similar to: [LLVMdev] Eliminating copies between overlapping register classes

On Feb 22, 2012, at 12:01 PM, Tom Stellard wrote: > Hi, > > I have two register classes A and B, where A contains a subset of the > registers in B: > > A = [R0, R1, R2, ... R128] > > B = [RO, R1, R2, ... R128, > T0, T1, T2, ... T128] > > I am using the Greedy Register Allocator, and I would expect the register > allocator to eliminate this copy: >

On Wed, Feb 22, 2012 at 07:00:55PM -0800, Jakob Stoklund Olesen wrote: > > On Feb 22, 2012, at 12:01 PM, Tom Stellard wrote: > > > Hi, > > > > I have two register classes A and B, where A contains a subset of the > > registers in B: > > > > A = [R0, R1, R2, ... R128] > > > > B = [RO, R1, R2, ... R128, > > T0, T1, T2, ... T128]

Vincent, File a bug report so you can get a fix for it. Ivan On 25/10/2012 23:01, Vincent Lejeune wrote: > Thank for your help. You're right, merging vreg32 and vreg48 is perfectly fine, sorry I missed that. > I "brute force" debuged by adding MachineFunction dump after each join, I think I found the issue : it's when vreg32 and vreg10 are merged. > vreg10 only

> I expected that this optimization would be picked > up in a cse, gvn, machine-cse or even peepholing pass. > > Comments? At the LLVM IR level this is represented as define i64 @caller() #0 { entry: store i64* @val, i64** @p, align 8, !tbaa !0 store i64 12345123400, i64* @val, align 8, !tbaa !3 %call = tail call i64 @xtr(i64 12345123400) #2 ret i64 %call } Which is

On Aug 2, 2013, at 1:37 PM, Rafael Espíndola <rafael.espindola at gmail.com> wrote: >> I expected that this optimization would be picked >> up in a cse, gvn, machine-cse or even peepholing pass. >> >> Comments? > > > At the LLVM IR level this is represented as > > define i64 @caller() #0 { > entry: > store i64* @val, i64** @p, align 8, !tbaa

Thank for your help. You're right, merging vreg32 and vreg48 is perfectly fine, sorry I missed that. I "brute force" debuged by adding MachineFunction dump after each join, I think I found the issue : it's when vreg32 and vreg10 are merged. vreg10 only appears in BB#3, and the join only occurs in BB#3 apparently even if vreg32 lives in the 4 machine blocks After joining, there

Currently it will always spill / restore the whole vreg but only spilling the parts that are actually live would be a nice addition in the future. Looking at r192119': if "mtlo" writes to $LO and sets $HI to an unpredictable value, then it should just have an additional (dead) def operand for $hi, shouldn't it? Greetings Matthias Am 10/8/13, 11:03 AM, schrieb Akira

What I didn't mention in r192119 is that mthi/lo clobbers the other sub-register only if the contents of hi and lo are produced by mult or other arithmetic instructions (div, madd, etc.) It doesn't have this side-effect if it is produced by another mthi/lo. So I don't think making mthi/lo clobber the other half would work. For example, this is an illegal sequence of instructions,

I've been working on patches to improve subregister liveness tracking on llvm and I wanted to inform the llvm community about the overal design/motivation for them. I will send the patches to llvm-commits later today. Greetings Matthias Braun Subregisters in llvm ==================== Some targets can access registers in different ways resulting in wider or narrower accesses. For

On Oct 8, 2013, at 2:06 PM, Akira Hatanaka <ahatanak at gmail.com> wrote: > What I didn't mention in r192119 is that mthi/lo clobbers the other sub-register only if the contents of hi and lo are produced by mult or other arithmetic instructions (div, madd, etc.) It doesn't have this side-effect if it is produced by another mthi/lo. So I don't think making mthi/lo clobber the

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.

Ok, after a long detour I am back to where I have started. I think there is a problem at dep DAG construction. Let me try to convince you. Here is the C code we are dealing with: push () { struct xx_stack *stack, *top; for (stack = xx_stack; stack; stack = stack->next) top = stack; yy_instr = top->first; } If the loop never iterates, "top" will have

Sergei, Absolutely right, the copy/ldriw should not be reordered. I was attempting to explain that I consider it a phi-elimination bug, not a DAG builder bug. Liveness will also have problems with this code in the long run. To avoid confusion, I filed PR13112: Phi elimination generates uninitialized vreg uses, and disabled the SSA check until its fixes in r158461. However, your C code is also

Hi, I am working on a simple copy propagation pass for the R600 backend that propagates immediates rather than registers. For example, I want to transform: ... %vreg1 = V_MOV_B32 1 %vreg2 = V_ADD_I32 %vreg1, %vreg0 ... into: %vreg1 = V_MOV_B32 1 ; <- Only delete this if it is dead %vreg2 = V_ADD_I32 1, %vreg0 For best results, I am trying to run this pass after the TwoAddressInstruction

On Jun 13, 2012, at 1:15 PM, Sergei Larin <slarin at codeaurora.org> wrote: > Andy, > > You are probably right here – look at this – before phi elimination this code looks much more sane: > > # *** IR Dump After Live Variable Analysis ***: > # Machine code for function push: SSA > Function Live Outs: %R0 > > BB#0: derived from LLVM BB %entry >

This code lives in DAGCombiner.cpp: ------------- // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr' if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) { // NOTE: If the original store is volatile, this transform must not increase // the number of stores. For example, on x86-32 an f64 can be stored in one // processor operation but

Hi, We recently found a testcase showing that simplifications in instcombine sometimes change the instruction without reducing the instruction cost, but causing problems in TwoAddressInstruction pass. And it looks like the problem is generic and other simplification may have the same issue. I want to get some ideas about what is the best way to fix such kind of problem. The testcase:

Andy, You are probably right here - look at this - before phi elimination this code looks much more sane: # *** IR Dump After Live Variable Analysis ***: # Machine code for function push: SSA Function Live Outs: %R0 BB#0: derived from LLVM BB %entry %vreg5<def> = IMPLICIT_DEF; IntRegs:%vreg5 %vreg4<def> = TFRI_V4 <ga:@xx_stack>; IntRegs:%vreg4

Hi all, Our target does not have native support for 64-bit integers, so we rely on library calls for certain operations (like sdiv). We recently ran into a problem where these operations that are expanded to library calls aren't maintaining the proper ordering in relation to other chains in the DAG. The following snippet of a DAG demonstrates the problem. t0: ch = EntryToken t2:

Is there any way to get it to delay this optimization where it goes from this: Initial selection DAG: BB#0 'bclr64:entry' SelectionDAG has 14 nodes: t0: ch = EntryToken t2: i64,ch = CopyFromReg t0, Register:i64 %vreg0 t4: i64,ch = CopyFromReg t0, Register:i64 %vreg1 t6: i64 = sub t4, Constant:i64<1> t7: i64 = shl Constant:i64<1>, t6