[llvm] [DAG] Fold nested add(add(reduce(a), b), add(reduce(c), d)) (PR #115150)

David Green via llvm-commits llvm-commits at lists.llvm.org
Thu Nov 21 05:07:14 PST 2024


https://github.com/davemgreen updated https://github.com/llvm/llvm-project/pull/115150

>From 27a6fee926d46d33c3cae1be75805b819cee1514 Mon Sep 17 00:00:00 2001
From: David Green <david.green at arm.com>
Date: Thu, 21 Nov 2024 13:01:43 +0000
Subject: [PATCH] [DAG] Fold nested add(add(reduce(a), b), add(reduce(c), d))

This patch reassociates add(add(vecreduce(a), b), add(vecreduce(c), d)) into
add(vecreduce(add(a, c)), add(b, d)), to combine the reductions into a single
node. This comes up after unrolling vectorized loops.

There is another small change to move reassociateReduction inside fadd outside
of a AllowNewConst block, as new constants will not be created and it should be
OK to perform the combine later after legalization.
---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp |  37 +++-
 llvm/test/CodeGen/AArch64/double_reduct.ll    | 149 ++++++---------
 llvm/test/CodeGen/Thumb2/mve-doublereduct.ll  | 176 +++++++-----------
 3 files changed, 154 insertions(+), 208 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 6059229cd6d9a4..0c06ab631e04a0 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -1329,6 +1329,38 @@ SDValue DAGCombiner::reassociateReduction(unsigned RedOpc, unsigned Opc,
                        DAG.getNode(Opc, DL, N0.getOperand(0).getValueType(),
                                    N0.getOperand(0), N1.getOperand(0)));
   }
+
+  // Reassociate op(op(vecreduce(a), b), op(vecreduce(c), d)) into
+  // op(vecreduce(op(a, c)), op(b, d)), to combine the reductions into a
+  // single node.
+  SDValue A, B, C, D, RedA, RedB;
+  if (sd_match(N0, m_OneUse(m_c_BinOp(
+                       Opc,
+                       m_AllOf(m_OneUse(m_UnaryOp(RedOpc, m_Value(A))),
+                               m_Value(RedA)),
+                       m_Value(B)))) &&
+      sd_match(N1, m_OneUse(m_c_BinOp(
+                       Opc,
+                       m_AllOf(m_OneUse(m_UnaryOp(RedOpc, m_Value(C))),
+                               m_Value(RedB)),
+                       m_Value(D)))) &&
+      !sd_match(B, m_UnaryOp(RedOpc, m_Value())) &&
+      !sd_match(D, m_UnaryOp(RedOpc, m_Value())) &&
+      A.getValueType() == C.getValueType() &&
+      hasOperation(Opc, A.getValueType()) &&
+      TLI.shouldReassociateReduction(RedOpc, VT)) {
+    if ((Opc == ISD::FADD || Opc == ISD::FMUL) &&
+        (!N0->getFlags().hasAllowReassociation() ||
+         !N1->getFlags().hasAllowReassociation() ||
+         !RedA->getFlags().hasAllowReassociation() ||
+         !RedB->getFlags().hasAllowReassociation()))
+      return SDValue();
+    SelectionDAG::FlagInserter FlagsInserter(DAG, Flags);
+    SDValue Op = DAG.getNode(Opc, DL, A.getValueType(), A, C);
+    SDValue Red = DAG.getNode(RedOpc, DL, VT, Op);
+    SDValue Op2 = DAG.getNode(Opc, DL, VT, B, D);
+    return DAG.getNode(Opc, DL, VT, Red, Op2);
+  }
   return SDValue();
 }
 
@@ -17107,12 +17139,15 @@ SDValue DAGCombiner::visitFADD(SDNode *N) {
                            DAG.getConstantFP(4.0, DL, VT));
       }
     }
+  } // enable-unsafe-fp-math && AllowNewConst
 
+  if (((Options.UnsafeFPMath && Options.NoSignedZerosFPMath) ||
+       (Flags.hasAllowReassociation() && Flags.hasNoSignedZeros()))) {
     // Fold fadd(vecreduce(x), vecreduce(y)) -> vecreduce(fadd(x, y))
