aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--lib/Target/Hexagon/HexagonPeephole.cpp35
-rw-r--r--test/CodeGen/Hexagon/remove_lsr.ll80
2 files changed, 115 insertions, 0 deletions
diff --git a/lib/Target/Hexagon/HexagonPeephole.cpp b/lib/Target/Hexagon/HexagonPeephole.cpp
index 55cbc09..a295015 100644
--- a/lib/Target/Hexagon/HexagonPeephole.cpp
+++ b/lib/Target/Hexagon/HexagonPeephole.cpp
@@ -109,6 +109,7 @@ bool HexagonPeephole::runOnMachineFunction(MachineFunction &MF) {
MRI = &MF.getRegInfo();
DenseMap<unsigned, unsigned> PeepholeMap;
+ DenseMap<unsigned, std::pair<unsigned, unsigned> > PeepholeDoubleRegsMap;
if (DisableHexagonPeephole) return false;
@@ -117,6 +118,7 @@ bool HexagonPeephole::runOnMachineFunction(MachineFunction &MF) {
MBBb != MBBe; ++MBBb) {
MachineBasicBlock* MBB = MBBb;
PeepholeMap.clear();
+ PeepholeDoubleRegsMap.clear();
// Traverse the basic block.
for (MachineBasicBlock::iterator MII = MBB->begin(); MII != MBB->end();
@@ -140,6 +142,24 @@ bool HexagonPeephole::runOnMachineFunction(MachineFunction &MF) {
}
}
+ // Look for this sequence below
+ // %vregDoubleReg1 = LSRd_ri %vregDoubleReg0, 32
+ // %vregIntReg = COPY %vregDoubleReg1:subreg_loreg.
+ // and convert into
+ // %vregIntReg = COPY %vregDoubleReg0:subreg_hireg.
+ if (MI->getOpcode() == Hexagon::LSRd_ri) {
+ assert(MI->getNumOperands() == 3);
+ MachineOperand &Dst = MI->getOperand(0);
+ MachineOperand &Src1 = MI->getOperand(1);
+ MachineOperand &Src2 = MI->getOperand(2);
+ if (Src2.getImm() != 32)
+ continue;
+ unsigned DstReg = Dst.getReg();
+ unsigned SrcReg = Src1.getReg();
+ PeepholeDoubleRegsMap[DstReg] =
+ std::make_pair(*&SrcReg, 1/*Hexagon::subreg_hireg*/);
+ }
+
// Look for P=NOT(P).
if (!DisablePNotP &&
(MI->getOpcode() == Hexagon::NOT_p)) {
@@ -178,6 +198,21 @@ bool HexagonPeephole::runOnMachineFunction(MachineFunction &MF) {
// Change the 1st operand.
MI->RemoveOperand(1);
MI->addOperand(MachineOperand::CreateReg(PeepholeSrc, false));
+ } else {
+ DenseMap<unsigned, std::pair<unsigned, unsigned> >::iterator DI =
+ PeepholeDoubleRegsMap.find(SrcReg);
+ if (DI != PeepholeDoubleRegsMap.end()) {
+ std::pair<unsigned,unsigned> PeepholeSrc = DI->second;
+ MI->RemoveOperand(1);
+ MI->addOperand(MachineOperand::CreateReg(PeepholeSrc.first,
+ false /*isDef*/,
+ false /*isImp*/,
+ false /*isKill*/,
+ false /*isDead*/,
+ false /*isUndef*/,
+ false /*isEarlyClobber*/,
+ PeepholeSrc.second));
+ }
}
}
}
diff --git a/test/CodeGen/Hexagon/remove_lsr.ll b/test/CodeGen/Hexagon/remove_lsr.ll
new file mode 100644
index 0000000..79b5f4a
--- /dev/null
+++ b/test/CodeGen/Hexagon/remove_lsr.ll
@@ -0,0 +1,80 @@
+; Test fix for PR-13709.
+; RUN: llc -march=hexagon -mcpu=hexagonv4 < %s | FileCheck %s
+; CHECK: foo
+; CHECK-NOT: lsr(r{{[0-9]+}}:{{[0-9]+}}, #32)
+; CHECK-NOT: lsr(r{{[0-9]+}}:{{[0-9]+}}, #32)
+
+; Convert the sequence
+; r17:16 = lsr(r11:10, #32)
+; .. = r16
+; into
+; r17:16 = lsr(r11:10, #32)
+; .. = r11
+; This makes the lsr instruction dead and it gets removed subsequently
+; by a dead code removal pass.
