[AArch64] Add combine for add(udot(0, x, y), z) -> udot(z, x, y).

Given a zero input for a udot, an add can be folded in to take the place
of the input, using thte addition that the instruction naturally
performs.

Differential Revision: https://reviews.llvm.org/D97188

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 2e8752f..3cc809e 100644 (file)
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -13217,6 +13217,29 @@ static SDValue performUADDVCombine(SDNode *N, SelectionDAG &DAG) {
                      DAG.getConstant(0, DL, MVT::i64));
 }
 
+// ADD(UDOT(zero, x, y), A) -->  UDOT(A, x, y)
+static SDValue performAddDotCombine(SDNode *N, SelectionDAG &DAG) {
+  EVT VT = N->getValueType(0);
+  if (N->getOpcode() != ISD::ADD)
+    return SDValue();
+
+  SDValue Dot = N->getOperand(0);
+  SDValue A = N->getOperand(1);
+  // Handle commutivity
+  auto isZeroDot = [](SDValue Dot) {
+    return (Dot.getOpcode() == AArch64ISD::UDOT ||
+            Dot.getOpcode() == AArch64ISD::SDOT) &&
+           ISD::isBuildVectorAllZeros(Dot.getOperand(0).getNode());
+  };
+  if (!isZeroDot(Dot))
+    std::swap(Dot, A);
+  if (!isZeroDot(Dot))
+    return SDValue();
+
+  return DAG.getNode(Dot.getOpcode(), SDLoc(N), VT, A, Dot.getOperand(1),
+                     Dot.getOperand(2));
+}
+
 // The basic add/sub long vector instructions have variants with "2" on the end
 // which act on the high-half of their inputs. They are normally matched by
 // patterns like:
@@ -13276,6 +13299,8 @@ static SDValue performAddSubCombine(SDNode *N,
   // Try to change sum of two reductions.
   if (SDValue Val = performUADDVCombine(N, DAG))
     return Val;
+  if (SDValue Val = performAddDotCombine(N, DAG))
+    return Val;
 
   return performAddSubLongCombine(N, DCI, DAG);
 }

diff --git a/llvm/test/CodeGen/AArch64/neon-dot-product.ll b/llvm/test/CodeGen/AArch64/neon-dot-product.ll
index 1d82f93..6c76d4e 100644 (file)
--- a/llvm/test/CodeGen/AArch64/neon-dot-product.ll
+++ b/llvm/test/CodeGen/AArch64/neon-dot-product.ll
@@ -55,9 +55,7 @@ entry:
 define <2 x i32> @test_vdot_u32_zero(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c) #0 {
 ; CHECK-LABEL: test_vdot_u32_zero:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    movi v3.2d, #0000000000000000
-; CHECK-NEXT:    udot v3.2s, v1.8b, v2.8b
-; CHECK-NEXT:    add v0.2s, v3.2s, v0.2s
+; CHECK-NEXT:    udot v0.2s, v1.8b, v2.8b
 ; CHECK-NEXT:    ret
 entry:
   %vdot1.i = call <2 x i32> @llvm.aarch64.neon.udot.v2i32.v8i8(<2 x i32> zeroinitializer, <8 x i8> %b, <8 x i8> %c) #2
@@ -68,9 +66,7 @@ entry:
 define <4 x i32> @test_vdotq_u32_zero(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c) #0 {
 ; CHECK-LABEL: test_vdotq_u32_zero:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    movi v3.2d, #0000000000000000
-; CHECK-NEXT:    udot v3.4s, v1.16b, v2.16b
-; CHECK-NEXT:    add v0.4s, v3.4s, v0.4s
+; CHECK-NEXT:    udot v0.4s, v1.16b, v2.16b
 ; CHECK-NEXT:    ret
 entry:
   %vdot1.i = call <4 x i32> @llvm.aarch64.neon.udot.v4i32.v16i8(<4 x i32> zeroinitializer, <16 x i8> %b, <16 x i8> %c) #2
@@ -81,9 +77,7 @@ entry:
 define <2 x i32> @test_vdot_s32_zero(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c) #0 {
 ; CHECK-LABEL: test_vdot_s32_zero:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    movi v3.2d, #0000000000000000
-; CHECK-NEXT:    sdot v3.2s, v1.8b, v2.8b
-; CHECK-NEXT:    add v0.2s, v3.2s, v0.2s
+; CHECK-NEXT:    sdot v0.2s, v1.8b, v2.8b
 ; CHECK-NEXT:    ret
 entry:
   %vdot1.i = call <2 x i32> @llvm.aarch64.neon.sdot.v2i32.v8i8(<2 x i32> zeroinitializer, <8 x i8> %b, <8 x i8> %c) #2
@@ -94,9 +88,7 @@ entry:
 define <4 x i32> @test_vdotq_s32_zero(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c) #0 {
 ; CHECK-LABEL: test_vdotq_s32_zero:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    movi v3.2d, #0000000000000000
-; CHECK-NEXT:    sdot v3.4s, v1.16b, v2.16b
-; CHECK-NEXT:    add v0.4s, v3.4s, v0.4s
+; CHECK-NEXT:    sdot v0.4s, v1.16b, v2.16b
 ; CHECK-NEXT:    ret
 entry:
   %vdot1.i = call <4 x i32> @llvm.aarch64.neon.sdot.v4i32.v16i8(<4 x i32> zeroinitializer, <16 x i8> %b, <16 x i8> %c) #2
@@ -161,6 +153,11 @@ entry:
 