     if (SDValue SD = reassociateReduction(ISD::VECREDUCE_FADD, ISD::FADD, DL,
                                           VT, N0, N1, Flags))
       return SD;
-  } // enable-unsafe-fp-math
+  }
 
   // FADD -> FMA combines:
   if (SDValue Fused = visitFADDForFMACombine<EmptyMatchContext>(N)) {
diff --git a/llvm/test/CodeGen/AArch64/double_reduct.ll b/llvm/test/CodeGen/AArch64/double_reduct.ll
index 0c356b1d98287f..b04e34a54af475 100644
--- a/llvm/test/CodeGen/AArch64/double_reduct.ll
+++ b/llvm/test/CodeGen/AArch64/double_reduct.ll
@@ -288,13 +288,11 @@ define i32 @smax_i32(<8 x i32> %a, <4 x i32> %b) {
 define float @nested_fadd_f32(<4 x float> %a, <4 x float> %b, float %c, float %d) {
 ; CHECK-LABEL: nested_fadd_f32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    faddp v1.4s, v1.4s, v1.4s
+; CHECK-NEXT:    fadd v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    fadd s2, s2, s3
 ; CHECK-NEXT:    faddp v0.4s, v0.4s, v0.4s
-; CHECK-NEXT:    faddp s1, v1.2s
 ; CHECK-NEXT:    faddp s0, v0.2s
-; CHECK-NEXT:    fadd s1, s1, s3
 ; CHECK-NEXT:    fadd s0, s0, s2
-; CHECK-NEXT:    fadd s0, s0, s1
 ; CHECK-NEXT:    ret
   %r1 = call fast float @llvm.vector.reduce.fadd.f32.v4f32(float -0.0, <4 x float> %a)
   %a1 = fadd fast float %r1, %c
@@ -332,15 +330,12 @@ define float @nested_fadd_f32_slow(<4 x float> %a, <4 x float> %b, float %c, flo
 define float @nested_mul_f32(<4 x float> %a, <4 x float> %b, float %c, float %d) {
 ; CHECK-LABEL: nested_mul_f32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ext v4.16b, v1.16b, v1.16b, #8
-; CHECK-NEXT:    ext v5.16b, v0.16b, v0.16b, #8
-; CHECK-NEXT:    fmul v1.2s, v1.2s, v4.2s
-; CHECK-NEXT:    fmul v0.2s, v0.2s, v5.2s
-; CHECK-NEXT:    fmul s1, s1, v1.s[1]
+; CHECK-NEXT:    fmul v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    fmul s2, s2, s3
+; CHECK-NEXT:    ext v1.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    fmul v0.2s, v0.2s, v1.2s
 ; CHECK-NEXT:    fmul s0, s0, v0.s[1]
-; CHECK-NEXT:    fmul s1, s1, s3
 ; CHECK-NEXT:    fmul s0, s0, s2
-; CHECK-NEXT:    fmul s0, s0, s1
 ; CHECK-NEXT:    ret
   %r1 = call fast float @llvm.vector.reduce.fmul.f32.v4f32(float 1.0, <4 x float> %a)
   %a1 = fmul fast float %r1, %c
@@ -353,12 +348,10 @@ define float @nested_mul_f32(<4 x float> %a, <4 x float> %b, float %c, float %d)
 define i32 @nested_add_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_add_i32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    addv s1, v1.4s
+; CHECK-NEXT:    add v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    add w8, w0, w1
 ; CHECK-NEXT:    addv s0, v0.4s
-; CHECK-NEXT:    fmov w8, s1
 ; CHECK-NEXT:    fmov w9, s0
-; CHECK-NEXT:    add w9, w9, w0
-; CHECK-NEXT:    add w8, w8, w1
 ; CHECK-NEXT:    add w0, w9, w8
 ; CHECK-NEXT:    ret
   %r1 = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %a)
@@ -372,12 +365,10 @@ define i32 @nested_add_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_add_c1_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_add_c1_i32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    addv s1, v1.4s
+; CHECK-NEXT:    add v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    add w8, w0, w1
 ; CHECK-NEXT:    addv s0, v0.4s
-; CHECK-NEXT:    fmov w8, s1
 ; CHECK-NEXT:    fmov w9, s0
-; CHECK-NEXT:    add w9, w0, w9
-; CHECK-NEXT:    add w8, w8, w1
 ; CHECK-NEXT:    add w0, w9, w8
 ; CHECK-NEXT:    ret
   %r1 = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %a)
