<html>
<head>
<meta content="text/html; charset=utf-8" http-equiv="Content-Type">
</head>
<body bgcolor="#FFFFFF" text="#000000">
Thank you, guys! I got it to work by setting the CPU features like
you suggested.<br>
Frank<br>
<br>
<div class="moz-cite-prefix">On 06/23/2016 01:09 PM, Craig Topper
wrote:<br>
</div>
<blockquote
cite="mid:CAF7ks-MSZmnMtWipLLMBaD2SxcH5jDnLJ6Tzjb14pGAO=oQjmw@mail.gmail.com"
type="cite">
<div dir="ltr">I think there's a bug in 3.8 where skylake as cpu
implies AVX512. I think later a skylake-avx512(or something
similar) was added and avx512 feature was removed from the
skylake cpu name.
<div><br>
</div>
<div>The right way to fix this is to call getHostCPUFeatures as
well. This will return a StringMap of true and false values
for each feature. Iterate through that and build a string of
"+feature,-feature" based on each feature name its true/false
value. Then pass that string to engineBuilder.setMAttrs.</div>
<div><br>
</div>
<div>This will protect against other issues such as low end
versions of SandyBridge, Haswell, and SkyLake processesors not
supporting AVX at all.</div>
</div>
<div class="gmail_extra"><br>
<div class="gmail_quote">On Thu, Jun 23, 2016 at 10:07 AM, Keno
Fischer <span dir="ltr"><<a moz-do-not-send="true"
href="mailto:kfischer@college.harvard.edu" target="_blank">kfischer@college.harvard.edu</a>></span>
wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0
.8ex;border-left:1px #ccc solid;padding-left:1ex">You likely
haven't set the cpu features correctly. See<br>
llvm::sys::getHostCPUFeatures. E.g. this is what we're doing
in julia:<br>
<a moz-do-not-send="true"
href="https://github.com/JuliaLang/julia/blob/59b253031af87f62e7d70a7d8848cdfd4a84288b/src/codegen.cpp#L5627"
rel="noreferrer" target="_blank">https://github.com/JuliaLang/julia/blob/59b253031af87f62e7d70a7d8848cdfd4a84288b/src/codegen.cpp#L5627</a><br>
<br>
On Thu, Jun 23, 2016 at 1:00 PM, Frank Winter via llvm-dev<br>
<div class="HOEnZb">
<div class="h5"><<a moz-do-not-send="true"
href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a>>
wrote:<br>
><br>
><br>
><br>
> On 06/23/2016 12:56 PM, Craig Topper wrote:<br>
><br>
> Can you check what value "getHostCPUName" returned?<br>
><br>
> getHostCPUName() = skylake<br>
><br>
><br>
> On Thu, Jun 23, 2016 at 9:53 AM, Frank Winter via
llvm-dev <<a moz-do-not-send="true"
href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a>>
wrote:<br>
>><br>
>> With LLVM 3.8 the JIT compiler engine generates
an AVX512 instruction although I target an 'avx2' CPU
(intel Core I7).<br>
>> I just downloaded the most recent 3.8 and still
it happens.<br>
>><br>
>> It happens with this input module:<br>
>><br>
>><br>
>> target datalayout =
"e-m:e-i64:64-f80:128-n8:16:32:64-S128"<br>
>><br>
>> define void @module_cFFEMJ(i64 %lo, i64 %hi,
i64 %myId, i1 %ordered, i64 %start, i32* noalias align
32 %arg0, i32* noalias align 32 %arg1) {<br>
>> entrypoint:<br>
>> %0 = add nsw i64 %lo, %start<br>
>> %1 = add nsw i64 %hi, %start<br>
>> %2 = select i1 %ordered, i64 %0, i64 %lo<br>
>> %3 = select i1 %ordered, i64 %1, i64 %hi<br>
