RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-19 11:02:59 +02:00
Code Issues Projects Releases Wiki Activity

CodeGenPrepare: Sink addressing modes for atomics

Browse Source 
llvm-svn: 297903
This commit is contained in:
Matt Arsenault 2017-03-15 22:35:20 +00:00
parent adb22c6868
commit fc2f0bb351
2 changed files with 112 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
lib/CodeGen/CodeGenPrepare.cpp
View File

@ -3868,11 +3868,28 @@ static bool FindAllMemoryUses(
if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
unsigned opNo = U.getOperandNo();
if (opNo == 0) return true; // Storing addr, not into addr.
if (opNo != StoreInst::getPointerOperandIndex())
return true; // Storing addr, not into addr.
MemoryUses.push_back(std::make_pair(SI, opNo));
continue;
}
if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) {
unsigned opNo = U.getOperandNo();
if (opNo != AtomicRMWInst::getPointerOperandIndex())
return true; // Storing addr, not into addr.
MemoryUses.push_back(std::make_pair(RMW, opNo));
continue;
}
if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) {
unsigned opNo = U.getOperandNo();
if (opNo != AtomicCmpXchgInst::getPointerOperandIndex())
return true; // Storing addr, not into addr.
MemoryUses.push_back(std::make_pair(CmpX, opNo));
continue;
}
if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
// If this is a cold call, we can sink the addressing calculation into
// the cold path. See optimizeCallInst
@ -5784,6 +5801,18 @@ bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
return false;
}
if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
unsigned AS = RMW->getPointerAddressSpace();
return optimizeMemoryInst(I, RMW->getPointerOperand(),
RMW->getType(), AS);
}
if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(I)) {
unsigned AS = CmpX->getPointerAddressSpace();
return optimizeMemoryInst(I, CmpX->getPointerOperand(),
CmpX->getCompareOperand()->getType(), AS);
}
BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
if (BinOp && (BinOp->getOpcode() == Instruction::And) &&

83
test/CodeGen/AMDGPU/cgp-addressing-modes.ll
View File

@ -5,6 +5,8 @@
; RUN: llc -march=amdgcn -mcpu=bonaire -mattr=-promote-alloca -amdgpu-sroa=0 < %s | FileCheck -check-prefix=GCN -check-prefix=CI %s
; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -mattr=-promote-alloca -amdgpu-sroa=0 < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
target datalayout = "e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32-p24:64:64-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64"
; OPT-LABEL: @test_sink_global_small_offset_i32(
; OPT-CI-NOT: getelementptr i32, i32 addrspace(1)* %in
; OPT-VI: getelementptr i32, i32 addrspace(1)* %in
@ -486,7 +488,7 @@ done:
%struct.foo = type { [3 x float], [3 x float] }
; OPT-LABEL: @sink_ds_address(
; OPT: ptrtoint %struct.foo addrspace(3)* %ptr to i64
; OPT: ptrtoint %struct.foo addrspace(3)* %ptr to i32
; GCN-LABEL: {{^}}sink_ds_address:
; GCN: s_load_dword [[SREG1:s[0-9]+]],
@ -541,6 +543,85 @@ done:
ret void
}
; OPT-LABEL: @test_sink_local_small_offset_atomicrmw_i32(
; OPT: %sunkaddr = ptrtoint i32 addrspace(3)* %in to i32
; OPT: %sunkaddr1 = add i32 %sunkaddr, 28
; OPT: %sunkaddr2 = inttoptr i32 %sunkaddr1 to i32 addrspace(3)*
; OPT: %tmp1 = atomicrmw add i32 addrspace(3)* %sunkaddr2, i32 2 seq_cst
define void @test_sink_local_small_offset_atomicrmw_i32(i32 addrspace(3)* %out, i32 addrspace(3)* %in) {
entry:
%out.gep = getelementptr i32, i32 addrspace(3)* %out, i32 999999
%in.gep = getelementptr i32, i32 addrspace(3)* %in, i32 7
%tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #0
%tmp0 = icmp eq i32 %tid, 0
br i1 %tmp0, label %endif, label %if
if:
%tmp1 = atomicrmw add i32 addrspace(3)* %in.gep, i32 2 seq_cst
br label %endif
endif:
%x = phi i32 [ %tmp1, %if ], [ 0, %entry ]
store i32 %x, i32 addrspace(3)* %out.gep
br label %done
done:
ret void
}
; OPT-LABEL: @test_sink_local_small_offset_cmpxchg_i32(
; OPT: %sunkaddr = ptrtoint i32 addrspace(3)* %in to i32
; OPT: %sunkaddr1 = add i32 %sunkaddr, 28
; OPT: %sunkaddr2 = inttoptr i32 %sunkaddr1 to i32 addrspace(3)*
; OPT: %tmp1.struct = cmpxchg i32 addrspace(3)* %sunkaddr2, i32 undef, i32 2 seq_cst monotonic
define void @test_sink_local_small_offset_cmpxchg_i32(i32 addrspace(3)* %out, i32 addrspace(3)* %in) {
entry:
%out.gep = getelementptr i32, i32 addrspace(3)* %out, i32 999999
%in.gep = getelementptr i32, i32 addrspace(3)* %in, i32 7
%tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #0
%tmp0 = icmp eq i32 %tid, 0
br i1 %tmp0, label %endif, label %if
if:
%tmp1.struct = cmpxchg i32 addrspace(3)* %in.gep, i32 undef, i32 2 seq_cst monotonic
%tmp1 = extractvalue { i32, i1 } %tmp1.struct, 0
br label %endif
endif:
%x = phi i32 [ %tmp1, %if ], [ 0, %entry ]
store i32 %x, i32 addrspace(3)* %out.gep
br label %done
done:
ret void
}
; OPT-LABEL: @test_wrong_operand_local_small_offset_cmpxchg_i32(
; OPT: %in.gep = getelementptr i32, i32 addrspace(3)* %in, i32 7
; OPT: br i1
; OPT: cmpxchg i32 addrspace(3)* addrspace(3)* undef, i32 addrspace(3)* %in.gep, i32 addrspace(3)* undef seq_cst monotonic
define void @test_wrong_operand_local_small_offset_cmpxchg_i32(i32 addrspace(3)* addrspace(3)* %out, i32 addrspace(3)* %in) {
entry:
%out.gep = getelementptr i32 addrspace(3)*, i32 addrspace(3)* addrspace(3)* %out, i32 999999
%in.gep = getelementptr i32, i32 addrspace(3)* %in, i32 7
%tid = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0) #0
%tmp0 = icmp eq i32 %tid, 0
br i1 %tmp0, label %endif, label %if
if:
%tmp1.struct = cmpxchg i32 addrspace(3)* addrspace(3)* undef, i32 addrspace(3)* %in.gep, i32 addrspace(3)* undef seq_cst monotonic
%tmp1 = extractvalue { i32 addrspace(3)*, i1 } %tmp1.struct, 0
br label %endif
endif:
%x = phi i32 addrspace(3)* [ %tmp1, %if ], [ null, %entry ]
store i32 addrspace(3)* %x, i32 addrspace(3)* addrspace(3)* %out.gep
br label %done
done:
ret void
}
declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #0
attributes #0 = { nounwind readnone }