@@ -391,12 +382,10 @@ define i32 @nested_add_c1_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_add_c2_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_add_c2_i32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    addv s1, v1.4s
+; CHECK-NEXT:    add v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    add w8, w0, w1
 ; CHECK-NEXT:    addv s0, v0.4s
-; CHECK-NEXT:    fmov w8, s1
 ; CHECK-NEXT:    fmov w9, s0
-; CHECK-NEXT:    add w9, w9, w0
-; CHECK-NEXT:    add w8, w1, w8
 ; CHECK-NEXT:    add w0, w9, w8
 ; CHECK-NEXT:    ret
   %r1 = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %a)
@@ -429,19 +418,14 @@ define i32 @nested_add_manyreduct_i32(<4 x i32> %a, <4 x i32> %b, <4 x i32> %c,
 define i32 @nested_mul_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_mul_i32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ext v3.16b, v0.16b, v0.16b, #8
-; CHECK-NEXT:    ext v2.16b, v1.16b, v1.16b, #8
-; CHECK-NEXT:    mul v0.2s, v0.2s, v3.2s
-; CHECK-NEXT:    mul v1.2s, v1.2s, v2.2s
-; CHECK-NEXT:    mov w8, v0.s[1]
+; CHECK-NEXT:    mul v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    mul w8, w0, w1
+; CHECK-NEXT:    ext v1.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    mul v0.2s, v0.2s, v1.2s
+; CHECK-NEXT:    mov w9, v0.s[1]
 ; CHECK-NEXT:    fmov w10, s0
-; CHECK-NEXT:    mov w9, v1.s[1]
-; CHECK-NEXT:    mul w8, w10, w8
-; CHECK-NEXT:    fmov w10, s1
 ; CHECK-NEXT:    mul w9, w10, w9
-; CHECK-NEXT:    mul w8, w8, w0
-; CHECK-NEXT:    mul w9, w9, w1
-; CHECK-NEXT:    mul w0, w8, w9
+; CHECK-NEXT:    mul w0, w9, w8
 ; CHECK-NEXT:    ret
   %r1 = call i32 @llvm.vector.reduce.mul.v4i32(<4 x i32> %a)
   %a1 = mul i32 %r1, %c
@@ -454,19 +438,14 @@ define i32 @nested_mul_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_and_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_and_i32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ext v2.16b, v1.16b, v1.16b, #8
-; CHECK-NEXT:    ext v3.16b, v0.16b, v0.16b, #8
-; CHECK-NEXT:    and v1.8b, v1.8b, v2.8b
-; CHECK-NEXT:    and v0.8b, v0.8b, v3.8b
-; CHECK-NEXT:    fmov x8, d1
+; CHECK-NEXT:    and v0.16b, v0.16b, v1.16b
+; CHECK-NEXT:    and w8, w0, w1
+; CHECK-NEXT:    ext v1.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    and v0.8b, v0.8b, v1.8b
 ; CHECK-NEXT:    fmov x9, d0
 ; CHECK-NEXT:    lsr x10, x9, #32
-; CHECK-NEXT:    lsr x11, x8, #32
-; CHECK-NEXT:    and w9, w9, w0
-; CHECK-NEXT:    and w8, w8, w1
-; CHECK-NEXT:    and w9, w9, w10
-; CHECK-NEXT:    and w8, w8, w11
-; CHECK-NEXT:    and w0, w9, w8
+; CHECK-NEXT:    and w8, w9, w8
+; CHECK-NEXT:    and w0, w8, w10
 ; CHECK-NEXT:    ret
   %r1 = call i32 @llvm.vector.reduce.and.v4i32(<4 x i32> %a)
   %a1 = and i32 %r1, %c
@@ -479,19 +458,14 @@ define i32 @nested_and_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_or_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_or_i32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ext v2.16b, v1.16b, v1.16b, #8
-; CHECK-NEXT:    ext v3.16b, v0.16b, v0.16b, #8
-; CHECK-NEXT:    orr v1.8b, v1.8b, v2.8b
-; CHECK-NEXT:    orr v0.8b, v0.8b, v3.8b
-; CHECK-NEXT:    fmov x8, d1
+; CHECK-NEXT:    orr v0.16b, v0.16b, v1.16b
+; CHECK-NEXT:    orr w8, w0, w1
+; CHECK-NEXT:    ext v1.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    orr v0.8b, v0.8b, v1.8b
 ; CHECK-NEXT:    fmov x9, d0
 ; CHECK-NEXT:    lsr x10, x9, #32
-; CHECK-NEXT:    lsr x11, x8, #32
-; CHECK-NEXT:    orr w9, w9, w0
-; CHECK-NEXT:    orr w8, w8, w1
