similar to: [LLVMdev] [TableGen] How to specify multiple types to one register class.

On Thu, May 24, 2012 at 10:25:36PM +0800, Lei Mou wrote: > Hi all, > > In our architecture, there is a set of general purpose registers which could be > used to store integer and floating point numbers. My problem is how to correctly > specify the type of this register class. For now, its type is > specified as i32, and > it is problematic if it is used in floating point

Hi, Our target only has native support for i32 and f32 types. For data types smaller than these, I have to custom lowering truncating store using two atomic load instruction (which have the same semantics as ISD::ATOMIC_LOAD_AND and ATOMIC_LOAD_OR, respectively). I run into a problem during the legalization process, where the legalizer complains that ISD::STORE and ISD::ATOMIC_LOAD_OR (generated

Hi Lei, Le 31/05/2012 03:44, Lei Mou a écrit : > Problem solved by returning the second result of the ATOMIC_LOAD_OR node... You got the chain instead of the loaded value. IMHO, a better solution would have been to add a Pat<> rule to match truncstores and expand them into target store/load/and/or. Pat : <(truncstore16 GPR:$val, MEM:$mem), (store MEM:$mem, (or

Problem solved by returning the second result of the ATOMIC_LOAD_OR node... On Wed, May 30, 2012 at 9:38 PM, Lei Mou <lei.mou.uu at gmail.com> wrote: > Hi, > > Our target only has native support for i32 and f32 types. For data > types smaller than these, I have to custom lowering truncating store > using two atomic load instruction (which have the same semantics as >

Hi all, I'm learning the PTX backend and confused by the following problem. In the constructor of PTXFrameLowering, StackAlignment and LocalAreaOffset are assigned 2 and -2, respectively. Since PTX has neither stack frame nor stack pointer, why StackAlignment and LocalAreaOffset are needed and where does 2 and -2 come from? Any explanation is appreciated. Thank you in advance! Yours

On Mon, Apr 16, 2012 at 3:48 AM, Lei Mou <lei.mou.uu at gmail.com> wrote: > Hi all, > > I'm learning the PTX backend and confused by the following problem. In the > constructor of PTXFrameLowering, StackAlignment and LocalAreaOffset are > assigned 2 and -2, respectively. Since PTX has neither stack frame nor > stack pointer, why StackAlignment and LocalAreaOffset are

Hi, I'm working on an LLVM backend for GPUs. One thing that is a little different about some GPUs is that instructions can take registers or 32-bit floating point immediates as arguments. I was wondering if there is a way to model this using tablegen, without having to define an instruction for each possible combination of registers and immediates (e.g. For ADD it would require four

Hi all, While working on a patch to improve codegen for fadd reductions on AArch64, I stumbled upon an outstanding issue with the experimental.vector.reduce.fadd/fmul intrinsics where the accumulator argument is ignored when fast-math is set on the intrinsic call. This behaviour is described in the LangRef (https://www.llvm.org/docs/LangRef.html#id1905) and is mentioned in

Hi, I'm writing to ask the differences between a "reserved" register and an "unallocable" register. In X86 backend, for example, the stack pointer register and instruction pointer are reserved but allocatable. In the Doxygen document of function llvm::TargetRegisterInfo::getReservedRegs, it says that a reserved register is one that *has particular uses and should be

On 05/04/2019 09:37, Simon Pilgrim via llvm-dev wrote: > On 04/04/2019 14:11, Sander De Smalen wrote: >> Proposed change: >> >> ---------------------------- >> >> In this RFC I propose changing the intrinsics for >> llvm.experimental.vector.reduce.fadd and >> llvm.experimental.vector.reduce.fmul (see options A and B). I also >> propose renaming

On Mon, 2011-12-19 at 23:20 -0800, Andrew Trick wrote: > > On Dec 19, 2011, at 10:53 PM, Hal Finkel wrote: > > > Here's my "thought experiment" (from PR11589): I have a bunch of > > load-fadd-store chains to schedule. A store takes two cycles to > > clear > > its last pipeline stage. The fadd takes longer to compute its result > > (say 5

