replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

DAGCombiner: Merge store/loads when we have extload/truncstores 

This is helps on architectures where i8,i16 are not legal but we have byte, and
short loads/stores. Allowing us to merge copies like the one below on ARM.

copy(char *a, char *b, int n) {
 do {
   int t0 = a[0];
   int t1 = a[1];
   b[0] = t0;
   b[1] = t1;

radar://13536387

llvm-svn: 178546
This commit is contained in:
Arnold Schwaighofer 2013-04-02 15:58:51 +00:00
parent ad36fe5c19
commit d6c6e868b2
2 changed files with 117 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
View File

@ -7980,6 +7980,14 @@ bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
if (TLI.isTypeLegal(StoreTy))
LastLegalType = i+1;
// Or check whether a truncstore is legal.
else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
TargetLowering::TypePromoteInteger) {
EVT LegalizedStoredValueTy =
TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
LastLegalType = i+1;
}
// Find a legal type for the vector store.
EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
@ -8163,6 +8171,17 @@ bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
if (TLI.isTypeLegal(StoreTy))
LastLegalIntegerType = i + 1;
// Or check whether a truncstore and extload is legal.
else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
TargetLowering::TypePromoteInteger) {
EVT LegalizedStoredValueTy =
TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) &&
TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) &&
TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy))
LastLegalIntegerType = i+1;
}
}
// Only use vector types if the vector type is larger than the integer type.

98
llvm/test/CodeGen/ARM/MergeConsecutiveStores.ll Normal file
View File

@ -0,0 +1,98 @@
; RUN: llc < %s -mtriple=armv7-apple-darwin | FileCheck %s
; Make sure that we merge the consecutive load/store sequence below and use a
; word (16 bit) instead of a byte copy.
; CHECK: MergeLoadStoreBaseIndexOffset
; CHECK: ldrh [[REG:r[0-9]+]], [{{.*}}]
; CHECK: strh [[REG]], [r1], #2
define void @MergeLoadStoreBaseIndexOffset(i32* %a, i8* %b, i8* %c, i32 %n) {
br label %1
; <label>:1
%.09 = phi i32 [ %n, %0 ], [ %11, %1 ]
%.08 = phi i8* [ %b, %0 ], [ %10, %1 ]
%.0 = phi i32* [ %a, %0 ], [ %2, %1 ]
%2 = getelementptr inbounds i32* %.0, i32 1
%3 = load i32* %.0, align 1
%4 = getelementptr inbounds i8* %c, i32 %3
%5 = load i8* %4, align 1
%6 = add i32 %3, 1
%7 = getelementptr inbounds i8* %c, i32 %6
%8 = load i8* %7, align 1
store i8 %5, i8* %.08, align 1
%9 = getelementptr inbounds i8* %.08, i32 1
store i8 %8, i8* %9, align 1
%10 = getelementptr inbounds i8* %.08, i32 2
%11 = add nsw i32 %.09, -1
%12 = icmp eq i32 %11, 0
br i1 %12, label %13, label %1
; <label>:13
ret void
}
; Make sure that we merge the consecutive load/store sequence below and use a
; word (16 bit) instead of a byte copy even if there are intermediate sign
; extensions.
; CHECK: MergeLoadStoreBaseIndexOffsetSext
; CHECK: ldrh [[REG:r[0-9]+]], [{{.*}}]
; CHECK: strh [[REG]], [r1], #2
define void @MergeLoadStoreBaseIndexOffsetSext(i8* %a, i8* %b, i8* %c, i32 %n) {
br label %1
; <label>:1
%.09 = phi i32 [ %n, %0 ], [ %12, %1 ]
%.08 = phi i8* [ %b, %0 ], [ %11, %1 ]
%.0 = phi i8* [ %a, %0 ], [ %2, %1 ]
%2 = getelementptr inbounds i8* %.0, i32 1
%3 = load i8* %.0, align 1
%4 = sext i8 %3 to i32
%5 = getelementptr inbounds i8* %c, i32 %4
%6 = load i8* %5, align 1
%7 = add i32 %4, 1
%8 = getelementptr inbounds i8* %c, i32 %7
%9 = load i8* %8, align 1
store i8 %6, i8* %.08, align 1
%10 = getelementptr inbounds i8* %.08, i32 1
store i8 %9, i8* %10, align 1
%11 = getelementptr inbounds i8* %.08, i32 2
%12 = add nsw i32 %.09, -1
%13 = icmp eq i32 %12, 0
br i1 %13, label %14, label %1
; <label>:14
ret void
}
; However, we can only merge ignore sign extensions when they are on all memory
; computations;
; CHECK: loadStoreBaseIndexOffsetSextNoSex
; CHECK-NOT: ldrh [[REG:r[0-9]+]], [{{.*}}]
; CHECK-NOT: strh [[REG]], [r1], #2
define void @loadStoreBaseIndexOffsetSextNoSex(i8* %a, i8* %b, i8* %c, i32 %n) {
br label %1
; <label>:1
%.09 = phi i32 [ %n, %0 ], [ %12, %1 ]
%.08 = phi i8* [ %b, %0 ], [ %11, %1 ]
%.0 = phi i8* [ %a, %0 ], [ %2, %1 ]
%2 = getelementptr inbounds i8* %.0, i32 1
%3 = load i8* %.0, align 1
%4 = sext i8 %3 to i32
%5 = getelementptr inbounds i8* %c, i32 %4
%6 = load i8* %5, align 1
%7 = add i8 %3, 1
%wrap.4 = sext i8 %7 to i32
%8 = getelementptr inbounds i8* %c, i32 %wrap.4
%9 = load i8* %8, align 1
store i8 %6, i8* %.08, align 1
%10 = getelementptr inbounds i8* %.08, i32 1
store i8 %9, i8* %10, align 1
%11 = getelementptr inbounds i8* %.08, i32 2
%12 = add nsw i32 %.09, -1
%13 = icmp eq i32 %12, 0
br i1 %13, label %14, label %1
; <label>:14
ret void
}