-; CHECK-NEXT:    orr w9, w9, w10
-; CHECK-NEXT:    orr w8, w8, w11
-; CHECK-NEXT:    orr w0, w9, w8
+; CHECK-NEXT:    orr w8, w9, w8
+; CHECK-NEXT:    orr w0, w8, w10
 ; CHECK-NEXT:    ret
   %r1 = call i32 @llvm.vector.reduce.or.v4i32(<4 x i32> %a)
   %a1 = or i32 %r1, %c
@@ -504,19 +478,14 @@ define i32 @nested_or_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_xor_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_xor_i32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ext v2.16b, v1.16b, v1.16b, #8
-; CHECK-NEXT:    ext v3.16b, v0.16b, v0.16b, #8
-; CHECK-NEXT:    eor v1.8b, v1.8b, v2.8b
-; CHECK-NEXT:    eor v0.8b, v0.8b, v3.8b
-; CHECK-NEXT:    fmov x8, d1
+; CHECK-NEXT:    eor v0.16b, v0.16b, v1.16b
+; CHECK-NEXT:    eor w8, w0, w1
+; CHECK-NEXT:    ext v1.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    eor v0.8b, v0.8b, v1.8b
 ; CHECK-NEXT:    fmov x9, d0
 ; CHECK-NEXT:    lsr x10, x9, #32
-; CHECK-NEXT:    lsr x11, x8, #32
-; CHECK-NEXT:    eor w9, w9, w0
-; CHECK-NEXT:    eor w8, w8, w1
-; CHECK-NEXT:    eor w9, w9, w10
-; CHECK-NEXT:    eor w8, w8, w11
-; CHECK-NEXT:    eor w0, w9, w8
+; CHECK-NEXT:    eor w8, w9, w8
+; CHECK-NEXT:    eor w0, w8, w10
 ; CHECK-NEXT:    ret
   %r1 = call i32 @llvm.vector.reduce.xor.v4i32(<4 x i32> %a)
   %a1 = xor i32 %r1, %c
@@ -529,14 +498,11 @@ define i32 @nested_xor_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_smin_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_smin_i32:
 ; CHECK:       // %bb.0:
+; CHECK-NEXT:    smin v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    cmp w0, w1
+; CHECK-NEXT:    csel w8, w0, w1, lt
 ; CHECK-NEXT:    sminv s0, v0.4s
-; CHECK-NEXT:    sminv s1, v1.4s
 ; CHECK-NEXT:    fmov w9, s0
-; CHECK-NEXT:    fmov w8, s1
-; CHECK-NEXT:    cmp w9, w0
-; CHECK-NEXT:    csel w9, w9, w0, lt
-; CHECK-NEXT:    cmp w8, w1
-; CHECK-NEXT:    csel w8, w8, w1, lt
 ; CHECK-NEXT:    cmp w9, w8
 ; CHECK-NEXT:    csel w0, w9, w8, lt
 ; CHECK-NEXT:    ret
@@ -551,14 +517,11 @@ define i32 @nested_smin_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_smax_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_smax_i32:
 ; CHECK:       // %bb.0:
+; CHECK-NEXT:    smax v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    cmp w0, w1
+; CHECK-NEXT:    csel w8, w0, w1, gt
 ; CHECK-NEXT:    smaxv s0, v0.4s
-; CHECK-NEXT:    smaxv s1, v1.4s
 ; CHECK-NEXT:    fmov w9, s0
-; CHECK-NEXT:    fmov w8, s1
-; CHECK-NEXT:    cmp w9, w0
-; CHECK-NEXT:    csel w9, w9, w0, gt
-; CHECK-NEXT:    cmp w8, w1
-; CHECK-NEXT:    csel w8, w8, w1, gt
 ; CHECK-NEXT:    cmp w9, w8
 ; CHECK-NEXT:    csel w0, w9, w8, gt
 ; CHECK-NEXT:    ret
@@ -573,14 +536,11 @@ define i32 @nested_smax_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_umin_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_umin_i32:
 ; CHECK:       // %bb.0:
+; CHECK-NEXT:    umin v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    cmp w0, w1
+; CHECK-NEXT:    csel w8, w0, w1, lo
 ; CHECK-NEXT:    uminv s0, v0.4s
-; CHECK-NEXT:    uminv s1, v1.4s
 ; CHECK-NEXT:    fmov w9, s0
-; CHECK-NEXT:    fmov w8, s1
-; CHECK-NEXT:    cmp w9, w0
-; CHECK-NEXT:    csel w9, w9, w0, lo
-; CHECK-NEXT:    cmp w8, w1
-; CHECK-NEXT:    csel w8, w8, w1, lo
 ; CHECK-NEXT:    cmp w9, w8
 ; CHECK-NEXT:    csel w0, w9, w8, lo
 ; CHECK-NEXT:    ret
@@ -595,14 +555,11 @@ define i32 @nested_umin_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_umax_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_umax_i32:
