1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-26 04:32:44 +01:00

[SystemZ] Replicate the load with most uses in buildVector()

Iterate over all elements and count the number of uses among them for each
used load. Then make sure to REPLICATE the load which has the most uses in
order to minimize the number of needed element insertions.

Review: Ulrich Weigand
https://reviews.llvm.org/D54322

llvm-svn: 346637
This commit is contained in:
Jonas Paulsson 2018-11-12 08:12:20 +00:00
parent eaae0059be
commit a41af2e2bb
2 changed files with 30 additions and 8 deletions

View File

@ -4490,18 +4490,21 @@ static SDValue buildVector(SelectionDAG &DAG, const SDLoc &DL, EVT VT,
// avoid a false dependency on any previous contents of the vector
// register.
// Use a VLREP if at least one element is a load.
unsigned LoadElIdx = UINT_MAX;
// Use a VLREP if at least one element is a load. Make sure to replicate
// the load with the most elements having its value.
std::map<const SDNode*, unsigned> UseCounts;
SDNode *LoadMaxUses = nullptr;
for (unsigned I = 0; I < NumElements; ++I)
if (Elems[I].getOpcode() == ISD::LOAD &&
cast<LoadSDNode>(Elems[I])->isUnindexed()) {
LoadElIdx = I;
break;
SDNode *Ld = Elems[I].getNode();
UseCounts[Ld]++;
if (LoadMaxUses == nullptr || UseCounts[LoadMaxUses] < UseCounts[Ld])
LoadMaxUses = Ld;
}
if (LoadElIdx != UINT_MAX) {
Result = DAG.getNode(SystemZISD::REPLICATE, DL, VT, Elems[LoadElIdx]);
Done[LoadElIdx] = true;
ReplicatedVal = Elems[LoadElIdx];
if (LoadMaxUses != nullptr) {
ReplicatedVal = SDValue(LoadMaxUses, 0);
Result = DAG.getNode(SystemZISD::REPLICATE, DL, VT, ReplicatedVal);
} else {
// Try to use VLVGP.
unsigned I1 = NumElements / 2 - 1;

View File

@ -0,0 +1,19 @@
; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z13 | FileCheck %s
; Test a vector which is built with elements from two loads replicates the
; load with most elements having its value.
; CHECK: vlef
; CHECK-NOT: vlvgf
define void @update(i32* %src1, i32* %src2, <4 x i32>* %dst) {
bb:
%tmp = load i32, i32* %src1
%tmp1 = load i32, i32* %src2
%tmp2 = insertelement <4 x i32> undef, i32 %tmp, i32 0
%tmp3 = insertelement <4 x i32> %tmp2, i32 %tmp1, i32 1
%tmp4 = insertelement <4 x i32> %tmp3, i32 %tmp1, i32 2
%tmp5 = insertelement <4 x i32> %tmp4, i32 %tmp1, i32 3
store <4 x i32> %tmp5, <4 x i32>* %dst
ret void
}