>> %4 = sdiv i64 %2, 4<br>
>> %5 = sdiv i64 %3, 4<br>
>> %6 = bitcast i32* %arg1 to i64*<br>
>> %7 = load i64, i64* %6, align 32<br>
>> %8 = trunc i64 %7 to i32<br>
>> %9 = getelementptr i32, i32* %arg1, i64 1<br>
>> %10 = lshr i64 %7, 32<br>
>> %11 = trunc i64 %10 to i32<br>
>> %12 = getelementptr i32, i32* %arg1, i64 2<br>
>> %13 = bitcast i32* %12 to i64*<br>
>> %14 = load i64, i64* %13, align 8<br>
>> %15 = trunc i64 %14 to i32<br>
>> %16 = getelementptr i32, i32* %arg1, i64 3<br>
>> %17 = lshr i64 %14, 32<br>
>> %18 = trunc i64 %17 to i32<br>
>> br label %L5<br>
>><br>
>> L5:
; preds = %L5, %entrypoint<br>
>> %19 = phi i64 [ %32, %L5 ], [ %4, %entrypoint
]<br>
>> %20 = shl i64 %19, 4<br>
>> %21 = or i64 %20, 4<br>
>> %22 = or i64 %20, 8<br>
>> %23 = or i64 %20, 12<br>
>> %broadcast.splatinsert9 = insertelement <4
x i32> undef, i32 %8, i32 0<br>
>> %broadcast.splat10 = shufflevector <4 x
i32> %broadcast.splatinsert9, <4 x i32> undef,
<4 x i32> zeroinitializer<br>
>> %broadcast.splatinsert11 = insertelement
<4 x i32> undef, i32 %11, i32 0<br>
>> %broadcast.splat12 = shufflevector <4 x
i32> %broadcast.splatinsert11, <4 x i32> undef,
<4 x i32> zeroinitializer<br>
>> %broadcast.splatinsert13 = insertelement
<4 x i32> undef, i32 %15, i32 0<br>
>> %broadcast.splat14 = shufflevector <4 x
i32> %broadcast.splatinsert13, <4 x i32> undef,
<4 x i32> zeroinitializer<br>
>> %broadcast.splatinsert15 = insertelement
<4 x i32> undef, i32 %18, i32 0<br>
>> %broadcast.splat16 = shufflevector <4 x
i32> %broadcast.splatinsert15, <4 x i32> undef,
<4 x i32> zeroinitializer<br>
>> %24 = getelementptr i32, i32* %arg0, i64 %20<br>
>> %25 = bitcast i32* %24 to <4 x i32>*<br>
>> store <4 x i32> %broadcast.splat10,
<4 x i32>* %25, align 16<br>
>> %26 = getelementptr i32, i32* %arg0, i64 %21<br>
>> %27 = bitcast i32* %26 to <4 x i32>*<br>
>> store <4 x i32> %broadcast.splat12,
<4 x i32>* %27, align 16<br>
>> %28 = getelementptr i32, i32* %arg0, i64 %22<br>
>> %29 = bitcast i32* %28 to <4 x i32>*<br>
>> store <4 x i32> %broadcast.splat14,
<4 x i32>* %29, align 16<br>
>> %30 = getelementptr i32, i32* %arg0, i64 %23<br>
>> %31 = bitcast i32* %30 to <4 x i32>*<br>
>> store <4 x i32> %broadcast.splat16,
<4 x i32>* %31, align 16<br>
>> %32 = add nsw i64 %19, 1<br>
>> %33 = icmp slt i64 %32, %5<br>
>> br i1 %33, label %L5, label %L6<br>
>><br>
>> L6:
; preds = %L5<br>
>> ret void<br>
>> }<br>
>><br>
>><br>
>> The following code line show how I call the JIT
compiler. ('Mod' is pointing to the module).<br>
>><br>
>> llvm::EngineBuilder
engineBuilder(std::move(std::unique_ptr<llvm::Module>(Mod)));<br>
>>
engineBuilder.setMCPU(llvm::sys::getHostCPUName());<br>
>>
engineBuilder.setEngineKind(llvm::EngineKind::JIT);<br>
>>
engineBuilder.setOptLevel(llvm::CodeGenOpt::Aggressive);<br>
>> engineBuilder.setErrorStr(&mcjit_error);<br>
>><br>
>> llvm::TargetOptions targetOptions;<br>
>> targetOptions.AllowFPOpFusion =
llvm::FPOpFusion::Fast;<br>
>> engineBuilder.setTargetOptions( targetOptions
);<br>
>><br>
>> TheExecutionEngine = engineBuilder.create();<br>