 ; CHECK:       // %bb.0:
+; CHECK-NEXT:    umax v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    cmp w0, w1
+; CHECK-NEXT:    csel w8, w0, w1, hi
 ; CHECK-NEXT:    umaxv s0, v0.4s
-; CHECK-NEXT:    umaxv s1, v1.4s
 ; CHECK-NEXT:    fmov w9, s0
-; CHECK-NEXT:    fmov w8, s1
-; CHECK-NEXT:    cmp w9, w0
-; CHECK-NEXT:    csel w9, w9, w0, hi
-; CHECK-NEXT:    cmp w8, w1
-; CHECK-NEXT:    csel w8, w8, w1, hi
 ; CHECK-NEXT:    cmp w9, w8
 ; CHECK-NEXT:    csel w0, w9, w8, hi
 ; CHECK-NEXT:    ret
@@ -617,11 +574,10 @@ define i32 @nested_umax_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define float @nested_fmin_float(<4 x float> %a, <4 x float> %b, float %c, float %d) {
 ; CHECK-LABEL: nested_fmin_float:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    fminnmv s1, v1.4s
+; CHECK-NEXT:    fminnm v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    fminnm s2, s2, s3
 ; CHECK-NEXT:    fminnmv s0, v0.4s
-; CHECK-NEXT:    fminnm s1, s1, s3
 ; CHECK-NEXT:    fminnm s0, s0, s2
-; CHECK-NEXT:    fminnm s0, s0, s1
 ; CHECK-NEXT:    ret
   %r1 = call float @llvm.vector.reduce.fmin.v4f32(<4 x float> %a)
   %a1 = call float @llvm.minnum.f32(float %r1, float %c)
@@ -634,11 +590,10 @@ define float @nested_fmin_float(<4 x float> %a, <4 x float> %b, float %c, float
 define float @nested_fmax_float(<4 x float> %a, <4 x float> %b, float %c, float %d) {
 ; CHECK-LABEL: nested_fmax_float:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    fmaxnmv s1, v1.4s
+; CHECK-NEXT:    fmaxnm v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    fmaxnm s2, s2, s3
 ; CHECK-NEXT:    fmaxnmv s0, v0.4s
-; CHECK-NEXT:    fmaxnm s1, s1, s3
 ; CHECK-NEXT:    fmaxnm s0, s0, s2
-; CHECK-NEXT:    fmaxnm s0, s0, s1
 ; CHECK-NEXT:    ret
   %r1 = call float @llvm.vector.reduce.fmax.v4f32(<4 x float> %a)
   %a1 = call float @llvm.maxnum.f32(float %r1, float %c)
diff --git a/llvm/test/CodeGen/Thumb2/mve-doublereduct.ll b/llvm/test/CodeGen/Thumb2/mve-doublereduct.ll
index 67723e8aa41ad7..1136246f6b14dd 100644
--- a/llvm/test/CodeGen/Thumb2/mve-doublereduct.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-doublereduct.ll
@@ -244,15 +244,12 @@ define i32 @smax_i32(<8 x i32> %a, <4 x i32> %b) {
 define float @nested_add_f32(<4 x float> %a, <4 x float> %b, float %c, float %d) {
 ; CHECK-LABEL: nested_add_f32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    vadd.f32 s6, s6, s7
-; CHECK-NEXT:    vadd.f32 s4, s4, s5
+; CHECK-NEXT:    vadd.f32 q0, q0, q1
+; CHECK-NEXT:    vadd.f32 s4, s8, s9
 ; CHECK-NEXT:    vadd.f32 s2, s2, s3
 ; CHECK-NEXT:    vadd.f32 s0, s0, s1
-; CHECK-NEXT:    vadd.f32 s4, s4, s6
-; CHECK-NEXT:    vadd.f32 s0, s0, s2
-; CHECK-NEXT:    vadd.f32 s2, s4, s9
-; CHECK-NEXT:    vadd.f32 s0, s0, s8
 ; CHECK-NEXT:    vadd.f32 s0, s0, s2
+; CHECK-NEXT:    vadd.f32 s0, s0, s4
 ; CHECK-NEXT:    bx lr
   %r1 = call fast float @llvm.vector.reduce.fadd.f32.v4f32(float -0.0, <4 x float> %a)
   %a1 = fadd fast float %r1, %c
@@ -265,15 +262,12 @@ define float @nested_add_f32(<4 x float> %a, <4 x float> %b, float %c, float %d)
 define float @nested_mul_f32(<4 x float> %a, <4 x float> %b, float %c, float %d) {
 ; CHECK-LABEL: nested_mul_f32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    vmul.f32 s6, s6, s7
-; CHECK-NEXT:    vmul.f32 s4, s4, s5
+; CHECK-NEXT:    vmul.f32 q0, q0, q1
+; CHECK-NEXT:    vmul.f32 s4, s8, s9
 ; CHECK-NEXT:    vmul.f32 s2, s2, s3
 ; CHECK-NEXT:    vmul.f32 s0, s0, s1
