similar to: avx512 JIT backend generates wrong code on <4 x float>

Hi Frank, I recommend trying trunk LLVM. AVX-512 development has been very active recently. -Hal ----- Original Message ----- > From: "Frank Winter via llvm-dev" <llvm-dev at lists.llvm.org> > To: "LLVM Dev" <llvm-dev at lists.llvm.org> > Sent: Wednesday, June 29, 2016 2:41:39 PM > Subject: [llvm-dev] avx512 JIT backend generates wrong code on <4

Hi Hal! Thanks, but unfortunately it didn't help. The exact same assembler instructions are generated for both 3.8 (yesterday) and trunk (from today). So, this really looks like a bug. Best, Frank On 06/29/2016 03:48 PM, Hal Finkel wrote: > Hi Frank, > > I recommend trying trunk LLVM. AVX-512 development has been very active recently. > > -Hal > > ----- Original

With LLVM 3.8 the JIT compiler engine generates an AVX512 instruction although I target an 'avx2' CPU (intel Core I7). I just downloaded the most recent 3.8 and still it happens. It happens with this input module: target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" define void @module_cFFEMJ(i64 %lo, i64 %hi, i64 %myId, i1 %ordered, i64 %start, i32* noalias align 32

On 06/23/2016 12:56 PM, Craig Topper wrote: > Can you check what value "getHostCPUName" returned? getHostCPUName() = skylake > > On Thu, Jun 23, 2016 at 9:53 AM, Frank Winter via llvm-dev > <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote: > > With LLVM 3.8 the JIT compiler engine generates an AVX512 > instruction although I

On 07/24/2015 03:42 AM, Benjamin Kramer wrote: >> On 24.07.2015, at 08:06, zhi chen <zchenhn at gmail.com> wrote: >> >> It seems that that it's hard to vectorize int64 in LLVM. For example, LLVM 3.4 generates very complicated code for the following IR. I am running on a Haswell processor. Is it because there is no alternative AVX/2 instructions for int64? The same thing

------------------------------------ IR ------------------------------------------------------------------ if.then.i.i.i.i.i.i: ; preds = %if.then4 %S25_D = zext <2 x i32> %splatLDS17_D.splat to <2 x i64> %umul_with_overflow.i.iS26_D = shl <2 x i64> %S25_D, <i64 3, i64 3> %extumul_with_overflow.i.iS26_D = extractelement <2 x i64>

This snippet of IR is interesting: %sub.ptr.div.iS37_D = sdiv <2 x i64> %sub.ptr.sub.iS36_D, <i64 24, i64 24> %cmp10S38_D = icmp ugt <2 x i64> %sub.ptr.div.iS37_D, %splatInsMapS1_D.splat %zextS39_D = sext <2 x i1> %cmp10S38_D to <2 x i64> %BCS39_D = bitcast <2 x i64> %zextS39_D to i128 %mskS39_D = icmp ne i128 %BCS39_D, 0 br i1 %mskS39_D,

On Jan 9, 2012, at 10:00 AM, Jakob Stoklund Olesen wrote: > > On Jan 8, 2012, at 11:18 PM, Demikhovsky, Elena wrote: > >> I'll explain what we see in the code. >> 1. The caller saves XMM registers across the call if needed (according to DEFS definition). >> YMMs are not in the set, so caller does not take care. > > This is not how the register allocator

Hi, While investigating a performance issue with an internal codebase I came across what looks to be poor register allocation. I have constructed a small(ish) reproducible which demonstrates the issue (see test.ll attached). I have spent some time going through the register allocator to understand what is happening. I have also experimented with some small changes to try and improve the

This is the wrong code: declare <16 x float> @foo(<16 x float>) define <16 x float> @test(<16 x float> %x, <16 x float> %y) nounwind { entry: %x1 = fadd <16 x float> %x, %y %call = call <16 x float> @foo(<16 x float> %x1) nounwind %y1 = fsub <16 x float> %call, %y ret <16 x float> %y1 } ./llc -mattr=+avx