+
+%union.vect64 = type { i64 }
+%union.vect32 = type { i32 }
+
+define void @foo(%union.vect64* nocapture %sss_extracted_bit_rx_data_ptr,
+ %union.vect32* nocapture %s_even, %union.vect32* nocapture %s_odd,
+ i8* nocapture %scr_s_even_code_ptr, i8* nocapture %scr_s_odd_code_ptr)
+ nounwind {
+entry:
+ %scevgep = getelementptr %union.vect64* %sss_extracted_bit_rx_data_ptr, i32 1
+ %scevgep28 = getelementptr %union.vect32* %s_odd, i32 1
+ %scevgep32 = getelementptr %union.vect32* %s_even, i32 1
+ %scevgep36 = getelementptr i8* %scr_s_odd_code_ptr, i32 1
+ %scevgep39 = getelementptr i8* %scr_s_even_code_ptr, i32 1
+ br label %for.body
+
+for.body: ; preds = %for.body, %entry
+ %lsr.iv42 = phi i32 [ %lsr.iv.next, %for.body ], [ 2, %entry ]
+ %lsr.iv40 = phi i8* [ %scevgep41, %for.body ], [ %scevgep39, %entry ]
+ %lsr.iv37 = phi i8* [ %scevgep38, %for.body ], [ %scevgep36, %entry ]
+ %lsr.iv33 = phi %union.vect32* [ %scevgep34, %for.body ], [ %scevgep32, %entry ]
+ %lsr.iv29 = phi %union.vect32* [ %scevgep30, %for.body ], [ %scevgep28, %entry ]
+ %lsr.iv = phi %union.vect64* [ %scevgep26, %for.body ], [ %scevgep, %entry ]
+ %predicate_1.023 = phi i8 [ undef, %entry ], [ %10, %for.body ]
+ %predicate.022 = phi i8 [ undef, %entry ], [ %9, %for.body ]
+ %val.021 = phi i64 [ undef, %entry ], [ %srcval, %for.body ]
+ %lsr.iv3335 = bitcast %union.vect32* %lsr.iv33 to i32*
+ %lsr.iv2931 = bitcast %union.vect32* %lsr.iv29 to i32*
+ %lsr.iv27 = bitcast %union.vect64* %lsr.iv to i64*
+ %0 = tail call i64 @llvm.hexagon.A2.vsubhs(i64 0, i64 %val.021)
+ %conv3 = sext i8 %predicate.022 to i32
+ %1 = trunc i64 %val.021 to i32
+ %2 = trunc i64 %0 to i32
+ %3 = tail call i32 @llvm.hexagon.C2.mux(i32 %conv3, i32 %1, i32 %2)
+ store i32 %3, i32* %lsr.iv3335, align 4, !tbaa !0
+ %conv8 = sext i8 %predicate_1.023 to i32
+ %4 = lshr i64 %val.021, 32
+ %5 = trunc i64 %4 to i32
+ %6 = lshr i64 %0, 32
+ %7 = trunc i64 %6 to i32
+ %8 = tail call i32 @llvm.hexagon.C2.mux(i32 %conv8, i32 %5, i32 %7)
+ store i32 %8, i32* %lsr.iv2931, align 4, !tbaa !0
+ %srcval = load i64* %lsr.iv27, align 8
+ %9 = load i8* %lsr.iv40, align 1, !tbaa !1
+ %10 = load i8* %lsr.iv37, align 1, !tbaa !1
+ %lftr.wideiv = trunc i32 %lsr.iv42 to i8
+ %exitcond = icmp eq i8 %lftr.wideiv, 32
+ %scevgep26 = getelementptr %union.vect64* %lsr.iv, i32 1
+ %scevgep30 = getelementptr %union.vect32* %lsr.iv29, i32 1
+ %scevgep34 = getelementptr %union.vect32* %lsr.iv33, i32 1
+ %scevgep38 = getelementptr i8* %lsr.iv37, i32 1
+ %scevgep41 = getelementptr i8* %lsr.iv40, i32 1
+ %lsr.iv.next = add i32 %lsr.iv42, 1
+ br i1 %exitcond, label %for.end, label %for.body
+
+for.end: ; preds = %for.body
+ ret void
+}
+
+declare i64 @llvm.hexagon.A2.vsubhs(i64, i64) nounwind readnone
+
+declare i32 @llvm.hexagon.C2.mux(i32, i32, i32) nounwind readnone
+
+!0 = metadata !{metadata !"long", metadata !1}
+!1 = metadata !{metadata !"omnipotent char", metadata !2}
+!2 = metadata !{metadata !"Simple C/C++ TBAA", null}
generated by cgit v1.1 at 2024-10-05 18:16:21 +0000