MCJIT when lowering to x86-64 generates a MOVAPS (Move Aligned Packed
Single-Precision Floating-Point Values) on a non-aligned memory address:
movaps 88(%rdx), %xmm0
where %rdx comes in as a function argument with only natural alignment
(float*). This x86 instruction requires the memory address to be 16 byte
aligned which 88 plus something aligned to 4 byte isn't.
Here the according IR code which was produced from the SLP vectorizer:
define void @func(float* noalias %arg0, float* noalias %arg1, float*
noalias %arg2) {
entrypoint:
...
%104 = getelementptr float* %arg0, i32 22
...
%204 = bitcast float* %104 to <4 x float>*
store <4 x float> %198, <4 x float>* %204
This in itself not wrong. However, shouldn't the lowering pass recognize
the wrong alignment?
I am using LLVM 3.4.2 as available as source code from llvm.org.
Thanks,
Frank
Arnold Schwaighofer
2014-Aug-07 20:29 UTC
[LLVMdev] MCJIT generates MOVAPS on unaligned address
> On Aug 7, 2014, at 12:42 PM, Frank Winter <fwinter at jlab.org> wrote: > > MCJIT when lowering to x86-64 generates a MOVAPS (Move Aligned Packed Single-Precision Floating-Point Values) on a non-aligned memory address: > > movaps 88(%rdx), %xmm0 > > where %rdx comes in as a function argument with only natural alignment (float*). This x86 instruction requires the memory address to be 16 byte aligned which 88 plus something aligned to 4 byte isn't. > > Here the according IR code which was produced from the SLP vectorizer: > > define void @func(float* noalias %arg0, float* noalias %arg1, float* noalias %arg2) { > entrypoint: > ... > %104 = getelementptr float* %arg0, i32 22 > ... > %204 = bitcast float* %104 to <4 x float>* > store <4 x float> %198, <4 x float>* %204 > > This in itself not wrong. However, shouldn't the lowering pass recognize the wrong alignment?The LLVM IR is wrong. Omitting the align directive on the store means abi alignment of the target. The backend is “right” wrt to LLVM IR semantics to produce the movaps. The error is in the producer (looks like the SLP vectorizer) of said vector store. Could you provide a full test case where running the SLP vectorizer (opt -slp-vectorize < t.ll) produces such an output? The following code in the SLP vectorizer should have made sure that we created an alignment of “4 bytes” given a data layout (http://llvm.org/docs/LangRef.html#data-layout) that specifies f32:32:32. case Instruction::Store: { StoreInst *SI = cast<StoreInst>(VL0); unsigned Alignment = SI->getAlignment(); ... StoreInst *S = Builder.CreateStore(VecValue, VecPtr); if (!Alignment) Alignment = DL->getABITypeAlignment(SI->getPointerOperand()->getType()); // << Get the 4byte alignment for the scalar float store from the data layout string. S->setAlignment(Alignment);
Makes sense. I overlooked the missing alignment tag in the IR. Attached the full test case. Frank On 08/07/2014 04:29 PM, Arnold Schwaighofer wrote:>> On Aug 7, 2014, at 12:42 PM, Frank Winter <fwinter at jlab.org> wrote: >> >> MCJIT when lowering to x86-64 generates a MOVAPS (Move Aligned Packed Single-Precision Floating-Point Values) on a non-aligned memory address: >> >> movaps 88(%rdx), %xmm0 >> >> where %rdx comes in as a function argument with only natural alignment (float*). This x86 instruction requires the memory address to be 16 byte aligned which 88 plus something aligned to 4 byte isn't. >> >> Here the according IR code which was produced from the SLP vectorizer: >> >> define void @func(float* noalias %arg0, float* noalias %arg1, float* noalias %arg2) { >> entrypoint: >> ... >> %104 = getelementptr float* %arg0, i32 22 >> ... >> %204 = bitcast float* %104 to <4 x float>* >> store <4 x float> %198, <4 x float>* %204 >> >> This in itself not wrong. However, shouldn't the lowering pass recognize the wrong alignment? > The LLVM IR is wrong. Omitting the align directive on the store means abi alignment of the target. The backend is “right” wrt to LLVM IR semantics to produce the movaps. > > The error is in the producer (looks like the SLP vectorizer) of said vector store. Could you provide a full test case where running the SLP vectorizer (opt -slp-vectorize < t.ll) produces such an output? > > The following code in the SLP vectorizer should have made sure that we created an