RPCS3/llvm
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm.git synced 2024-12-20 11:08:27 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

[AArch64] Fix halfword load merging for big-endian targets

Browse Source 
For big-endian targets, when we merge two halfword loads into a word load, the
order of the halfwords in the loaded value is reversed compared to
little-endian, so the load-store optimiser needs to swap the destination
registers.

This does not affect merging of two word loads, as we use ldp, which treats the
memory as two separate 32-bit words.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@252597 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Oliver Stannard 2015-11-10 11:04:18 +00:00
parent be73ba8c82
commit 2358986dd4
2 changed files with 35 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
lib/Target/AArch64/AArch64LoadStoreOptimizer.cpp
View File

@ -84,6 +84,7 @@ struct AArch64LoadStoreOpt : public MachineFunctionPass {
const AArch64InstrInfo *TII; const AArch64InstrInfo *TII;
const TargetRegisterInfo *TRI; const TargetRegisterInfo *TRI;
const AArch64Subtarget *Subtarget;
// Scan the instructions looking for a load/store that can be combined // Scan the instructions looking for a load/store that can be combined
// with the current instruction into a load/store pair. // with the current instruction into a load/store pair.
@ -537,6 +538,10 @@ AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I,
if (!IsUnscaled) if (!IsUnscaled)
OffsetImm /= 2; OffsetImm /= 2;
MachineInstr *RtNewDest = MergeForward ? I : Paired; MachineInstr *RtNewDest = MergeForward ? I : Paired;
// When merging small (< 32 bit) loads for big-endian targets, the order of
// the component parts gets swapped.
if (!Subtarget->isLittleEndian())
std::swap(RtMI, Rt2MI);
// Construct the new load instruction. // Construct the new load instruction.
// FIXME: currently we support only halfword unsigned load. We need to // FIXME: currently we support only halfword unsigned load. We need to
// handle byte type, signed, and store instructions as well. // handle byte type, signed, and store instructions as well.
@ -560,7 +565,7 @@ AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I,
DEBUG((NewMemMI)->print(dbgs())); DEBUG((NewMemMI)->print(dbgs()));
MachineInstr *ExtDestMI = MergeForward ? Paired : I; MachineInstr *ExtDestMI = MergeForward ? Paired : I;
if (ExtDestMI == Rt2MI) { if ((ExtDestMI == Rt2MI) == Subtarget->isLittleEndian()) {
// Create the bitfield extract for high half. // Create the bitfield extract for high half.
BitExtMI1 = BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(), BitExtMI1 = BuildMI(*I->getParent(), InsertionPoint, I->getDebugLoc(),
TII->get(AArch64::UBFMWri)) TII->get(AArch64::UBFMWri))
@ -1388,8 +1393,9 @@ bool AArch64LoadStoreOpt::enableNarrowLdMerge(MachineFunction &Fn) {
} }
bool AArch64LoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) { bool AArch64LoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) {
TII = static_cast<const AArch64InstrInfo *>(Fn.getSubtarget().getInstrInfo()); Subtarget = &static_cast<const AArch64Subtarget &>(Fn.getSubtarget());
TRI = Fn.getSubtarget().getRegisterInfo(); TII = static_cast<const AArch64InstrInfo *>(Subtarget->getInstrInfo());
TRI = Subtarget->getRegisterInfo();
bool Modified = false; bool Modified = false;
bool enableNarrowLdOpt = enableNarrowLdMerge(Fn); bool enableNarrowLdOpt = enableNarrowLdMerge(Fn);

