RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-25 20:23:11 +01:00
Code Issues Projects Releases Wiki Activity

[Hexagon] Emit enough stores when aligning vector addresses

Browse Source
This commit is contained in:
Krzysztof Parzyszek 2020-12-15 18:41:49 -06:00
parent f211f158cb
commit a31ff2abea
2 changed files with 27 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
lib/Target/Hexagon/HexagonVectorCombine.cpp
View File

@ -820,7 +820,9 @@ auto AlignVectors::realignGroup(const MoveGroup &Move) const -> bool {
return Builder.CreateBitCast(Val, VecTy);
};
for (int i = -1; i != NumSectors; ++i) {
// Create an extra "undef" sector at the beginning and at the end.
// They will be used as the left/right filler in the vlalign step.
for (int i = -1; i != NumSectors + 1; ++i) {
ByteSpan Section = VSpan.section(i * ScLen, ScLen).normalize();
Value *AccumV = UndefValue::get(SecTy);
Value *AccumM = HVC.getNullValue(SecTy);
@ -838,14 +840,14 @@ auto AlignVectors::realignGroup(const MoveGroup &Move) const -> bool {
}
// vlalign
for (int j = 1; j != NumSectors + 1; ++j) {
for (int j = 1; j != NumSectors + 2; ++j) {
ASpanV[j - 1].Seg.Val = HVC.vlalignb(Builder, ASpanV[j - 1].Seg.Val,
ASpanV[j].Seg.Val, AlignVal);
ASpanM[j - 1].Seg.Val = HVC.vlalignb(Builder, ASpanM[j - 1].Seg.Val,
ASpanM[j].Seg.Val, AlignVal);
}
for (int i = 0; i != NumSectors; ++i) {
for (int i = 0; i != NumSectors + 1; ++i) {
Value *Ptr = createAdjustedPointer(Builder, AlignAddr, SecTy, i * ScLen);
Value *Val = ASpanV[i].Seg.Val;
Value *Mask = ASpanM[i].Seg.Val; // bytes

22
test/CodeGen/Hexagon/autohvx/vector-align-store.ll Normal file
View File

@ -0,0 +1,22 @@
; RUN: llc -march=hexagon < %s | FileCheck %s
; Make sure we generate 3 aligned stores.
; CHECK: vmem({{.*}}) =
; CHECK: vmem({{.*}}) =
; CHECK: vmem({{.*}}) =
; CHECK-NOT: vmem
define void @f0(i16* %a0, i32 %a11, <64 x i16> %a22, <64 x i16> %a3) #0 {
b0:
%v0 = add i32 %a11, 64
%v1 = getelementptr i16, i16* %a0, i32 %v0
%v2 = bitcast i16* %v1 to <64 x i16>*
store <64 x i16> %a22, <64 x i16>* %v2, align 2
%v33 = add i32 %a11, 128
%v44 = getelementptr i16, i16* %a0, i32 %v33
%v5 = bitcast i16* %v44 to <64 x i16>*
store <64 x i16> %a3, <64 x i16>* %v5, align 2
ret void
}
attributes #0 = { nounwind "target-cpu"="hexagonv66" "target-features"="+hvx,+hvx-length128b" }