[llvm] r322534 - [BPF] Teach DAG2DAG AND elimination about load intrinsics
Yonghong Song via llvm-commits
llvm-commits at lists.llvm.org
Mon Jan 15 23:27:19 PST 2018
Author: yhs
Date: Mon Jan 15 23:27:19 2018
New Revision: 322534
URL: http://llvm.org/viewvc/llvm-project?rev=322534&view=rev
Log:
[BPF] Teach DAG2DAG AND elimination about load intrinsics
As commented on the existing code:
// The Reg operand should be a virtual register, which is defined
// outside the current basic block. DAG combiner has done a pretty
// good job in removing truncating inside a single basic block.
However, when the Reg operand comes from bpf_load_[byte | half | word]
intrinsics, the generic optimizer doesn't understand their results are
zero extended, so these single basic block elimination opportunities were
missed.
Acked-by: Jakub Kicinski <jakub.kicinski at netronome.com>
Acked-by: Yonghong Song <yhs at fb.com>
Signed-off-by: Jiong Wang <jiong.wang at netronome.com>
Added:
llvm/trunk/test/CodeGen/BPF/remove_truncate_4.ll
Modified:
llvm/trunk/lib/Target/BPF/BPFISelDAGToDAG.cpp
Modified: llvm/trunk/lib/Target/BPF/BPFISelDAGToDAG.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/BPF/BPFISelDAGToDAG.cpp?rev=322534&r1=322533&r2=322534&view=diff
==============================================================================
--- llvm/trunk/lib/Target/BPF/BPFISelDAGToDAG.cpp (original)
+++ llvm/trunk/lib/Target/BPF/BPFISelDAGToDAG.cpp Mon Jan 15 23:27:19 2018
@@ -519,6 +519,37 @@ void BPFDAGToDAGISel::PreprocessTrunc(SD
if (!MaskN)
return;
+ // The Reg operand should be a virtual register, which is defined
+ // outside the current basic block. DAG combiner has done a pretty
+ // good job in removing truncating inside a single basic block except
+ // when the Reg operand comes from bpf_load_[byte | half | word] for
+ // which the generic optimizer doesn't understand their results are
+ // zero extended.
+ SDValue BaseV = Node->getOperand(0);
+ if (BaseV.getOpcode() == ISD::INTRINSIC_W_CHAIN) {
+ unsigned IntNo = cast<ConstantSDNode>(BaseV->getOperand(1))->getZExtValue();
+ uint64_t MaskV = MaskN->getZExtValue();
+
+ if (!((IntNo == Intrinsic::bpf_load_byte && MaskV == 0xFF) ||
+ (IntNo == Intrinsic::bpf_load_half && MaskV == 0xFFFF) ||
+ (IntNo == Intrinsic::bpf_load_word && MaskV == 0xFFFFFFFF)))
+ return;
+
+ DEBUG(dbgs() << "Remove the redundant AND operation in: "; Node->dump();
+ dbgs() << '\n');
+
+ I--;
+ CurDAG->ReplaceAllUsesWith(SDValue(Node, 0), BaseV);
+ I++;
+ CurDAG->DeleteNode(Node);
+
+ return;
+ }
+
+ // Multiple basic blocks case.
+ if (BaseV.getOpcode() != ISD::CopyFromReg)
+ return;
+
unsigned match_load_op = 0;
switch (MaskN->getZExtValue()) {
default:
@@ -534,13 +565,6 @@ void BPFDAGToDAGISel::PreprocessTrunc(SD
break;
}
- // The Reg operand should be a virtual register, which is defined
- // outside the current basic block. DAG combiner has done a pretty
- // good job in removing truncating inside a single basic block.
- SDValue BaseV = Node->getOperand(0);
- if (BaseV.getOpcode() != ISD::CopyFromReg)
- return;
-
const RegisterSDNode *RegN =
dyn_cast<RegisterSDNode>(BaseV.getNode()->getOperand(1));
if (!RegN || !TargetRegisterInfo::isVirtualRegister(RegN->getReg()))
Added: llvm/trunk/test/CodeGen/BPF/remove_truncate_4.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/BPF/remove_truncate_4.ll?rev=322534&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/BPF/remove_truncate_4.ll (added)
+++ llvm/trunk/test/CodeGen/BPF/remove_truncate_4.ll Mon Jan 15 23:27:19 2018
@@ -0,0 +1,58 @@
+; RUN: llc < %s -march=bpf -verify-machineinstrs | FileCheck %s
+
+; Source code:
+;struct __sk_buff;
+;unsigned long long
+;load_byte(void *skb, unsigned long long off) asm("llvm.bpf.load.byte");
+;unsigned long long
+;load_half(void *skb, unsigned long long off) asm("llvm.bpf.load.half");
+;typedef unsigned char uint8_t;
+;typedef unsigned short uint16_t;
+;
+;int func_b(struct __sk_buff *skb)
+;{
+; uint8_t t = load_byte(skb, 0);
+; return t;
+;}
+;
+;int func_h(struct __sk_buff *skb)
+;{
+; uint16_t t = load_half(skb, 0);
+; return t;
+;}
+;
+;int func_w(struct __sk_buff *skb)
+;{
+; uint32_t t = load_word(skb, 0);
+; return t;
+;}
+
+%struct.__sk_buff = type opaque
+
+; Function Attrs: nounwind readonly
+define i32 @func_b(%struct.__sk_buff* %skb) local_unnamed_addr #0 {
+entry:
+ %0 = bitcast %struct.__sk_buff* %skb to i8*
+ %call = tail call i64 @llvm.bpf.load.byte(i8* %0, i64 0)
+ %conv = trunc i64 %call to i32
+ %conv1 = and i32 %conv, 255
+; CHECK-NOT: r0 &= 255
+ ret i32 %conv1
+}
+
+; Function Attrs: nounwind readonly
+declare i64 @llvm.bpf.load.byte(i8*, i64) #1
+
+; Function Attrs: nounwind readonly
+define i32 @func_h(%struct.__sk_buff* %skb) local_unnamed_addr #0 {
+entry:
+ %0 = bitcast %struct.__sk_buff* %skb to i8*
+ %call = tail call i64 @llvm.bpf.load.half(i8* %0, i64 0)
+ %conv = trunc i64 %call to i32
+ %conv1 = and i32 %conv, 65535
+; CHECK-NOT: r0 &= 65535
+ ret i32 %conv1
+}
+
+; Function Attrs: nounwind readonly
+declare i64 @llvm.bpf.load.half(i8*, i64) #1
More information about the llvm-commits
mailing list