summaryrefslogtreecommitdiff
path: root/lib
diff options
context:
space:
mode:
authorChad Rosier <mcrosier@apple.com>2011-11-17 07:15:58 +0000
committerChad Rosier <mcrosier@apple.com>2011-11-17 07:15:58 +0000
commit478b06c9801f0b75c5216ce0886ac3d630d4fc7b (patch)
tree54fa90701be13d4ab0be4cfb05352cfec4de49be /lib
parentec43d1f553cb440df1b435d3798063d0cba6a117 (diff)
downloadllvm-478b06c9801f0b75c5216ce0886ac3d630d4fc7b.tar.gz
llvm-478b06c9801f0b75c5216ce0886ac3d630d4fc7b.tar.bz2
llvm-478b06c9801f0b75c5216ce0886ac3d630d4fc7b.tar.xz
When fast iseling a GEP, accumulate the offset rather than emitting a series of
ADDs. MaxOffs is used as a threshold to limit the size of the offset. Tradeoffs being: (1) If we can't materialize the large constant then we'll cause fast-isel to bail. (2) Too large of an offset can't be directly encoded in the ADD resulting in a MOV+ADD. Generally not a bad thing because otherwise we would have had ADD+ADD, but on Thumb this turns into a MOVS+MOVT+ADD. Working on a fix for that. (3) Conversely, too low of a threshold we'll miss opportunities to coalesce ADDs. rdar://10412592 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@144886 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib')
-rw-r--r--lib/CodeGen/SelectionDAG/FastISel.cpp46
1 files changed, 35 insertions, 11 deletions
diff --git a/lib/CodeGen/SelectionDAG/FastISel.cpp b/lib/CodeGen/SelectionDAG/FastISel.cpp
index f0fe179e05..cff37c2c86 100644
--- a/lib/CodeGen/SelectionDAG/FastISel.cpp
+++ b/lib/CodeGen/SelectionDAG/FastISel.cpp
@@ -437,6 +437,11 @@ bool FastISel::SelectGetElementPtr(const User *I) {
bool NIsKill = hasTrivialKill(I->getOperand(0));
+ // Keep a running tab of the total offset to coalesce multiple N = N + Offset
+ // into a single N = N + TotalOffset.
+ uint64_t TotalOffs = 0;
+ // FIXME: What's a good SWAG number for MaxOffs?
+ uint64_t MaxOffs = 2048;
Type *Ty = I->getOperand(0)->getType();
MVT VT = TLI.getPointerTy();
for (GetElementPtrInst::const_op_iterator OI = I->op_begin()+1,
@@ -446,14 +451,15 @@ bool FastISel::SelectGetElementPtr(const User *I) {
unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
if (Field) {
// N = N + Offset
- uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field);
- // FIXME: This can be optimized by combining the add with a
- // subsequent one.
- N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT);
- if (N == 0)
- // Unhandled operand. Halt "fast" selection and bail.
- return false;
- NIsKill = true;
+ TotalOffs += TD.getStructLayout(StTy)->getElementOffset(Field);
+ if (TotalOffs >= MaxOffs) {
+ N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
+ if (N == 0)
+ // Unhandled operand. Halt "fast" selection and bail.
+ return false;
+ NIsKill = true;
+ TotalOffs = 0;
+ }
}
Ty = StTy->getElementType(Field);
} else {
@@ -462,14 +468,26 @@ bool FastISel::SelectGetElementPtr(const User *I) {
// If this is a constant subscript, handle it quickly.
if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
if (CI->isZero()) continue;
- uint64_t Offs =
+ // N = N + Offset
+ TotalOffs +=
TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
- N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT);
+ if (TotalOffs >= MaxOffs) {
+ N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
+ if (N == 0)
+ // Unhandled operand. Halt "fast" selection and bail.
+ return false;
+ NIsKill = true;
+ TotalOffs = 0;
+ }
+ continue;
+ }
+ if (TotalOffs) {
+ N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
if (N == 0)
// Unhandled operand. Halt "fast" selection and bail.
return false;
NIsKill = true;
- continue;
+ TotalOffs = 0;
}
// N = N + Idx * ElementSize;
@@ -494,6 +512,12 @@ bool FastISel::SelectGetElementPtr(const User *I) {
return false;
}
}
+ if (TotalOffs) {
+ N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
+ if (N == 0)
+ // Unhandled operand. Halt "fast" selection and bail.
+ return false;
+ }
// We successfully emitted code for the given LLVM Instruction.
UpdateValueMap(I, N);