mirror of
https://github.com/c64scene-ar/llvm-6502.git
synced 2024-12-30 17:33:24 +00:00
Speculatively revert commit 144124 (djg) in the hope that the 32 bit
dragonegg self-host buildbot will recover (it is complaining about object files differing between different build stages). Original commit message: Add a hack to the scheduler to disable pseudo-two-address dependencies in basic blocks containing calls. This works around a problem in which these artificial dependencies can get tied up in calling seqeunce scheduling in a way that makes the graph unschedulable with the current approach of using artificial physical register dependencies for calling sequences. This fixes PR11314. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@144188 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
parent
bb539bf973
commit
ef0b3ca3a8
@ -1666,7 +1666,7 @@ public:
|
||||
|
||||
protected:
|
||||
bool canClobber(const SUnit *SU, const SUnit *Op);
|
||||
void AddPseudoTwoAddrDeps(const TargetInstrInfo *TII);
|
||||
void AddPseudoTwoAddrDeps();
|
||||
void PrescheduleNodesWithMultipleUses();
|
||||
void CalculateSethiUllmanNumbers();
|
||||
};
|
||||
@ -2628,7 +2628,7 @@ bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
|
||||
void RegReductionPQBase::initNodes(std::vector<SUnit> &sunits) {
|
||||
SUnits = &sunits;
|
||||
// Add pseudo dependency edges for two-address nodes.
|
||||
AddPseudoTwoAddrDeps(TII);
|
||||
AddPseudoTwoAddrDeps();
|
||||
// Reroute edges to nodes with multiple uses.
|
||||
if (!TracksRegPressure)
|
||||
PrescheduleNodesWithMultipleUses();
|
||||
@ -2855,17 +2855,7 @@ void RegReductionPQBase::PrescheduleNodesWithMultipleUses() {
|
||||
/// one that has a CopyToReg use (more likely to be a loop induction update).
|
||||
/// If both are two-address, but one is commutable while the other is not
|
||||
/// commutable, favor the one that's not commutable.
|
||||
void RegReductionPQBase::AddPseudoTwoAddrDeps(const TargetInstrInfo *TII) {
|
||||
// If the graph contains any calls, disable this optimization.
|
||||
// FIXME: This is a kludge to work around the fact that the artificial edges
|
||||
// can combine with the way call sequences use physical register dependencies
|
||||
// to model their resource usage to create unschedulable graphs.
|
||||
for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
|
||||
for (SDNode *Node = (*SUnits)[i].getNode(); Node; Node = Node->getGluedNode())
|
||||
if (Node->isMachineOpcode() &&
|
||||
Node->getMachineOpcode() == (unsigned)TII->getCallFrameSetupOpcode())
|
||||
return;
|
||||
|
||||
void RegReductionPQBase::AddPseudoTwoAddrDeps() {
|
||||
for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
|
||||
SUnit *SU = &(*SUnits)[i];
|
||||
if (!SU->isTwoAddress)
|
||||
|
@ -3,9 +3,15 @@
|
||||
|
||||
; This testcase shouldn't need to spill the -1 value,
|
||||
; so it should just use pcmpeqd to materialize an all-ones vector.
|
||||
; For i386, cp load of -1 are folded.
|
||||
|
||||
; With -regalloc=greedy, the live range is split before spilling, so the first
|
||||
; pcmpeq doesn't get folded as a constant pool load.
|
||||
|
||||
; I386: pcmpeqd
|
||||
; I386-NOT: pcmpeqd
|
||||
; I386: orps LCPI0_2, %xmm
|
||||
; I386-NOT: pcmpeqd
|
||||
; I386: orps LCPI0_2, %xmm
|
||||
|
||||
; X86-64: pcmpeqd
|
||||
; X86-64-NOT: pcmpeqd
|
||||
|
@ -1,16 +0,0 @@
|
||||
; RUN: llc -march=x86 -mcpu=pentium4 -mtriple=i686-none-linux < %s
|
||||
; PR11314
|
||||
|
||||
; Make sure the scheduler's hack to insert artificial dependencies to optimize
|
||||
; two-address instruction scheduling doesn't interfere with the scheduler's
|
||||
; hack to model call sequences as artificial physical registers.
|
||||
|
||||
define inreg { i64, i64 } @sscanf(i32 inreg %base.1.i) nounwind {
|
||||
entry:
|
||||
%conv38.i92.i = sext i32 %base.1.i to i64
|
||||
%rem.i93.i = urem i64 10, %conv38.i92.i
|
||||
%div.i94.i = udiv i64 10, %conv38.i92.i
|
||||
%a = insertvalue { i64, i64 } undef, i64 %rem.i93.i, 0
|
||||
%b = insertvalue { i64, i64 } %a, i64 %div.i94.i, 1
|
||||
ret { i64, i64 } %b
|
||||
}
|
Loading…
Reference in New Issue
Block a user