Hi, Please consider the following loop: using v4f32 = float __attribute__((__vector_size__(16))); void fct6(v4f32 *x) { #pragma clang loop vectorize(enable) for (int i = 0; i < 256; ++i) x[i] = 7 * x[i]; } After compiling it with: clang++ -O3 -march=native -mtune=native \ -Rpass=loop-vectorize,slp-vectorize -Rpass-missed=loop-vectorize,slp-vectorize

Hi, I'm reading "http://llvm.org/docs/Vectorizers.html" and have few question. Hope someone has answers on it. The Loop Vectorizer can vectorize code that becomes a sequence of scalar instructions that scatter/gathers memory. ( http://llvm.org/docs/Vectorizers.html#scatter-gather) int foo(int *A, int *B, int n, int k) { for (int i = 0; i < n; ++i) A[i*7] += B[i*k]; } I

Just looked up the numbers from Agner Fog for Sandy Bridge for vmovaps/etc for loading/storing from memory. vmovaps - load takes 1 load mu op, 3 latency, with a reciprocal throughput of 0.5. vmovaps - store takes 1 store mu op, 1 load mu op for address calculation, 3 latency, with a reciprocal throughput of 1. He does not list vmovsd, but movsd has the same stats as vmovaps, so I feel it is a

Hey Michael, Thanks for the legwork! It appears that the stats you listed are for movaps [SSE], not vmovaps [AVX]. I would *assume* that vmovaps(m128) is closer to vmovaps(m256), since they are both AVX instructions. Although, yes, I agree that this is not clear from Agner's report. Please correct me if I am misunderstanding. As I am sure you are aware, we cannot use SSE (movaps)

We're migrating to LLVM 3.1 and trying to use the upstream FMA patterns. Why is VFMADDSD4 defined with vector types? Is this simply because the gcc intrinsic uses vector types? It's quite unnatural if you have a compiler that generates FMAs as opposed to requiring user intrinsics. -Dave

Ah, bad example. This is a general problem for all (maybe most) SSE and AVX SS/SD patterns though, which is why I mentioned Sandybridge. You can swap out VFMADDSD in my example for VADDSD or whatever you like. I have a lion's share of such a change implemented already and performance is greatly affected. If the community is interested in this change, I would be happy to prepare a patch.

On Wed, Jan 28, 2015 at 4:05 PM, Ahmed Bougacha <ahmed.bougacha at gmail.com> wrote: > Hi Chandler, > > I've been looking at the regressions Quentin mentioned, and filed a PR > for the most egregious one: http://llvm.org/bugs/show_bug.cgi?id=22377 > > As for the others, I'm working on reducing them, but for now, here are > some raw observations, in case any of

In order to vectorize code like this LLVM needs to prove that “A[i*7]” does not wrap in the address space. It fails to do so and so LLVM doesn’t vectorize this loop even if we try to force it. The following loop will be vectorized if we force it: int foo(int * A, int * B, int n, int k) { for (int i = 0; i < 1024; ++i) A[i] += B[i*k]; } So will this loop: int foo(int * restrict A, int

Hey Jan and Dave, It's not obvious, but there is a significant scalar performance issue following the GCC intrinsics. Let's look at the VFMADDSD pattern. We're operating on scalars with undefineds as the remaining vector elements of the operands. This sounds okay, but when one looks closer... vmovsd fp4_+1088(%rip), %xmm3 # fpppp.f:647 vmovaps %xmm3, 18560(%rsp)

I filed a couple more, in case they're actually different issues: - http://llvm.org/bugs/show_bug.cgi?id=22412 - http://llvm.org/bugs/show_bug.cgi?id=22413 And that's pretty much it for internal changes. I'm fine with flipping the switch; Quentin, are you? Also, just to have an idea, do you (or someone else!) plan to tackle these in the near future? -Ahmed On Thu, Jan 29, 2015 at