I think this is because of float fRec0[2]; which the LLVM optimizer tries to optimize into loads/stores of an i64 (one load/store instead of 2). Then it gets split up into load/stores of i32 plus reinterpreting. I've seen this in box2d as well, which has Vec2 with an array of two floats.
We should get rid of the reinterpret, even if has very low runtime cost, it wastes space in the binary. It's possible that LLVM codegen opts we don't run would optimize that further, so that's one option. Another is to do this in the binaryen optimizer, I'll look into that. It might make a difference to replace the array of 2 floats with float x, float y; On Thu, Oct 19, 2017 at 5:44 AM, <[email protected]> wrote: > Hi, > > Our Faust DSP audio language (faust.grame.fr) can generate C,C++ ... or > wast/wasm code. On a simple example (signal integrator), we are comparing > the wasm code directly generated from our wast/wasm backend, to the > C backed later on compiled/optimized in wasm with EMCC. Here are the > results: > > 1) in C we get : > > #ifndef FAUSTFLOAT > #define FAUSTFLOAT float > #endif > > typedef struct { > > float fRec0[2]; > int fSamplingFreq; > > } mydsp; > > void computemydsp(mydsp* dsp, int count, FAUSTFLOAT** inputs, FAUSTFLOAT** > outputs) { > FAUSTFLOAT* input0 = inputs[0]; > FAUSTFLOAT* output0 = outputs[0]; > /* C99 loop */ > { > int i; > for (i = 0; (i < count); i = (i + 1)) { > dsp->fRec0[0] = ((float)input0[i] + dsp->fRec0[1]); > output0[i] = (FAUSTFLOAT)dsp->fRec0[0]; > dsp->fRec0[1] = dsp->fRec0[0]; > > } > } > > } > > 2) in wast from Faust, we get: > > (func $compute (param $dsp i32) (param $count i32) (param $inputs i32) > (param $outputs i32) > (local $input0 i32) > (local $output0 i32) > (local $i i32) > (set_local $input0 (i32.const 0)) > (set_local $output0 (i32.const 0)) > (set_local $i (i32.const 0)) > (set_local $input0 (i32.load (i32.add (get_local $inputs) (i32.shl > (i32.const 0) (i32.const 2))))) > (set_local $output0 (i32.load (i32.add (get_local $outputs) (i32.shl > (i32.const 0) (i32.const 2))))) > (set_local $i (i32.const 0)) > (loop $for-in-i > (block $for-out-i > (f32.store (i32.const 0) (f32.add (f32.load (i32.add (get_local $input0) > (i32.shl (get_local $i) (i32.const 2)))) (f32.load (i32.const 4)))) > (f32.store (i32.add (get_local $output0) (i32.shl (get_local $i) > (i32.const 2))) (f32.load (i32.const 0))) > (f32.store (i32.const 4) (f32.load (i32.const 0))) > (set_local $i (i32.add (get_local $i) (i32.const 1))) > (if (i32.lt_s (get_local $i) (get_local $count)) (br $for-in-i) (br > $for-out-i)) > ) > ) > ) > > > ==> so we get some f32.load/f32.store in the object fRec0 field that could > be possibly optimized using an intermediate local variable. > > 3) With emcc -O1 (emcc testWASM.cpp -O2 -s WASM=1 -s SIDE_MODULE=1 -o > testWASM.wasm) > > (func $0 (type $0) (param $var$0 i32) (param $var$1 i32) (param $var$2 > i32) (param $var$3 i32) > (local $var$4 i32) > (local $var$5 i32) > (local $var$6 f32) > (block $label$1 > (set_local $var$4 > (i32.load > (get_local $var$2) > ) > ) > (set_local $var$5 > (i32.load > (get_local $var$3) > ) > ) > (if > (i32.le_s > (get_local $var$1) > (i32.const 0) > ) > (return) > ) > (set_local $var$3 > (i32.add > (get_local $var$0) > (i32.const 4) > ) > ) > (set_local $var$2 > (i32.const 0) > ) > (loop $label$3 > (f32.store > (get_local $var$0) > (tee_local $var$6 > (f32.add > (f32.load > (i32.add > (get_local $var$4) > (i32.shl > (get_local $var$2) > (i32.const 2) > ) > ) > ) > (f32.load > (get_local $var$3) > ) > ) > ) > ) > (f32.store > (i32.add > (get_local $var$5) > (i32.shl > (get_local $var$2) > (i32.const 2) > ) > ) > (get_local $var$6) > ) > (i32.store > (get_local $var$3) > (i32.load > (get_local $var$0) > ) > ) > (br_if $label$3 > (i32.ne > (tee_local $var$2 > (i32.add > (get_local $var$2) > (i32.const 1) > ) > ) > (get_local $var$1) > ) > ) > ) > ) > ) > > 4) With emcc -O3 > > (func $0 (type $0) (param $var$0 i32) (param $var$1 i32) (param $var$2 > i32) (param $var$3 i32) > (local $var$4 i32) > (local $var$5 i32) > (local $var$6 i32) > (local $var$7 f32) > (block $label$1 > (set_local $var$4 > (i32.load > (get_local $var$2) > ) > ) > (set_local $var$3 > (i32.load > (get_local $var$3) > ) > ) > (if > (i32.le_s > (get_local $var$1) > (i32.const 0) > ) > (return) > ) > (set_local $var$2 > (i32.const 0) > ) > (set_local $var$7 > (f32.load > (tee_local $var$5 > (i32.add > (get_local $var$0) > (i32.const 4) > ) > ) > ) > ) > (loop $label$3 > (f32.store > (get_local $var$0) > (tee_local $var$7 > (f32.add > (f32.load > (i32.add > (get_local $var$4) > (i32.shl > (get_local $var$2) > (i32.const 2) > ) > ) > ) > (get_local $var$7) > ) > ) > ) > (f32.store > (i32.add > (get_local $var$3) > (i32.shl > (get_local $var$2) > (i32.const 2) > ) > ) > (get_local $var$7) > ) > (i32.store > (get_local $var$5) > (tee_local $var$6 > (i32.load > (get_local $var$0) > ) > ) > ) > (set_local $var$7 > (f32.reinterpret/i32 > (get_local $var$6) > ) > ) > (br_if $label$3 > (i32.ne > (tee_local $var$2 > (i32.add > (get_local $var$2) > (i32.const 1) > ) > ) > (get_local $var$1) > ) > ) > ) > ) > ) > > The Emcc optimized code seems rather complicated: one of the f32.store > is transformed in a i32.store at -O1 level. In -O3 mode, the i32 $var$6 > intermediate variable is even "reinterpreted" as a f32 type with the > f32.reinterpret/i32 > instruction. Is this supposed to be faster at the end ? > > Thanks for any good explanation. > > -- > You received this message because you are subscribed to the Google Groups > "emscripten-discuss" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > For more options, visit https://groups.google.com/d/optout. > -- You received this message because you are subscribed to the Google Groups "emscripten-discuss" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
