6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2025-02-07 14:33:15 +00:00
Code Issues Projects Releases Wiki Activity

Fix a bug with insertelement on SPU.

Browse Source 
The previous algorithm in LowerVECTOR_SHUFFLE 
didn't check all requirements for "monotonic" shuffles.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@111361 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Kalle Raiskila 2010-08-18 10:20:29 +00:00
parent 86a791284a
commit ca9460f5a0
2 changed files with 26 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
lib/Target/CellSPU/SPUISelLowering.cpp
View File

@ -1755,11 +1755,12 @@ static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
// If we have a single element being moved from V1 to V2, this can be handled
// using the C*[DX] compute mask instructions, but the vector elements have
// to be monotonically increasing with one exception element.
// to be monotonically increasing with one exception element, and the source
// slot of the element to move must be the same as the destination.
EVT VecVT = V1.getValueType();
EVT EltVT = VecVT.getVectorElementType();
unsigned EltsFromV2 = 0;
unsigned V2Elt = 0;
unsigned V2EltOffset = 0;
unsigned V2EltIdx0 = 0;
unsigned CurrElt = 0;
unsigned MaxElts = VecVT.getVectorNumElements();
@ -1792,9 +1793,13 @@ static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
if (monotonic) {
if (SrcElt >= V2EltIdx0) {
if (1 >= (++EltsFromV2)) {
V2Elt = (V2EltIdx0 - SrcElt) << 2;
}
// TODO: optimize for the monotonic case when several consecutive
// elements are taken form V2. Do we ever get such a case?
if (EltsFromV2 == 0 && CurrElt == (SrcElt - V2EltIdx0))
V2EltOffset = (SrcElt - V2EltIdx0) * (EltVT.getSizeInBits()/8);
else
monotonic = false;
++EltsFromV2;
} else if (CurrElt != SrcElt) {
monotonic = false;
}
@ -1830,7 +1835,7 @@ static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
// R1 ($sp) is used here only as it is guaranteed to have last bits zero
SDValue Pointer = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
DAG.getRegister(SPU::R1, PtrVT),
DAG.getConstant(V2Elt, MVT::i32));
DAG.getConstant(V2EltOffset, MVT::i32));
SDValue ShufMaskOp = DAG.getNode(SPUISD::SHUFFLE_MASK, dl,
maskVT, Pointer);

15
test/CodeGen/CellSPU/shuffles.ll
View File

@ -16,3 +16,18 @@ define <4 x float> @splat(float %param1) {
ret <4 x float> %val
}
define void @test_insert( <2 x float>* %ptr, float %val1, float %val2 ) {
%sl2_17_tmp1 = insertelement <2 x float> zeroinitializer, float %val1, i32 0
;CHECK: lqa $6,
;CHECK: shufb $4, $4, $5, $6
%sl2_17 = insertelement <2 x float> %sl2_17_tmp1, float %val2, i32 1
;CHECK: cdd $5, 0($3)
;CHECK: lqd $6, 0($3)
;CHECK: shufb $4, $4, $6, $5
;CHECK: stqd $4, 0($3)
;CHECK: bi $lr
store <2 x float> %sl2_17, <2 x float>* %ptr
ret void
}