37
test/CodeGen/AArch64/arm64-ldr-merge.ll
View File

@ -1,36 +1,51 @@
; RUN: llc < %s -march=arm64 -mcpu=cortex-a57 -verify-machineinstrs | FileCheck %s ; RUN: llc < %s -march=arm64 -mcpu=cortex-a57 -verify-machineinstrs | FileCheck %s --check-prefix=CHECK --check-prefix=LE
; RUN: llc < %s -march=aarch64_be -mcpu=cortex-a57 -verify-machineinstrs | FileCheck %s --check-prefix=CHECK --check-prefix=BE
; CHECK-LABEL: Ldrh_merge ; CHECK-LABEL: Ldrh_merge
; CHECK-NOT: ldrh ; CHECK-NOT: ldrh
; CHECK: ldr [[NEW_DEST:w[0-9]+]] ; CHECK: ldr [[NEW_DEST:w[0-9]+]]
; CHECK: and w{{[0-9]+}}, [[NEW_DEST]], #0xffff ; CHECK-DAG: and [[LO_PART:w[0-9]+]], [[NEW_DEST]], #0xffff
; CHECK: lsr w{{[0-9]+}}, [[NEW_DEST]] ; CHECK-DAG: lsr [[HI_PART:w[0-9]+]], [[NEW_DEST]], #16
; LE: sub {{w[0-9]+}}, [[LO_PART]], [[HI_PART]]
; BE: sub {{w[0-9]+}}, [[HI_PART]], [[LO_PART]]
define i16 @Ldrh_merge(i16* nocapture readonly %p) { define i16 @Ldrh_merge(i16* nocapture readonly %p) {
%1 = load i16, i16* %p, align 2 %1 = load i16, i16* %p, align 2
%arrayidx2 = getelementptr inbounds i16, i16* %p, i64 1 %arrayidx2 = getelementptr inbounds i16, i16* %p, i64 1
%2 = load i16, i16* %arrayidx2, align 2 %2 = load i16, i16* %arrayidx2, align 2
%add = add nuw nsw i16 %1, %2 %add = sub nuw nsw i16 %1, %2
ret i16 %add ret i16 %add
} }
; CHECK-LABEL: Ldurh_merge ; CHECK-LABEL: Ldurh_merge
; CHECK-NOT: ldurh ; CHECK-NOT: ldurh
; CHECK: ldur [[NEW_DEST:w[0-9]+]] ; CHECK: ldur [[NEW_DEST:w[0-9]+]]
; CHECK: and w{{[0-9]+}}, [[NEW_DEST]], #0xffff ; CHECK-DAG: and [[LO_PART:w[0-9]+]], [[NEW_DEST]], #0xffff
; CHECK: lsr w{{[0-9]+}}, [[NEW_DEST]] ; CHECK-DAG: lsr [[HI_PART:w[0-9]+]], [[NEW_DEST]]
; LE: sub {{w[0-9]+}}, [[LO_PART]], [[HI_PART]]
; BE: sub {{w[0-9]+}}, [[HI_PART]], [[LO_PART]]
define i16 @Ldurh_merge(i16* nocapture readonly %p) { define i16 @Ldurh_merge(i16* nocapture readonly %p) {
entry: entry:
%arrayidx = getelementptr inbounds i16, i16* %p, i64 -2 %arrayidx = getelementptr inbounds i16, i16* %p, i64 -2
%0 = load i16, i16* %arrayidx %0 = load i16, i16* %arrayidx
%arrayidx3 = getelementptr inbounds i16, i16* %p, i64 -1 %arrayidx3 = getelementptr inbounds i16, i16* %p, i64 -1
%1 = load i16, i16* %arrayidx3 %1 = load i16, i16* %arrayidx3
%add = add nuw nsw i16 %0, %1 %add = sub nuw nsw i16 %0, %1
ret i16 %add ret i16 %add
} }
; CHECK-LABEL: Ldrh_4_merge ; CHECK-LABEL: Ldrh_4_merge
; CHECK-NOT: ldrh ; CHECK-NOT: ldrh
; CHECK: ldp [[NEW_DEST:w[0-9]+]] ; CHECK: ldp [[WORD1:w[0-9]+]], [[WORD2:w[0-9]+]], [x0]
; CHECK-DAG: and [[WORD1LO:w[0-9]+]], [[WORD1]], #0xffff
; CHECK-DAG: lsr [[WORD1HI:w[0-9]+]], [[WORD1]], #16
; CHECK-DAG: and [[WORD2LO:w[0-9]+]], [[WORD2]], #0xffff
; CHECK-DAG: lsr [[WORD2HI:w[0-9]+]], [[WORD2]], #16
; LE-DAG: sub [[TEMP1:w[0-9]+]], [[WORD1HI]], [[WORD1LO]]
; BE-DAG: sub [[TEMP1:w[0-9]+]], [[WORD1LO]], [[WORD1HI]]
; LE: udiv [[TEMP2:w[0-9]+]], [[TEMP1]], [[WORD2LO]]
; BE: udiv [[TEMP2:w[0-9]+]], [[TEMP1]], [[WORD2HI]]
; LE: sub w0, [[TEMP2]], [[WORD2HI]]
; BE: sub w0, [[TEMP2]], [[WORD2LO]]
define i16 @Ldrh_4_merge(i16* nocapture readonly %P) { define i16 @Ldrh_4_merge(i16* nocapture readonly %P) {
%arrayidx = getelementptr inbounds i16, i16* %P, i64 0 %arrayidx = getelementptr inbounds i16, i16* %P, i64 0
%l0 = load i16, i16* %arrayidx %l0 = load i16, i16* %arrayidx
@ -40,8 +55,8 @@ define i16 @Ldrh_4_merge(i16* nocapture readonly %P) {
%l2 = load i16, i16* %arrayidx7 %l2 = load i16, i16* %arrayidx7
%arrayidx12 = getelementptr inbounds i16, i16* %P, i64 3 %arrayidx12 = getelementptr inbounds i16, i16* %P, i64 3
%l3 = load i16, i16* %arrayidx12 %l3 = load i16, i16* %arrayidx12
%add4 = add nuw nsw i16 %l1, %l0 %add4 = sub nuw nsw i16 %l1, %l0
%add9 = add nuw nsw i16 %add4, %l2 %add9 = udiv i16 %add4, %l2
%add14 = add nuw nsw i16 %add9, %l3 %add14 = sub nuw nsw i16 %add9, %l3
ret i16 %add14 ret i16 %add14
} }