-; CHECK-NEXT:    vmul.f32 s4, s4, s6
-; CHECK-NEXT:    vmul.f32 s0, s0, s2
-; CHECK-NEXT:    vmul.f32 s2, s4, s9
-; CHECK-NEXT:    vmul.f32 s0, s0, s8
 ; CHECK-NEXT:    vmul.f32 s0, s0, s2
+; CHECK-NEXT:    vmul.f32 s0, s0, s4
 ; CHECK-NEXT:    bx lr
   %r1 = call fast float @llvm.vector.reduce.fmul.f32.v4f32(float 1.0, <4 x float> %a)
   %a1 = fmul fast float %r1, %c
@@ -301,22 +295,17 @@ define i32 @nested_add_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_mul_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_mul_i32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    .save {r4, r5, r6, r7, r8, lr}
-; CHECK-NEXT:    push.w {r4, r5, r6, r7, r8, lr}
-; CHECK-NEXT:    vmov r8, r3, d2
-; CHECK-NEXT:    vmov r4, r5, d1
-; CHECK-NEXT:    vmov r6, r7, d0
-; CHECK-NEXT:    vmov r12, lr, d3
-; CHECK-NEXT:    mul r3, r8, r3
-; CHECK-NEXT:    muls r5, r4, r5
-; CHECK-NEXT:    mul r2, r12, lr
-; CHECK-NEXT:    muls r7, r6, r7
+; CHECK-NEXT:    .save {r7, lr}
+; CHECK-NEXT:    push {r7, lr}
+; CHECK-NEXT:    vmul.i32 q0, q0, q1
+; CHECK-NEXT:    muls r0, r1, r0
+; CHECK-NEXT:    vmov r12, lr, d1
+; CHECK-NEXT:    vmov r2, r3, d0
+; CHECK-NEXT:    mul r12, r12, lr
 ; CHECK-NEXT:    muls r2, r3, r2
-; CHECK-NEXT:    mul r3, r7, r5
-; CHECK-NEXT:    muls r1, r2, r1
-; CHECK-NEXT:    muls r0, r3, r0
+; CHECK-NEXT:    mul r1, r2, r12
 ; CHECK-NEXT:    muls r0, r1, r0
-; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, pc}
+; CHECK-NEXT:    pop {r7, pc}
   %r1 = call i32 @llvm.vector.reduce.mul.v4i32(<4 x i32> %a)
   %a1 = mul i32 %r1, %c
   %r2 = call i32 @llvm.vector.reduce.mul.v4i32(<4 x i32> %b)
@@ -328,22 +317,17 @@ define i32 @nested_mul_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_and_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_and_i32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    .save {r4, r5, r6, r7, r8, lr}
-; CHECK-NEXT:    push.w {r4, r5, r6, r7, r8, lr}
-; CHECK-NEXT:    vmov r2, r3, d2
-; CHECK-NEXT:    vmov r12, lr, d3
-; CHECK-NEXT:    vmov r8, r5, d1
-; CHECK-NEXT:    vmov r6, r7, d0
-; CHECK-NEXT:    ands r2, r3
-; CHECK-NEXT:    and.w r4, r12, lr
-; CHECK-NEXT:    ands r2, r4
-; CHECK-NEXT:    ands r1, r2
-; CHECK-NEXT:    and.w r2, r8, r5
-; CHECK-NEXT:    and.w r3, r6, r7
+; CHECK-NEXT:    .save {r7, lr}
+; CHECK-NEXT:    push {r7, lr}
+; CHECK-NEXT:    vand q0, q0, q1
+; CHECK-NEXT:    ands r0, r1
+; CHECK-NEXT:    vmov r12, lr, d1
+; CHECK-NEXT:    vmov r2, r3, d0
+; CHECK-NEXT:    and.w r12, r12, lr
 ; CHECK-NEXT:    ands r2, r3
+; CHECK-NEXT:    and.w r2, r2, r12
 ; CHECK-NEXT:    ands r0, r2
-; CHECK-NEXT:    ands r0, r1
-; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, pc}
+; CHECK-NEXT:    pop {r7, pc}
   %r1 = call i32 @llvm.vector.reduce.and.v4i32(<4 x i32> %a)
   %a1 = and i32 %r1, %c
   %r2 = call i32 @llvm.vector.reduce.and.v4i32(<4 x i32> %b)
@@ -355,22 +339,17 @@ define i32 @nested_and_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_or_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_or_i32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    .save {r4, r5, r6, r7, r8, lr}
-; CHECK-NEXT:    push.w {r4, r5, r6, r7, r8, lr}
-; CHECK-NEXT:    vmov r2, r3, d2
-; CHECK-NEXT:    vmov r12, lr, d3
-; CHECK-NEXT:    vmov r8, r5, d1
-; CHECK-NEXT:    vmov r6, r7, d0
-; CHECK-NEXT:    orrs r2, r3
-; CHECK-NEXT:    orr.w r4, r12, lr
-; CHECK-NEXT:    orrs r2, r4
-; CHECK-NEXT:    orrs r1, r2
-; CHECK-NEXT:    orr.w r2, r8, r5