alignment of “4 bytes” given a data layout (http://llvm.org/docs/LangRef.html#data-layout) that specifies f32:32:32. > > case Instruction::Store: { > StoreInst *SI = cast<StoreInst>(VL0); > unsigned Alignment = SI->getAlignment(); > ... > StoreInst *S = Builder.CreateStore(VecValue, VecPtr); > if (!Alignment) > Alignment = DL->getABITypeAlignment(SI->getPointerOperand()->getType()); // << Get the 4byte alignment for the scalar float store from the data layout string. > S->setAlignment(Alignment);-------------- next part -------------- ;; ModuleID = 'module' target triple = "x86_64-unknown-linux-gnu" declare float @sinf(float) declare float @acosf(float) declare float @asinf(float) declare float @atanf(float) declare float @ceilf(float) declare float @floorf(float) declare float @cosf(float) declare float @coshf(float) declare float @expf(float) declare float @logf(float) declare float @log10f(float) declare float @sinhf(float) declare float @tanf(float) declare float @tanhf(float) declare float @fabsf(float) declare float @sqrtf(float) declare float @powf(float, float) declare float @atan2f(float, float) declare double @sin(double) declare double @acos(double) declare double @asin(double) declare double @atan(double) declare double @ceil(double) declare double @floor(double) declare double @cos(double) declare double @cosh(double) declare double @exp(double) declare double @log(double) declare double @log10(double) declare double @sinh(double) declare double @tan(double) declare double @tanh(double) declare double @fabs(double) declare double @sqrt(double) declare double @pow(double, double) declare double @atan2(double, double) define void @main(float* noalias %arg0, float* noalias %arg1, float* noalias %arg2) { entrypoint: %0 = getelementptr float* %arg1, i32 0 %1 = load float* %0 %2 = getelementptr float* %arg1, i32 4 %3 = load float* %2 %4 = getelementptr float* %arg1, i32 8 %5 = load float* %4 %6 = getelementptr float* %arg1, i32 12 %7 = load float* %6 %8 = getelementptr float* %arg1, i32 16 %9 = load float* %8 %10 = getelementptr float* %arg1, i32 20 %11 = load float* %10 %12 = getelementptr float* %arg2, i32 0 %13 = load float* %12 %14 = getelementptr float* %arg2, i32 4 %15 = load float* %14 %16 = getelementptr float* %arg2, i32 8 %17 = load float* %16 %18 = getelementptr float* %arg2, i32 12 %19 = load float* %18 %20 = getelementptr float* %arg2, i32 16 %21 = load float* %20 %22 = getelementptr float* %arg2, i32 20 %23 = load float* %22 %24 = fadd float %15, %3 %25 = fadd float %13, %1 %26 = fadd float %19, %7 %27 = fadd float %17, %5 %28 = fadd float %23, %11 %29 = fadd float %21, %9 %30 = getelementptr float* %arg0, i32 0 store float %25, float* %30 %31 = getelementptr float* %arg0, i32 4 store float %24, float* %31 %32 = getelementptr float* %arg0, i32 8 store float %27, float* %32 %33 = getelementptr float* %arg0, i32 12 store float %26, float* %33 %34 = getelementptr float* %arg0, i32 16 store float %29, float* %34 %35 = getelementptr float* %arg0, i32 20 store float %28, float* %35 %36 = getelementptr float* %arg1, i32 1 %37 = load float* %36 %38 = getelementptr float* %arg1, i32 5 %39 = load float* %38 %40 = getelementptr float* %arg1, i32 9 %41 = load float* %40 %42 = getelementptr float* %arg1, i32 13 %43 = load float* %42 %44 = getelementptr float* %arg1, i32 17 %45 = load float* %44 %46 = getelementptr float* %arg1, i32 21 %47 = load float* %46 %48 = getelementptr float* %arg2, i32 1 %49 = load float* %48 %50 = getelementptr float* %arg2, i32 5 %51 = load float* %50 %52 = getelementptr float* %arg2, i32 9 %53 = load float* %52 %54 = getelementptr float* %arg2, i32 13 %55 = load float* %54 %56 = getelementptr float* %arg2, i32 17 %57 = load float* %56 %58 = getelementptr float* %arg2, i32 21 %59 = load float* %58 %60 = fadd float %51, %39 %61 = fadd float %49, %37 %62 = fadd float %55, %43 %63 = fadd float %53, %41 %64 = fadd float %59, %47 %65 = fadd float %57, %45 %66 = getelementptr float* %arg0, i32 1 store float %61, float* %66 %67 = getelementptr float* %arg0, i32 5 store float %60, float* %67 %68 = getelementptr float* %arg0, i32 9 store float %63, float* %68 %69 = getelementptr float* %arg0, i32 13 store float %62, float* %69 %70 = getelementptr float* %arg0, i32 