similar to: [LLVMdev] Scheduling question (memory dependency)

Hi Bill, Which scheduler do you use? MI or SDNode one? In either case the problem is likely the same, but cause might be in a different place... The way I see it, you have an issue with the alias analyzer, not scheduler. When scheduling DAG is constructed, AA is checked for pairs of mem accessing objects, and if no potential interference is flagged by the AA the chain edge is _not_ inserted.

Here's another data point that may be useful. [Scheduling experts, please help! :) ] If the two-byte bitfield is replaced by a two-byte struct (replace "short i:8" with "short i", etc.), the scheduler properly generates a dependency between the store and the load. For this case, a GEP is used instead of a bitcast:

On Fri, 2012-09-21 at 11:34 -0500, William J. Schmidt wrote: > Hi Sergei, > > Thanks for the response! We just discovered there is likely a bug > happening during post-RA list scheduling. There's an invalid successor > index in the scheduling graph that is probably supposed to be the > missing arc. Starting to investigate further now. This is recorded in >

Hi Sergei, Thanks for the response! We just discovered there is likely a bug happening during post-RA list scheduling. There's an invalid successor index in the scheduling graph that is probably supposed to be the missing arc. Starting to investigate further now. This is recorded in http://llvm.org/bugs/show_bug.cgi?id=13891. Thanks, Bill On Fri, 2012-09-21 at 11:15 -0500, Sergei Larin

OK, finally found it. The AliasChain in ScheduleDAGInstrs::buildSchedGraph is not acting as a chain for loads and stores (the head of the chain is not being updated as they are encountered, so dependencies aren't being added solely on the basis of may-aliasing in some cases). Will test a patch. On Fri, 2012-09-21 at 13:04 -0500, William J. Schmidt wrote: > On Fri, 2012-09-21 at 11:34

I am having trouble trying to enable post RA scheduler for the Mips backend. This is the bit code of the function I am compiling: (gdb) p MF.Fn->dump() define void @PointToHPoint(%struct.HPointStruct* noalias sret %agg.result, %struct.ObjPointStruct* byval %P) nounwind { entry: %res = alloca %struct.HPointStruct, align 8 %x2 = bitcast %struct.ObjPointStruct* %P to double* %0 = load

I filed a bug report (Bug 12205). Please take a look when you have time. Per your suggestion, I also attached a patch which attaches to load or store nodes a machinepointerinfo that points to a stack frame object when it can infer they are actually reading from or writing to the stack. The test that was failing passes if I apply this patch, but I doubt this is the right approach, because this

On Mar 6, 2012, at 5:05 PM, Akira Hatanaka <ahatanak at gmail.com> wrote: > I am having trouble trying to enable post RA scheduler for the Mips backend. > > This is the bit code of the function I am compiling: > > (gdb) p MF.Fn->dump() > > define void @PointToHPoint(%struct.HPointStruct* noalias sret > %agg.result, %struct.ObjPointStruct* byval %P) nounwind {

On Mar 7, 2012, at 11:34 AM, Akira Hatanaka <ahatanak at gmail.com> wrote: > I filed a bug report (Bug 12205). > Please take a look when you have time. > > Per your suggestion, I also attached a patch which attaches to load or > store nodes a machinepointerinfo that points to a stack frame object > when it can infer they are actually reading from or writing to the >

Thanks for spelling it out, now I understand. On Jun 5, 2011, at 6:11 AM, Galanov, Sergey wrote: > Well, the point is CMOV_GR* are marked clobbering EFLAGS conservatively just in case they turn out to be lowered into a sequence containing XOR %reg,%reg which indeed clobbers EFLAGS. This means there might not be any instruction which actually uses this EFLAGS value. This actually looks like a

On Jun 3, 2011, at 2:59 AM, Galanov, Sergey wrote: > Hi, Bill and Jakob. > > I don't quite understand. I am talking about CMOV_GR* instructions which are conservatively marked as clobbering EFLAGS in X86InstrCompiler.td. Doesn't that mean there cannot be any use of EFLAGS in subsequent instructions before it is defined by some other instruction? > > I also don't

Hi, while writing my register allocator, I have come across a case which confuses me because the llvm definition cannot be mapped to machine code. For instance I come across (1) and I reduce it to (2). However a copy instruction cannot move from EDX to CX. What mechanics in LLVM will tell me that I cannot make this move during register allocation, or how can I tell from (1) that I cannot execute

Well, the point is CMOV_GR* are marked clobbering EFLAGS conservatively just in case they turn out to be lowered into a sequence containing XOR %reg,%reg which indeed clobbers EFLAGS. This means there might not be any instruction which actually uses this EFLAGS value. For an example we can look no further than the actual test which has been disabled after the fix

On Nov 8, 2010, at 12:04 PM, Jeff Kunkel wrote: > Hi, while writing my register allocator, I have come across a case which confuses me because the llvm definition cannot be mapped to machine code. > > For instance I come across (1) and I reduce it to (2). However a copy instruction cannot move from EDX to CX. What mechanics in LLVM will tell me that I cannot make this move during

Hi, I've been looking at the Machine Scheduler on Power PC. I am looking only at the pre-RA machine scheduler and I am running it in the default bi-directional mode (so, both top down and bottom up queues are considered). I've come across an example where the scheduler picks a poor ordering for the instructions which results in very high register pressure which results in spills.

I am wondering about register allocation when there is a kill flag on the MachineOperand. Do I need to remove the kill flag? This code below is just an example from test\CodeGen\X86\xor.ll # Machine code for function test3: Frame Objects: fi#-2: size=4, align=4, fixed, at location [SP+8] fi#-1: size=4, align=8, fixed, at location [SP+4] Function Live Outs: %EAX BB#0: derived from LLVM BB

Andy, Lang, Thanks for the suggestion. I have spent more time with it today, and I do see some strange things in liveness update. I am not at the actual cause yet, but here is what I got so far: I have the following live ranges when I start scheduling a region: R2 = [0B,48r:0)[352r,416r:5)... R3 = [0B,48r:0)[368r,416r:5)... R4 = [0B,32r:0)[384r,416r:4)... R5 = [0B,32r:0)[400r,416r:4)...

On Aug 30, 2012, at 1:20 PM, Arnold Schwaighofer <arnolds at codeaurora.org> wrote: > The code in collectRanges() does: > > // Collect ranges for register units. These live ranges are computed on > // demand, so just skip any that haven't been computed yet. > if (TargetRegisterInfo::isPhysicalRegister(Reg)) { > for (MCRegUnitIterator Units(Reg,

Hi Sergei, Andy, Sorry - I got distracted with some other work. I'm looking into this and PR13719 now. I'll let you know what I find out. Sergei - thanks very much for the investigation. That should help me pin this down. Cheers, Lang. On Tue, Aug 28, 2012 at 2:33 PM, Sergei Larin <slarin at codeaurora.org> wrote: > Andy, Lang, > > Thanks for the suggestion. >

The code in collectRanges() does: // Collect ranges for register units. These live ranges are computed on // demand, so just skip any that haven't been computed yet. if (TargetRegisterInfo::isPhysicalRegister(Reg)) { for (MCRegUnitIterator Units(Reg, &TRI); Units.isValid(); ++Units) if (LiveInterval *LI = LIS.getCachedRegUnit(*Units))