Frank Winter via llvm-dev
2016-Jun-29 19:41 UTC
[llvm-dev] avx512 JIT backend generates wrong code on <4 x float>
Hi!
When compiling the attached module with the JIT engine on an Intel KNL I
see wrong code getting emitted. I attach a complete exploit program
which shows the bug in LLVM 3.8. It loads and JIT compiles the module
and prints the assembler. I stumbled on this since the result of an
actual calculation was wrong. So, it's not only the text version of the
assembler also the machine assembler is wrong.
When I execute the exploit program on an Intel KNL the following output
is produced:
CPU name = knl
-sse4a,-avx512bw,cx16,-tbm,xsave,-fma4,-avx512vl,prfchw,bmi2,adx,-xsavec,fsgsbase,avx,avx512cd,avx512pf,-rtm,popcnt,fma,bmi,aes,rdrnd,-xsaves,sse4.1,sse4.2,avx2,avx512er,sse,lzcnt,pclmul,avx512f,f16c,ssse3,mmx,-pku,cmov,-xop,rdseed,movbe,-hle,xsaveopt,-sha,sse2,sse3,-avx512dq,
Assembly:
.text
.file "module_KFxOBX_i4_after.ll"
.globl adjmul
.align 16, 0x90
.type adjmul, at function
adjmul:
.cfi_startproc
leaq (%rdi,%r8), %rdx
addq %rsi, %r8
testb $1, %cl
cmoveq %rdi, %rdx
cmoveq %rsi, %r8
movq %rdx, %rax
sarq $63, %rax
shrq $62, %rax
addq %rdx, %rax
sarq $2, %rax
movq %r8, %rcx
sarq $63, %rcx
shrq $62, %rcx
addq %r8, %rcx
sarq $2, %rcx
movq %rax, %rdx
shlq $5, %rdx
leaq 16(%r9,%rdx), %rsi
orq $16, %rdx
movq 16(%rsp), %rdi
addq %rdx, %rdi
addq 8(%rsp), %rdx
.align 16, 0x90
.LBB0_1:
vmovaps -16(%rdx), %xmm0
vmovaps (%rdx), %xmm1
vmovaps -16(%rdi), %xmm2
vmovaps (%rdi), %xmm3
vmulps %xmm3, %xmm1, %xmm4
vmulps %xmm2, %xmm1, %xmm1
vfmadd213ss %xmm4, %xmm0, %xmm2
vfmsub213ss %xmm1, %xmm0, %xmm3
vmovaps %xmm2, -16(%rsi)
vmovaps %xmm3, (%rsi)
addq $1, %rax
addq $32, %rsi
addq $32, %rdi
addq $32, %rdx
cmpq %rcx, %rax
jl .LBB0_1
retq
.Lfunc_end0:
.size adjmul, .Lfunc_end0-adjmul
.cfi_endproc
.section ".note.GNU-stack","", at progbits
end assembly!
The instructions 'vfmadd213ss' are 'Fused Multiply-Add of Scalar
Single-Precision Floating-Point'. Those should be SIMD vector
instructions. Note that the KNL has 16 wide float SIMD, while the
exploit module uses only 4. However, the backend should be able to
handle this.
Unless I receive further ideas I will file an official bug report.
Frank
-------------- next part --------------
LLVMPATH=/home/fwinter/toolchain/install/llvm-3.8-recent
LLVMCONF=$(LLVMPATH)/bin/llvm-config
CXXFLAGS=$(shell $(LLVMCONF) --cxxflags)
LIBS=$(shell $(LLVMCONF) --libs core mcjit native scalaropts vectorize irreader
linker) -ldl
LDFLAGS=$(shell $(LLVMCONF) --ldflags)
CXX=g++
OBJS=main.o
TARGET=main
all: $(TARGET)
main: $(OBJS)
$(CXX) -g -o $@ $^ $(LIBS) $(LDFLAGS)
%.o: %.cc
$(CXX) $(CXXFLAGS) -c -g $<
clean:
rm -rf $(TARGET) $(OBJS) *~
-------------- next part --------------
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
define void @adjmul(i64 %lo, i64 %hi, i64 %myId, i1 %ordered, i64 %start, float*
noalias align 64 %arg0, float* noalias align 64 %arg1, float* noalias align 64
%arg2) {
entrypoint:
%0 = add nsw i64 %lo, %start
%1 = add nsw i64 %hi, %start
%2 = select i1 %ordered, i64 %0, i64 %lo
%3 = select i1 %ordered, i64 %1, i64 %hi
%4 = sdiv i64 %2, 4
%5 = sdiv i64 %3, 4
br label %L5
L5: ; preds = %L5, %entrypoint
%6 = phi i64 [ %27, %L5 ], [ %4, %entrypoint ]
%7 = shl i64 %6, 3
%8 = or i64 %7, 4
%9 = getelementptr float, float* %arg1, i64 %7
%10 = bitcast float* %9 to <4 x float>*
%wide.load = load <4 x float>, <4 x float>* %10, align 16
%11 = getelementptr float, float* %arg1, i64 %8