-; CHECK-NEXT:    orr.w r3, r6, r7
+; CHECK-NEXT:    .save {r7, lr}
+; CHECK-NEXT:    push {r7, lr}
+; CHECK-NEXT:    vorr q0, q0, q1
+; CHECK-NEXT:    orrs r0, r1
+; CHECK-NEXT:    vmov r12, lr, d1
+; CHECK-NEXT:    vmov r2, r3, d0
+; CHECK-NEXT:    orr.w r12, r12, lr
 ; CHECK-NEXT:    orrs r2, r3
+; CHECK-NEXT:    orr.w r2, r2, r12
 ; CHECK-NEXT:    orrs r0, r2
-; CHECK-NEXT:    orrs r0, r1
-; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, pc}
+; CHECK-NEXT:    pop {r7, pc}
   %r1 = call i32 @llvm.vector.reduce.or.v4i32(<4 x i32> %a)
   %a1 = or i32 %r1, %c
   %r2 = call i32 @llvm.vector.reduce.or.v4i32(<4 x i32> %b)
@@ -382,22 +361,17 @@ define i32 @nested_or_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_xor_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_xor_i32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    .save {r4, r5, r6, r7, r8, lr}
-; CHECK-NEXT:    push.w {r4, r5, r6, r7, r8, lr}
-; CHECK-NEXT:    vmov r2, r3, d2
-; CHECK-NEXT:    vmov r12, lr, d3
-; CHECK-NEXT:    vmov r8, r5, d1
-; CHECK-NEXT:    vmov r6, r7, d0
-; CHECK-NEXT:    eors r2, r3
-; CHECK-NEXT:    eor.w r4, r12, lr
-; CHECK-NEXT:    eors r2, r4
-; CHECK-NEXT:    eors r1, r2
-; CHECK-NEXT:    eor.w r2, r8, r5
-; CHECK-NEXT:    eor.w r3, r6, r7
+; CHECK-NEXT:    .save {r7, lr}
+; CHECK-NEXT:    push {r7, lr}
+; CHECK-NEXT:    veor q0, q0, q1
+; CHECK-NEXT:    eors r0, r1
+; CHECK-NEXT:    vmov r12, lr, d1
+; CHECK-NEXT:    vmov r2, r3, d0
+; CHECK-NEXT:    eor.w r12, r12, lr
 ; CHECK-NEXT:    eors r2, r3
+; CHECK-NEXT:    eor.w r2, r2, r12
 ; CHECK-NEXT:    eors r0, r2
-; CHECK-NEXT:    eors r0, r1
-; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, pc}
+; CHECK-NEXT:    pop {r7, pc}
   %r1 = call i32 @llvm.vector.reduce.xor.v4i32(<4 x i32> %a)
   %a1 = xor i32 %r1, %c
   %r2 = call i32 @llvm.vector.reduce.xor.v4i32(<4 x i32> %b)
@@ -409,16 +383,13 @@ define i32 @nested_xor_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_smin_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_smin_i32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    mvn r3, #-2147483648
-; CHECK-NEXT:    mvn r2, #-2147483648
-; CHECK-NEXT:    vminv.s32 r3, q0
-; CHECK-NEXT:    vminv.s32 r2, q1
-; CHECK-NEXT:    cmp r3, r0
-; CHECK-NEXT:    csel r0, r3, r0, lt
-; CHECK-NEXT:    cmp r2, r1
-; CHECK-NEXT:    csel r1, r2, r1, lt
 ; CHECK-NEXT:    cmp r0, r1
+; CHECK-NEXT:    mvn r2, #-2147483648
+; CHECK-NEXT:    vmin.s32 q0, q0, q1
 ; CHECK-NEXT:    csel r0, r0, r1, lt
+; CHECK-NEXT:    vminv.s32 r2, q0
+; CHECK-NEXT:    cmp r2, r0
+; CHECK-NEXT:    csel r0, r2, r0, lt
 ; CHECK-NEXT:    bx lr
   %r1 = call i32 @llvm.vector.reduce.smin.v4i32(<4 x i32> %a)
   %a1 = call i32 @llvm.smin.i32(i32 %r1, i32 %c)
@@ -431,16 +402,13 @@ define i32 @nested_smin_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_smax_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_smax_i32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    mov.w r3, #-2147483648
-; CHECK-NEXT:    mov.w r2, #-2147483648
-; CHECK-NEXT:    vmaxv.s32 r3, q0
-; CHECK-NEXT:    vmaxv.s32 r2, q1
-; CHECK-NEXT:    cmp r3, r0
-; CHECK-NEXT:    csel r0, r3, r0, gt
-; CHECK-NEXT:    cmp r2, r1
-; CHECK-NEXT:    csel r1, r2, r1, gt
 ; CHECK-NEXT:    cmp r0, r1
+; CHECK-NEXT:    mov.w r2, #-2147483648
+; CHECK-NEXT:    vmax.s32 q0, q0, q1
 ; CHECK-NEXT:    csel r0, r0, r1, gt
+; CHECK-NEXT:    vmaxv.s32 r2, q0
+; CHECK-NEXT:    cmp r2, r0
+; CHECK-NEXT:    csel r0, r2, r0, gt
 ; CHECK-NEXT:    bx lr