 
 define <2 x i32> @test_vdot_lane_u32_zero(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c) {
+; CHECK-LABEL: test_vdot_lane_u32_zero:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    // kill: def $d2 killed $d2 def $q2
+; CHECK-NEXT:    udot v0.2s, v1.8b, v2.4b[1]
+; CHECK-NEXT:    ret
 entry:
   %.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>
@@ -171,6 +168,11 @@ entry:
 }
 
 define <4 x i32> @test_vdotq_lane_u32_zero(<4 x i32> %a, <16 x i8> %b, <8 x i8> %c) {
+; CHECK-LABEL: test_vdotq_lane_u32_zero:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    // kill: def $d2 killed $d2 def $q2
+; CHECK-NEXT:    udot v0.4s, v1.16b, v2.4b[1]
+; CHECK-NEXT:    ret
 entry:
   %.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>
@@ -183,9 +185,7 @@ entry:
 define <2 x i32> @test_vdot_laneq_u32_zero(<2 x i32> %a, <8 x i8> %b, <16 x i8> %c) {
 ; CHECK-LABEL: test_vdot_laneq_u32_zero:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    movi v3.2d, #0000000000000000
-; CHECK-NEXT:    udot v3.2s, v1.8b, v2.4b[1]
-; CHECK-NEXT:    add v0.2s, v3.2s, v0.2s
+; CHECK-NEXT:    udot v0.2s, v1.8b, v2.4b[1]
 ; CHECK-NEXT:    ret
 entry:
   %.cast = bitcast <16 x i8> %c to <4 x i32>
@@ -199,9 +199,7 @@ entry:
 define <4 x i32> @test_vdotq_laneq_u32_zero(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c) {
 ; CHECK-LABEL: test_vdotq_laneq_u32_zero:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    movi v3.2d, #0000000000000000
-; CHECK-NEXT:    udot v3.4s, v1.16b, v2.4b[1]
-; CHECK-NEXT:    add v0.4s, v3.4s, v0.4s
+; CHECK-NEXT:    udot v0.4s, v1.16b, v2.4b[1]
 ; CHECK-NEXT:    ret
 entry:
   %.cast = bitcast <16 x i8> %c to <4 x i32>
@@ -269,6 +267,11 @@ entry:
 
 
 define <2 x i32> @test_vdot_lane_s32_zero(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c) {
+; CHECK-LABEL: test_vdot_lane_s32_zero:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    // kill: def $d2 killed $d2 def $q2
+; CHECK-NEXT:    sdot v0.2s, v1.8b, v2.4b[1]
+; CHECK-NEXT:    ret
 entry:
   %.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>
@@ -279,6 +282,11 @@ entry:
 }
 