%12 = bitcast float* %11 to <4 x float>*
%wide.load5 = load <4 x float>, <4 x float>* %12, align 16
%13 = getelementptr float, float* %arg2, i64 %7
%14 = bitcast float* %13 to <4 x float>*
%wide.load6 = load <4 x float>, <4 x float>* %14, align 16
%15 = getelementptr float, float* %arg2, i64 %8
%16 = bitcast float* %15 to <4 x float>*
%wide.load7 = load <4 x float>, <4 x float>* %16, align 16
%17 = fmul <4 x float> %wide.load, %wide.load6
%18 = fmul <4 x float> %wide.load5, %wide.load7
%19 = fadd <4 x float> %17, %18
%20 = fmul <4 x float> %wide.load, %wide.load7
%21 = fmul <4 x float> %wide.load5, %wide.load6
%22 = fsub <4 x float> %20, %21
%23 = getelementptr float, float* %arg0, i64 %7
%24 = bitcast float* %23 to <4 x float>*
store <4 x float> %19, <4 x float>* %24, align 16
%25 = getelementptr float, float* %arg0, i64 %8
%26 = bitcast float* %25 to <4 x float>*
store <4 x float> %22, <4 x float>* %26, align 16
%27 = add nsw i64 %6, 1
%28 = icmp slt i64 %27, %5
br i1 %28, label %L5, label %L6
L6: ; preds = %L5
ret void
}
-------------- next part --------------
A non-text attachment was scrubbed...
Name: main.cc
Type: text/x-c++src
Size: 5059 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160629/8008aa24/attachment.cc>
Hal Finkel via llvm-dev
2016-Jun-29 19:48 UTC
[llvm-dev] 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 x float> > > Hi! > > When compiling the attached module with the JIT engine on an Intel > KNL I > see wrong code getting emitted. I attach a complete exploit program > which shows the bug in LLVM 3.8. It loads and JIT compiles the module > and prints the assembler. I stumbled on this since the result of an > actual calculation was wrong. So, it's not only the text version of > the > assembler also the machine assembler is wrong. > > When I execute the exploit program on an Intel KNL the following > output > is produced: > > CPU name = knl > -sse4a,-avx512bw,cx16,-tbm,xsave,-fma4,-avx512vl,prfchw,bmi2,adx,-xsavec,fsgsbase,avx,avx512cd,avx512pf,-rtm,popcnt,fma,bmi,aes,rdrnd,-xsaves,sse4.1,sse4.2,avx2,avx512er,sse,lzcnt,pclmul,avx512f,f16c,ssse3,mmx,-pku,cmov,-xop,rdseed,movbe,-hle,xsaveopt,-sha,sse2,sse3,-avx512dq, > Assembly: > .text > .file "module_KFxOBX_i4_after.ll" > .globl adjmul > .align 16, 0x90 > .type adjmul, at function > adjmul: > .cfi_startproc > leaq (%rdi,%r8), %rdx > addq %rsi, %r8 > testb $1, %cl > cmoveq %rdi, %rdx > cmoveq %rsi, %r8 > movq %rdx, %rax > sarq $63, %rax > shrq $62, %rax > addq %rdx, %rax > sarq $2, %rax > movq %r8, %rcx > sarq $63, %rcx > shrq $62, %rcx > addq %r8, %rcx > sarq $2, %rcx > movq %rax, %rdx > shlq $5, %rdx > leaq 16(%r9,%rdx), %rsi > orq $16, %rdx > movq 16(%rsp), %rdi > addq %rdx, %rdi > addq 8(%rsp), %rdx > .align 16, 0x90 > .LBB0_1: > vmovaps -16(%rdx), %xmm0 > vmovaps (%rdx), %xmm1 > vmovaps -16(%rdi), %xmm2 > vmovaps (%rdi), %xmm3 > vmulps %xmm3, %xmm1, %xmm4 > vmulps %xmm2, %xmm1, %xmm1 > vfmadd213ss %xmm4, %xmm0, %xmm2 > vfmsub213ss %xmm1, %xmm0, %xmm3 > vmovaps %xmm2, -16(%rsi) > vmovaps %xmm3, (%rsi) > addq $1, %rax > addq $32, %rsi > addq $32, %rdi > addq $32, %rdx > cmpq %rcx, %rax > jl .LBB0_1 > retq > .Lfunc_end0: > .size adjmul, .Lfunc_end0-adjmul > .cfi_endproc > > > .section ".note.GNU-stack","", at progbits > > end assembly! > > > The instructions 'vfmadd213ss' are 'Fused Multiply-Add of Scalar > Single-Precision Floating-Point'. Those should be SIMD vector > instructions. Note that the KNL has 16 wide float SIMD, while the > exploit module uses only 4. However, the backend should be able to > handle this. > > Unless I receive further ideas I will file an official bug report. > > Frank > > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >-- Hal Finkel Assistant Computational Scientist Leadership Computing Facility Argonne National Laboratory