   %r1 = call i32 @llvm.vector.reduce.smax.v4i32(<4 x i32> %a)
   %a1 = call i32 @llvm.smax.i32(i32 %r1, i32 %c)
@@ -453,16 +421,13 @@ define i32 @nested_smax_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_umin_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_umin_i32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    mov.w r3, #-1
-; CHECK-NEXT:    mov.w r2, #-1
-; CHECK-NEXT:    vminv.u32 r3, q0
-; CHECK-NEXT:    vminv.u32 r2, q1
-; CHECK-NEXT:    cmp r3, r0
-; CHECK-NEXT:    csel r0, r3, r0, lo
-; CHECK-NEXT:    cmp r2, r1
-; CHECK-NEXT:    csel r1, r2, r1, lo
 ; CHECK-NEXT:    cmp r0, r1
+; CHECK-NEXT:    mov.w r2, #-1
+; CHECK-NEXT:    vmin.u32 q0, q0, q1
 ; CHECK-NEXT:    csel r0, r0, r1, lo
+; CHECK-NEXT:    vminv.u32 r2, q0
+; CHECK-NEXT:    cmp r2, r0
+; CHECK-NEXT:    csel r0, r2, r0, lo
 ; CHECK-NEXT:    bx lr
   %r1 = call i32 @llvm.vector.reduce.umin.v4i32(<4 x i32> %a)
   %a1 = call i32 @llvm.umin.i32(i32 %r1, i32 %c)
@@ -475,16 +440,13 @@ define i32 @nested_umin_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define i32 @nested_umax_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 ; CHECK-LABEL: nested_umax_i32:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    movs r3, #0
-; CHECK-NEXT:    movs r2, #0
-; CHECK-NEXT:    vmaxv.u32 r3, q0
-; CHECK-NEXT:    vmaxv.u32 r2, q1
-; CHECK-NEXT:    cmp r3, r0
-; CHECK-NEXT:    csel r0, r3, r0, hi
-; CHECK-NEXT:    cmp r2, r1
-; CHECK-NEXT:    csel r1, r2, r1, hi
 ; CHECK-NEXT:    cmp r0, r1
+; CHECK-NEXT:    mov.w r2, #0
+; CHECK-NEXT:    vmax.u32 q0, q0, q1
 ; CHECK-NEXT:    csel r0, r0, r1, hi
+; CHECK-NEXT:    vmaxv.u32 r2, q0
+; CHECK-NEXT:    cmp r2, r0
+; CHECK-NEXT:    csel r0, r2, r0, hi
 ; CHECK-NEXT:    bx lr
   %r1 = call i32 @llvm.vector.reduce.umax.v4i32(<4 x i32> %a)
   %a1 = call i32 @llvm.umax.i32(i32 %r1, i32 %c)
@@ -497,14 +459,11 @@ define i32 @nested_umax_i32(<4 x i32> %a, <4 x i32> %b, i32 %c, i32 %d) {
 define float @nested_fmin_float(<4 x float> %a, <4 x float> %b, float %c, float %d) {
 ; CHECK-LABEL: nested_fmin_float:
 ; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vminnm.f32 q0, q0, q1
 ; CHECK-NEXT:    vminnm.f32 s2, s2, s3
 ; CHECK-NEXT:    vminnm.f32 s0, s0, s1
 ; CHECK-NEXT:    vminnm.f32 s0, s0, s2
-; CHECK-NEXT:    vminnm.f32 s2, s6, s7
-; CHECK-NEXT:    vminnm.f32 s4, s4, s5
-; CHECK-NEXT:    vminnm.f32 s0, s0, s8
-; CHECK-NEXT:    vminnm.f32 s2, s4, s2
-; CHECK-NEXT:    vminnm.f32 s2, s2, s9
+; CHECK-NEXT:    vminnm.f32 s2, s8, s9
 ; CHECK-NEXT:    vminnm.f32 s0, s0, s2
 ; CHECK-NEXT:    bx lr
   %r1 = call float @llvm.vector.reduce.fmin.v4f32(<4 x float> %a)
@@ -518,14 +477,11 @@ define float @nested_fmin_float(<4 x float> %a, <4 x float> %b, float %c, float
 define float @nested_fmax_float(<4 x float> %a, <4 x float> %b, float %c, float %d) {
 ; CHECK-LABEL: nested_fmax_float:
 ; CHECK:       @ %bb.0:
+; CHECK-NEXT:    vmaxnm.f32 q0, q0, q1
 ; CHECK-NEXT:    vmaxnm.f32 s2, s2, s3
 ; CHECK-NEXT:    vmaxnm.f32 s0, s0, s1
 ; CHECK-NEXT:    vmaxnm.f32 s0, s0, s2
-; CHECK-NEXT:    vmaxnm.f32 s2, s6, s7
-; CHECK-NEXT:    vmaxnm.f32 s4, s4, s5
-; CHECK-NEXT:    vmaxnm.f32 s0, s0, s8
-; CHECK-NEXT:    vmaxnm.f32 s2, s4, s2
-; CHECK-NEXT:    vmaxnm.f32 s2, s2, s9
+; CHECK-NEXT:    vmaxnm.f32 s2, s8, s9
 ; CHECK-NEXT:    vmaxnm.f32 s0, s0, s2
 ; CHECK-NEXT:    bx lr
   %r1 = call float @llvm.vector.reduce.fmax.v4f32(<4 x float> %a)



More information about the llvm-commits mailing list