On Dec 19, 2011, at 10:53 PM, Hal Finkel wrote: > Here's my "thought experiment" (from PR11589): I have a bunch of > load-fadd-store chains to schedule. A store takes two cycles to clear > its last pipeline stage. The fadd takes longer to compute its result > (say 5 cycles), but can sustain a rate of 1 independent add per cycle. > As the scheduling is bottom-up, it

On Apr 17, 2012, at 9:09 PM, Lei Mou wrote: > I'm writing to ask the differences between a "reserved" register and an "unallocable" register. In X86 backend, for example, the stack pointer register and instruction pointer are reserved but allocatable. In the Doxygen document of function llvm::TargetRegisterInfo::getReservedRegs, it says that a reserved register is one

I'm trying to find a way to emulate a floating point instruction, say a floating point add. My understanding is that in order to do that I need to execute setOperationAction(ISD::FADD, (MVT::f32, Expand); setOperationAction(ISD::FADD, (MVT::f64, Expand); in MyTargetISelLowering.cpp, MyTargetLowering::MyTargetLowering(...). However for some reason I'm still seeing a floating point add in

Tom, There is no way to do this that I know of. Maybe David Greene or someone who hacks on Tablegen a lot would know. Micah > -----Original Message----- > From: llvmdev-bounces at cs.uiuc.edu [mailto:llvmdev-bounces at cs.uiuc.edu] > On Behalf Of Tom Stellard > Sent: Thursday, November 03, 2011 10:23 AM > To: LLVM Developers Mailing List > Subject: [LLVMdev] Tablegen:

Hello, We recently tracked down a bug on Tesla GPUs (i.e. G80-GT218) whereby it appears that instructions like 00000028: b5000409 08000780 add rn f32 $r2 $r2 neg c0[$a1] 00000040: b500060d 08004780 add rn f32 $r3 $r3 neg c0[$a1+0x4] or with nvdisasm: .headerflags @"EF_CUDA_SM12 EF_CUDA_PTX_SM(EF_CUDA_SM12)" /*0000*/ FADD R2, R2, -c[0x0][A1+0x0]; /*

On Tue, 2011-12-20 at 10:35 -0600, Hal Finkel wrote: > On Mon, 2011-12-19 at 23:20 -0800, Andrew Trick wrote: > > > > On Dec 19, 2011, at 10:53 PM, Hal Finkel wrote: > > > > > Here's my "thought experiment" (from PR11589): I have a bunch of > > > load-fadd-store chains to schedule. A store takes two cycles to > > > clear > >

On Apr 16, 2012, at 11:50 PM, Duncan Sands <baldrick at free.fr> wrote: > Hi Dan, > >> I realize that some of these thoughts apply equally to the >> prior !fpaccuracy metadata that's been around a while, but I >> hadn't looked at it closely until now. >> >> The !fpmath metadata violates the original spirit of >> metadata, which is that

On Dec 20, 2011, at 10:29 AM, Hal Finkel wrote: > On Tue, 2011-12-20 at 10:35 -0600, Hal Finkel wrote: >> On Mon, 2011-12-19 at 23:20 -0800, Andrew Trick wrote: >>> >>> On Dec 19, 2011, at 10:53 PM, Hal Finkel wrote: >>> >>>> Here's my "thought experiment" (from PR11589): I have a bunch of >>>> load-fadd-store chains to

On Mon, 2011-12-19 at 22:14 -0800, Andrew Trick wrote: > On Dec 19, 2011, at 3:19 PM, Hal Finkel wrote: > > > On Mon, 2011-12-19 at 07:41 -0800, Andrew Trick wrote: > >> On Dec 19, 2011, at 6:51 AM, Hal Finkel <hfinkel at anl.gov> wrote: > >> > >>> On Tue, 2011-10-25 at 21:00 -0700, Andrew Trick wrote: > >>> Now, to generate the best