch64.neon.sqshl.v4i16(<4 x i16> %tmp1, <4 x i16> <i16 1, i16 1, i16 1, i16 1>)
         ret <4 x i16> %tmp3
 }
 
 define <2 x i32> @sqshli2s(<2 x i32>* %A) nounwind {
 ;CHECK-LABEL: sqshli2s:
 ;CHECK: sqshl.2s v0, {{v[0-9]+}}, #1
         %tmp1 = load <2 x i32>, <2 x i32>* %A
         %tmp3 = call <2 x i32> @llvm.aarch64.neon.sqshl.v2i32(<2 x i32> %tmp1, <2 x i32> <i32 1, i32 1>)
         ret <2 x i32> %tmp3
 }
 
 define <16 x i8> @sqshli16b(<16 x i8>* %A) nounwind {