RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-19 19:12:56 +02:00
Code Issues Projects Releases Wiki Activity

TwoAddressInstructionPass doesn't really know how to merge live intervals when

Browse Source 
lowering REG_SEQUENCE instructions.

Insert copies for REG_SEQUENCE sources not killed to avoid breaking later passes.

llvm-svn: 104146
This commit is contained in:
Jakob Stoklund Olesen 2010-05-19 20:08:00 +00:00
parent e3e13b67e2
commit 6a2bfde3c8
2 changed files with 23 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
lib/CodeGen/TwoAddressInstructionPass.cpp
View File

@ -1275,9 +1275,10 @@ bool TwoAddressInstructionPass::EliminateRegSequences() {
if (!Seen.insert(SrcReg) ||
MI->getParent() != DefMI->getParent() ||
!MI->getOperand(i).isKill() ||
HasOtherRegSequenceUses(SrcReg, MI, MRI)) {
// REG_SEQUENCE cannot have duplicated operands, add a copy.
// Also add an copy if the source if live-in the block. We don't want
// Also add an copy if the source is live-in the block. We don't want
// to end up with a partial-redef of a livein, e.g.
// BB0:
// reg1051:10<def> =
@ -1287,6 +1288,10 @@ bool TwoAddressInstructionPass::EliminateRegSequences() {
// BB2:
// reg1051:9<def> =
// LiveIntervalAnalysis won't like it.
//
// If the REG_SEQUENCE doesn't kill its source, keeping live variables
// correctly up to date becomes very difficult. Insert a copy.
//
const TargetRegisterClass *RC = MRI->getRegClass(SrcReg);
unsigned NewReg = MRI->createVirtualRegister(RC);
MachineBasicBlock::iterator InsertLoc = MI;

17
test/CodeGen/ARM/reg_sequence.ll
View File

@ -302,6 +302,23 @@ bb14: ; preds = %bb6
ret i32 0
}
; This test crashes the coalescer because live variables were not updated properly.
define <8 x i8> @t11(i8* %A1, i8* %A2, i8* %A3, i8* %A4, i8* %A5, i8* %A6, i8* %A7, i8* %A8, i8* %B) nounwind {
%tmp1d = call %struct.__neon_int8x8x3_t @llvm.arm.neon.vld3.v8i8(i8* %A4) ; <%struct.__neon_int8x8x3_t> [#uses=1]
%tmp2d = extractvalue %struct.__neon_int8x8x3_t %tmp1d, 0 ; <<8 x i8>> [#uses=1]
%tmp1f = call %struct.__neon_int8x8x3_t @llvm.arm.neon.vld3.v8i8(i8* %A6) ; <%struct.__neon_int8x8x3_t> [#uses=1]
%tmp2f = extractvalue %struct.__neon_int8x8x3_t %tmp1f, 0 ; <<8 x i8>> [#uses=1]
%tmp2bd = add <8 x i8> zeroinitializer, %tmp2d ; <<8 x i8>> [#uses=1]
%tmp2abcd = mul <8 x i8> zeroinitializer, %tmp2bd ; <<8 x i8>> [#uses=1]
%tmp2ef = sub <8 x i8> zeroinitializer, %tmp2f ; <<8 x i8>> [#uses=1]
%tmp2efgh = mul <8 x i8> %tmp2ef, undef ; <<8 x i8>> [#uses=2]
call void @llvm.arm.neon.vst3.v8i8(i8* %A2, <8 x i8> undef, <8 x i8> undef, <8 x i8> %tmp2efgh)
%tmp2 = sub <8 x i8> %tmp2efgh, %tmp2abcd ; <<8 x i8>> [#uses=1]
%tmp7 = mul <8 x i8> undef, %tmp2 ; <<8 x i8>> [#uses=1]
tail call void @llvm.arm.neon.vst3.v8i8(i8* %B, <8 x i8> undef, <8 x i8> undef, <8 x i8> %tmp7)
ret <8 x i8> undef
}
declare <4 x i32> @llvm.arm.neon.vld1.v4i32(i8*) nounwind readonly
declare <8 x i16> @llvm.arm.neon.vld1.v8i16(i8*) nounwind readonly