>><br>
>> targetMachine = engineBuilder.selectTarget();<br>
>> Mod->setDataLayout(
targetMachine->createDataLayout() );<br>
>><br>
>> TheExecutionEngine->finalizeObject(); //
MCJIT<br>
>> fptr_mainFunc_extern =
TheExecutionEngine->getPointerToFunction(
mainFunc_extern );<br>
>><br>
>><br>
>> When calling the function an 'illegal
instruction' is raised.<br>
>> Looking at the assembler reveals an AVX512
instruction which shouldn't be there.<br>
>><br>
>> Assembly:<br>
>> .text<br>
>> .file "module"<br>
>> .globl main<br>
>> .align 16, 0x90<br>
>> .type main,@function<br>
>> main:<br>
>> .cfi_startproc<br>
>> movq 8(%rsp), %r10<br>
>> leaq (%rdi,%r8), %rdx<br>
>> addq %rsi, %r8<br>
>> testb $1, %cl<br>
>> cmoveq %rdi, %rdx<br>
>> cmoveq %rsi, %r8<br>
>> movq %rdx, %rax<br>
>> sarq $63, %rax<br>
>> shrq $62, %rax<br>
>> addq %rdx, %rax<br>
>> sarq $2, %rax<br>
>> movq %r8, %rcx<br>
>> sarq $63, %rcx<br>
>> shrq $62, %rcx<br>
>> addq %r8, %rcx<br>
>> sarq $2, %rcx<br>
>> movq (%r10), %r8<br>
>> movq 8(%r10), %r10<br>
>> movq %r8, %rdi<br>
>> shrq $32, %rdi<br>
>> movq %r10, %rsi<br>
>> shrq $32, %rsi<br>
>> movq %rax, %rdx<br>
>> shlq $6, %rdx<br>
>> leaq 48(%rdx,%r9), %rdx<br>
>> .align 16, 0x90<br>
>> .LBB0_1:<br>
>> vmovd %r8d, %xmm0<br>
>> vpbroadcastd %xmm0, %xmm0<br>
>> vmovd %edi, %xmm1<br>
>> vpbroadcastd %xmm1, %xmm1<br>
>> vmovd %r10d, %xmm2<br>
>> vpbroadcastd %xmm2, %xmm2<br>
>> vmovd %esi, %xmm3<br>
>> vpbroadcastd %xmm3, %xmm3<br>
>> vmovdqa32 %xmm0, -48(%rdx)<br>
>> vmovdqa32 %xmm1, -32(%rdx)<br>
>> vmovdqa32 %xmm2, -16(%rdx)<br>
>> vmovdqa32 %xmm3, (%rdx)<br>
>> addq $1, %rax<br>
>> addq $64, %rdx<br>
>> cmpq %rcx, %rax<br>
>> jl .LBB0_1<br>
>> retq<br>
>> .Lfunc_end0:<br>
>> .size main, .Lfunc_end0-main<br>
>> .cfi_endproc<br>
>><br>
>><br>
>> .section ".note.GNU-stack","",@progbits<br>
>><br>
>> end assembly!<br>
>><br>
>> I am not sure what instruction is the offending
one, but the 'vmovdqa32' looks avx512.<br>
>><br>
>> I wasn't able to reproduce this with 'opt' - it
generates avx2 instructions. And when I force it to use
e.g. avx512f it rejects the CPU type.<br>
>><br>
>> Any ideas?<br>
>><br>
>><br>
>> Frank<br>
>> _______________________________________________<br>
>> LLVM Developers mailing list<br>
>> <a moz-do-not-send="true"
href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a><br>
>> <a moz-do-not-send="true"
href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev"
rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
><br>
><br>
><br>
><br>
> --<br>
> ~Craig<br>
><br>
><br>
><br>
> _______________________________________________<br>
> LLVM Developers mailing list<br>
> <a moz-do-not-send="true"
href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a><br>
> <a moz-do-not-send="true"
href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev"
rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
><br>
</div>
</div>
</blockquote>
</div>
<br>
<br clear="all">
<div><br>
</div>
-- <br>
<div class="gmail_signature" data-smartmail="gmail_signature">~Craig</div>
</div>
</blockquote>
<br>
</body>
</html>