Frank Winter via llvm-dev
2016-Jun-30 16:49 UTC
[llvm-dev] avx512 JIT backend generates wrong code on <4 x float>
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 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 x float> >> >> Hi! >> >> When compiling the attached module with the JIT engine on an Intel >> KNL I >> see wrong code getting emitted. I attach a complete exploit program >> which shows the bug in LLVM 3.8. It loads and JIT compiles the module >> and prints the assembler. I stumbled on this since the result of an >> actual calculation was wrong. So, it's not only the text version of >> the >> assembler also the machine assembler is wrong. >> >> When I execute the exploit program on an Intel KNL the following >> output >> is produced: >> >> CPU name = knl >> -sse4a,-avx512bw,cx16,-tbm,xsave,-fma4,-avx512vl,prfchw,bmi2,adx,-xsavec,fsgsbase,avx,avx512cd,avx512pf,-rtm,popcnt,fma,bmi,aes,rdrnd,-xsaves,sse4.1,sse4.2,avx2,avx512er,sse,lzcnt,pclmul,avx512f,f16c,ssse3,mmx,-pku,cmov,-xop,rdseed,movbe,-hle,xsaveopt,-sha,sse2,sse3,-avx512dq, >> Assembly: >> .text >> .file "module_KFxOBX_i4_after.ll" >> .globl adjmul >> .align 16, 0x90 >> .type adjmul, at function >> adjmul: >> .cfi_startproc >> leaq (%rdi,%r8), %rdx >> addq %rsi, %r8 >> testb $1, %cl >> cmoveq %rdi, %rdx >> cmoveq %rsi, %r8 >> movq %rdx, %rax >> sarq $63, %rax >> shrq $62, %rax >> addq %rdx, %rax >> sarq $2, %rax >> movq %r8, %rcx >> sarq $63, %rcx >> shrq $62, %rcx >> addq %r8, %rcx >> sarq $2, %rcx >> movq %rax, %rdx >> shlq $5, %rdx >> leaq 16(%r9,%rdx), %rsi >> orq $16, %rdx >> movq 16(%rsp), %rdi >> addq %rdx, %rdi >> addq 8(%rsp), %rdx >> .align 16, 0x90 >> .LBB0_1: >> vmovaps -16(%rdx), %xmm0 >> vmovaps (%rdx), %xmm1 >> vmovaps -16(%rdi), %xmm2 >> vmovaps (%rdi), %xmm3 >> vmulps %xmm3, %xmm1, %xmm4 >> vmulps %xmm2, %xmm1, %xmm1 >> vfmadd213ss %xmm4, %xmm0, %xmm2 >> vfmsub213ss %xmm1, %xmm0, %xmm3 >> vmovaps %xmm2, -16(%rsi) >> vmovaps %xmm3, (%rsi) >> addq $1, %rax >> addq $32, %rsi >> addq $32, %rdi >> addq $32, %rdx >> cmpq %rcx, %rax >> jl .LBB0_1 >> retq >> .Lfunc_end0: >> .size adjmul, .Lfunc_end0-adjmul >> .cfi_endproc >> >> >> .section ".note.GNU-stack","", at progbits >> >> end assembly! >> >> >> The instructions 'vfmadd213ss' are 'Fused Multiply-Add of Scalar >> Single-Precision Floating-Point'. Those should be SIMD vector >> instructions. Note that the KNL has 16 wide float SIMD, while the >> exploit module uses only 4. However, the backend should be able to >> handle this. >> >> Unless I receive further ideas I will file an official bug report. >> >> Frank >> >> >> _______________________________________________ >> LLVM Developers mailing list >> llvm-dev at lists.llvm.org >> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >>
Reasonably Related Threads
- avx512 JIT backend generates wrong code on <4 x float>
- avx512 JIT backend generates wrong code on <4 x float>
- AVX512 instruction generated when JIT compiling for an avx2 architecture
- AVX512 instruction generated when JIT compiling for an avx2 architecture
- [LLVMdev] SIMD for sdiv <2 x i64>