17 store float %65, float* %70 %71 = getelementptr float* %arg0, i32 21 store float %64, float* %71 %72 = getelementptr float* %arg1, i32 2 %73 = load float* %72 %74 = getelementptr float* %arg1, i32 6 %75 = load float* %74 %76 = getelementptr float* %arg1, i32 10 %77 = load float* %76 %78 = getelementptr float* %arg1, i32 14 %79 = load float* %78 %80 = getelementptr float* %arg1, i32 18 %81 = load float* %80 %82 = getelementptr float* %arg1, i32 22 %83 = getelementptr float* %arg2, i32 2 %84 = load float* %83 %85 = getelementptr float* %arg2, i32 6 %86 = load float* %85 %87 = getelementptr float* %arg2, i32 10 %88 = load float* %87 %89 = getelementptr float* %arg2, i32 14 %90 = load float* %89 %91 = getelementptr float* %arg2, i32 18 %92 = load float* %91 %93 = getelementptr float* %arg2, i32 22 %94 = fadd float %86, %75 %95 = fadd float %84, %73 %96 = fadd float %90, %79 %97 = fadd float %88, %77 %98 = fadd float %92, %81 %99 = getelementptr float* %arg0, i32 2 store float %95, float* %99 %100 = getelementptr float* %arg0, i32 6 store float %94, float* %100 %101 = getelementptr float* %arg0, i32 10 store float %97, float* %101 %102 = getelementptr float* %arg0, i32 14 store float %96, float* %102 %103 = getelementptr float* %arg0, i32 18 store float %98, float* %103 %104 = getelementptr float* %arg0, i32 22 %105 = getelementptr float* %arg1, i32 3 %106 = load float* %105 %107 = getelementptr float* %arg1, i32 7 %108 = load float* %107 %109 = getelementptr float* %arg1, i32 11 %110 = load float* %109 %111 = getelementptr float* %arg1, i32 15 %112 = load float* %111 %113 = getelementptr float* %arg1, i32 19 %114 = load float* %113 %115 = getelementptr float* %arg1, i32 23 %116 = getelementptr float* %arg2, i32 3 %117 = load float* %116 %118 = getelementptr float* %arg2, i32 7 %119 = load float* %118 %120 = getelementptr float* %arg2, i32 11 %121 = load float* %120 %122 = getelementptr float* %arg2, i32 15 %123 = load float* %122 %124 = getelementptr float* %arg2, i32 19 %125 = load float* %124 %126 = getelementptr float* %arg2, i32 23 %127 = fadd float %119, %108 %128 = fadd float %117, %106 %129 = fadd float %123, %112 %130 = fadd float %121, %110 %131 = fadd float %125, %114 %132 = getelementptr float* %arg0, i32 3 store float %128, float* %132 %133 = getelementptr float* %arg0, i32 7 store float %127, float* %133 %134 = getelementptr float* %arg0, i32 11 store float %130, float* %134 %135 = getelementptr float* %arg0, i32 15 store float %129, float* %135 %136 = getelementptr float* %arg0, i32 19 store float %131, float* %136 %137 = getelementptr float* %arg0, i32 23 %138 = getelementptr float* %arg1, i32 24 %139 = getelementptr float* %arg1, i32 28 %140 = load float* %139 %141 = getelementptr float* %arg1, i32 32 %142 = load float* %141 %143 = getelementptr float* %arg1, i32 36 %144 = load float* %143 %145 = getelementptr float* %arg1, i32 40 %146 = load float* %145 %147 = getelementptr float* %arg1, i32 44 %148 = load float* %147 %149 = getelementptr float* %arg2, i32 24 %150 = getelementptr float* %arg2, i32 28 %151 = load float* %150 %152 = getelementptr float* %arg2, i32 32 %153 = load float* %152 %154 = getelementptr float* %arg2, i32 36 %155 = load float* %154 %156 = getelementptr float* %arg2, i32 40 %157 = load float* %156 %158 = getelementptr float* %arg2, i32 44 %159 = load float* %158 %160 = fadd float %151, %140 %161 = fadd float %155, %144 %162 = fadd float %153, %142 %163 = fadd float %159, %148 %164 = fadd float %157, %146 %165 = getelementptr float* %arg0, i32 24 %166 = getelementptr float* %arg0, i32 28 store float %160, float* %166 %167 = getelementptr float* %arg0, i32 32 store float %162, float* %167 %168 = getelementptr float* %arg0, i32 36 store float %161, float* %168 %169 = getelementptr float* %arg0, i32 40 store float %164, float* %169 %170 = getelementptr float* %arg0, i32 44 store float %163, float* %170 %171 = getelementptr float* %arg1, i32 25 %172 = bitcast float* %82 to <4 x float>* %173 = load <4 x float>* %172 %174 = getelementptr float* %arg1, i32 29 %175 = load float* %174 %176 = getelementptr float* %arg1, i32 33 %177 = load float* %176 %178 = getelementptr float* %arg1, i32 37 %179 = load float* %178 %180 = getelementptr float* %arg1, i32 