 define <4 x i32> @test_vdotq_lane_s32_zero(<4 x i32> %a, <16 x i8> %b, <8 x i8> %c) {
+; CHECK-LABEL: test_vdotq_lane_s32_zero:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    // kill: def $d2 killed $d2 def $q2
+; CHECK-NEXT:    sdot v0.4s, v1.16b, v2.4b[1]
+; CHECK-NEXT:    ret
 entry:
   %.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>
@@ -291,9 +299,7 @@ entry:
 define <2 x i32> @test_vdot_laneq_s32_zero(<2 x i32> %a, <8 x i8> %b, <16 x i8> %c) {
 ; CHECK-LABEL: test_vdot_laneq_s32_zero:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    movi v3.2d, #0000000000000000
-; CHECK-NEXT:    sdot v3.2s, v1.8b, v2.4b[1]
-; CHECK-NEXT:    add v0.2s, v3.2s, v0.2s
+; CHECK-NEXT:    sdot v0.2s, v1.8b, v2.4b[1]
 ; CHECK-NEXT:    ret
 entry:
   %.cast = bitcast <16 x i8> %c to <4 x i32>
@@ -307,9 +313,7 @@ entry:
 define <4 x i32> @test_vdotq_laneq_s32_zero(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c) {
 ; CHECK-LABEL: test_vdotq_laneq_s32_zero:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    movi v3.2d, #0000000000000000
-; CHECK-NEXT:    sdot v3.4s, v1.16b, v2.4b[1]
-; CHECK-NEXT:    add v0.4s, v3.4s, v0.4s
+; CHECK-NEXT:    sdot v0.4s, v1.16b, v2.4b[1]
 ; CHECK-NEXT:    ret
 entry:
   %.cast = bitcast <16 x i8> %c to <4 x i32>

diff --git a/llvm/test/CodeGen/AArch64/neon-dotreduce.ll b/llvm/test/CodeGen/AArch64/neon-dotreduce.ll
index 2b399f2..55bfd06 100644 (file)
--- a/llvm/test/CodeGen/AArch64/neon-dotreduce.ll
+++ b/llvm/test/CodeGen/AArch64/neon-dotreduce.ll
@@ -273,11 +273,9 @@ define i32 @test_udot_v16i8_double_nomla(<16 x i8> %a, <16 x i8> %b, <16 x i8> %
 ; CHECK:       // %bb.0: // %entry
 ; CHECK-NEXT:    movi v1.16b, #1
 ; CHECK-NEXT:    movi v3.2d, #0000000000000000
-; CHECK-NEXT:    movi v4.2d, #0000000000000000
-; CHECK-NEXT:    udot v4.4s, v1.16b, v0.16b
 ; CHECK-NEXT:    udot v3.4s, v1.16b, v2.16b
-; CHECK-NEXT:    add v0.4s, v4.4s, v3.4s
-; CHECK-NEXT:    addv s0, v0.4s
+; CHECK-NEXT:    udot v3.4s, v1.16b, v0.16b
+; CHECK-NEXT:    addv s0, v3.4s
 ; CHECK-NEXT:    fmov w0, s0
 ; CHECK-NEXT:    ret
 entry:
@@ -390,11 +388,9 @@ define i32 @test_sdot_v16i8_double_nomla(<16 x i8> %a, <16 x i8> %b, <16 x i8> %
 ; CHECK:       // %bb.0: // %entry
 ; CHECK-NEXT:    movi v1.16b, #1
 ; CHECK-NEXT:    movi v3.2d, #0000000000000000
-; CHECK-NEXT:    movi v4.2d, #0000000000000000
-; CHECK-NEXT:    sdot v4.4s, v1.16b, v0.16b
 ; CHECK-NEXT:    sdot v3.4s, v1.16b, v2.16b
-; CHECK-NEXT:    add v0.4s, v4.4s, v3.4s
-; CHECK-NEXT:    addv s0, v0.4s
+; CHECK-NEXT:    sdot v3.4s, v1.16b, v0.16b
+; CHECK-NEXT:    addv s0, v3.4s
 ; CHECK-NEXT:    fmov w0, s0
 ; CHECK-NEXT:    ret
 entry: