RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-26 04:32:44 +01:00
Code Issues Projects Releases Wiki Activity

[LegalizeTypes] Teach PromoteIntRes_BITCAST to better handle a bitcast with vector output type and a vector input type that needs to be widened

Browse Source 
Summary: Previously if we had a bitcast vector output type that needs promotion and a vector input type that needs widening we would just do a stack store and load to handle the conversion. We can do a little better if we can widen the bitcast to a legal vector type the same size as the widened input type. Then we can do the bitcast between this widened type and the widened input type. Afterwards we can extract_subvector back to the original output and any_extend that. Type legalization will then circle back and handle promotion of the extract_subvector and the any_extend will just be removed. This will avoid going through the stack and allows us to remove a custom version of this legalization from X86.

Reviewers: efriedma, RKSimon

Reviewed By: efriedma

Subscribers: javed.absar, llvm-commits

Differential Revision: https://reviews.llvm.org/D53229

llvm-svn: 345567
This commit is contained in:
Craig Topper 2018-10-30 03:27:15 +00:00
parent d7e457ed36
commit e3e2aba431
3 changed files with 27 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
View File

@ -310,6 +310,26 @@ SDValue DAGTypeLegalizer::PromoteIntRes_BITCAST(SDNode *N) {
// make us bitcast between two vectors which are legalized in different ways.
if (NOutVT.bitsEq(NInVT) && !NOutVT.isVector())
return DAG.getNode(ISD::BITCAST, dl, NOutVT, GetWidenedVector(InOp));
// If the output type is also a vector and widening it to the same size
// as the widened input type would be a legal type, we can widen the bitcast
// and handle the promotion after.
if (NOutVT.isVector()) {
unsigned WidenInSize = NInVT.getSizeInBits();
unsigned OutSize = OutVT.getSizeInBits();
if (WidenInSize % OutSize == 0) {
unsigned Scale = WidenInSize / OutSize;
EVT WideOutVT = EVT::getVectorVT(*DAG.getContext(),
OutVT.getVectorElementType(),
OutVT.getVectorNumElements() * Scale);
if (isTypeLegal(WideOutVT)) {
InOp = DAG.getBitcast(WideOutVT, GetWidenedVector(InOp));
MVT IdxTy = TLI.getVectorIdxTy(DAG.getDataLayout());
InOp = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OutVT, InOp,
DAG.getConstant(0, dl, IdxTy));
return DAG.getNode(ISD::ANY_EXTEND, dl, NOutVT, InOp);
}
}
}
}
return DAG.getNode(ISD::ANY_EXTEND, dl, NOutVT,

10
lib/Target/X86/X86ISelLowering.cpp
View File

@ -26338,7 +26338,7 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
return;
}
if ((SrcVT != MVT::f64 && SrcVT != MVT::v2f32) ||
if (SrcVT != MVT::f64 ||
(DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8) ||
getTypeAction(*DAG.getContext(), DstVT) == TypeWidenVector)
return;
@ -26347,13 +26347,7 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
EVT SVT = DstVT.getVectorElementType();
EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
SDValue Res;
if (SrcVT == MVT::f64)
Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
MVT::v2f64, N->getOperand(0));
else
Res = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32, N->getOperand(0),
DAG.getUNDEF(MVT::v2f32));
Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, N->getOperand(0));
Res = DAG.getBitcast(WiderVT, Res);
Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DstVT, Res,
DAG.getIntPtrConstant(0, dl));

14
test/CodeGen/AArch64/bitcast-promote-widen.ll
View File

@ -6,16 +6,12 @@
define <2 x i16> @bitcast_v2i16_v2f16(<2 x half> %x) {
; CHECK-LABEL: bitcast_v2i16_v2f16:
; CHECK: // %bb.0:
; CHECK-NEXT: sub sp, sp, #16 // =16
; CHECK-NEXT: .cfi_def_cfa_offset 16
; CHECK-NEXT: // kill: def $d0 killed $d0 def $q0
; CHECK-NEXT: str s0, [sp, #12]
; CHECK-NEXT: ldrh w8, [sp, #12]
; CHECK-NEXT: ldrh w9, [sp, #14]
; CHECK-NEXT: fmov s0, w8
; CHECK-NEXT: mov v0.s[1], w9
; CHECK-NEXT: // kill: def $d0 killed $d0 killed $q0
; CHECK-NEXT: add sp, sp, #16 // =16
; CHECK-NEXT: umov w8, v0.h[0]
; CHECK-NEXT: fmov s1, w8
; CHECK-NEXT: umov w8, v0.h[1]
; CHECK-NEXT: mov v1.s[1], w8
; CHECK-NEXT: mov v0.16b, v1.16b
; CHECK-NEXT: ret
%y = bitcast <2 x half> %x to <2 x i16>
ret <2 x i16> %y