41 %181 = load float* %180 %182 = getelementptr float* %arg1, i32 45 %183 = load float* %182 %184 = getelementptr float* %arg2, i32 25 %185 = bitcast float* %93 to <4 x float>* %186 = load <4 x float>* %185 %187 = getelementptr float* %arg2, i32 29 %188 = load float* %187 %189 = getelementptr float* %arg2, i32 33 %190 = load float* %189 %191 = getelementptr float* %arg2, i32 37 %192 = load float* %191 %193 = getelementptr float* %arg2, i32 41 %194 = load float* %193 %195 = getelementptr float* %arg2, i32 45 %196 = load float* %195 %197 = fadd float %188, %175 %198 = fadd <4 x float> %186, %173 %199 = fadd float %192, %179 %200 = fadd float %190, %177 %201 = fadd float %196, %183 %202 = fadd float %194, %181 %203 = getelementptr float* %arg0, i32 25 %204 = bitcast float* %104 to <4 x float>* store <4 x float> %198, <4 x float>* %204 %205 = getelementptr float* %arg0, i32 29 store float %197, float* %205 %206 = getelementptr float* %arg0, i32 33 store float %200, float* %206 %207 = getelementptr float* %arg0, i32 37 store float %199, float* %207 %208 = getelementptr float* %arg0, i32 41 store float %202, float* %208 %209 = getelementptr float* %arg0, i32 45 store float %201, float* %209 ret void } define void @main_extern([8 x i8]* %arg_ptr) { entrypoint: %0 = getelementptr [8 x i8]* %arg_ptr, i32 0 %1 = bitcast [8 x i8]* %0 to float** %2 = load float** %1 %3 = getelementptr [8 x i8]* %arg_ptr, i32 1 %4 = bitcast [8 x i8]* %3 to float** %5 = load float** %4 %6 = getelementptr [8 x i8]* %arg_ptr, i32 2 %7 = bitcast [8 x i8]* %6 to float** %8 = load float** %7 call void @main(float* %2, float* %5, float* %8) ret void }
It's not reproducible with 'opt'. I call the SLP pass from my application and only then the wrong IR gets generated. On the attached module I call via the function pass manager: 1) TargetLibraryInfo with the target triple 2) Set the data layout 3) Basic Alias Analysis 4) SLP vectorizer This produces the wrong IR. On the other hand running the attached module through 'opt -slp-vectorizer' results in no code changes. What could I be missing here? Frank On 08/07/2014 04:29 PM, Arnold Schwaighofer wrote:>> On Aug 7, 2014, at 12:42 PM, Frank Winter <fwinter at jlab.org> wrote: >> >> MCJIT when lowering to x86-64 generates a MOVAPS (Move Aligned Packed Single-Precision Floating-Point Values) on a non-aligned memory address: >> >> movaps 88(%rdx), %xmm0 >> >> where %rdx comes in as a function argument with only natural alignment (float*). This x86 instruction requires the memory address to be 16 byte aligned which 88 plus something aligned to 4 byte isn't. >> >> Here the according IR code which was produced from the SLP vectorizer: >> >> define void @func(float* noalias %arg0, float* noalias %arg1, float* noalias %arg2) { >> entrypoint: >> ... >> %104 = getelementptr float* %arg0, i32 22 >> ... >> %204 = bitcast float* %104 to <4 x float>* >> store <4 x float> %198, <4 x float>* %204 >> >> This in itself not wrong. However, shouldn't the lowering pass recognize the wrong alignment? > The LLVM IR is wrong. Omitting the align directive on the store means abi alignment of the target. The backend is “right” wrt to LLVM IR semantics to produce the movaps. > > The error is in the producer (looks like the SLP vectorizer) of said vector store. Could you provide a full test case where running the SLP vectorizer (opt -slp-vectorize < t.ll) produces such an output? > > The following code in the SLP vectorizer should have made sure that we created an alignment of “4 bytes” given a data layout (http://llvm.org/docs/LangRef.html#data-layout) that specifies f32:32:32. > > case Instruction::Store: { > StoreInst *SI = cast<StoreInst>(VL0); > unsigned Alignment = SI->getAlignment(); > ... > StoreInst *S = Builder.CreateStore(VecValue, VecPtr); > if (!Alignment) > Alignment = DL->getABITypeAlignment(SI->getPointerOperand()->getType()); // << Get the 4byte alignment for the scalar float store from the data layout string. > S->setAlignment(Alignment);-------------- next part -------------- A non-text attachment was scrubbed... Name: module_H7ktW0.ll.gz Type: application/x-gzip Size: 65781 bytes Desc: not available URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140807/a42fe088/attachment.bin>