wasm output contains compiler_rt exports

[daurnimator]

foo.c:

int foo(void) {
    return 42;
}

$ zig build-lib -target wasm32-freestanding --c-source foo.c
$ wasm-dis foo.wasm  | grep '(func' | tail
 (func $__udivmodti4 (; 265 ;) (type $50) (param $0 i32) (param $1 i64) (param $2 i64) (param $3 i64) (param $4 i64) (param $5 i32)
 (func $__umodti3 (; 266 ;) (type $28) (param $0 i32) (param $1 i64) (param $2 i64) (param $3 i64) (param $4 i64)
 (func $__muloti4 (; 267 ;) (type $50) (param $0 i32) (param $1 i64) (param $2 i64) (param $3 i64) (param $4 i64) (param $5 i32)
 (func $__mulodi4 (; 268 ;) (type $23) (param $0 i64) (param $1 i64) (param $2 i32) (result i64)
 (func $compiler_rt.mulXf3.S.loWord (; 269 ;) (type $35) (param $0 i64) (result i64)
 (func $compiler_rt.mulXf3.S.hiWord (; 270 ;) (type $35) (param $0 i64) (result i64)
 (func $compiler_rt.mulXf3.S.Word_1 (; 271 ;) (type $48) (param $0 i64) (param $1 i64) (result i64)
 (func $compiler_rt.mulXf3.S.Word_2 (; 272 ;) (type $48) (param $0 i64) (param $1 i64) (result i64)
 (func $compiler_rt.mulXf3.S.Word_3 (; 273 ;) (type $48) (param $0 i64) (param $1 i64) (result i64)
 (func $compiler_rt.mulXf3.S.Word_4 (; 274 ;) (type $48) (param $0 i64) (param $1 i64) (result i64)

Related #2062 #2290 #2369

[andrewrk]

That's correct. Unused functions are not garbage collected until linking. Linking happens with build_exe.

[daurnimator]

That's correct. Unused functions are not garbage collected until linking. Linking happens with build_exe.

Ah! so I should be using build-exe instead?
That results in an output file foo (rather than foo.wasm), but it seems to have the right exports:

$ wasm-dis foo | grep '(func'
 (type $0 (func))
 (type $1 (func (result i32)))
 (export "__wasm_call_ctors" (func $__wasm_call_ctors))
 (export "foo" (func $foo))
 (func $__wasm_call_ctors (; 0 ;) (type $0)
 (func $foo (; 1 ;) (type $1) (result i32)

I guess that __wasm_call_ctors function is there because https://reviews.llvm.org/D58806 isn't in LLVM 8?

[andrewrk]

I'm going to leave this open to investigate because my memory is fuzzy about how this is supposed to work.

[daurnimator]

From #2499 (thanks @fengb)

WebAssembly: build-exe is for executables which have a main(). build-lib is for building libraries of functions to use from, for example, a web browser environment.

I think this conflicts with your statement above?

[s-ol]

The documentation also suggests to compoile as follows:

$ zig build-lib math.zig -target wasm32-freestanding

https://ziglang.org/documentation/master/#Freestanding

As of 0.7.0, compiling the example with -o ReleaseSmall results in a 51kb .wasm file containing a lot of seemingly useless symbols, not only from compiler_rt:

% wasm-objdump -d math.wasm | grep func
000db7 func[1] <__wasm_call_ctors>:
000dba func[2] <add>:
000dc8 func[3] <memset>:
000df3 func[4] <__memset>:
000e1e func[5] <memcpy>:
000e4f func[6] <memmove>:
000ebf func[7] <memcmp>:
000f0b func[8] <bcmp>:
000f4a func[9] <fmodf>:
001123 func[10] <fmod>:
001328 func[11] <floorf>:
0013a8 func[12] <ceilf>:
00142a func[13] <floor>:
0014e5 func[14] <ceil>:
0015a0 func[15] <fma>:
001b44 func[16] <std.math.frexp.frexp64>:
001c85 func[17] <fmaf>:
001c95 func[18] <sin>:
001f32 func[19] <sinf>:
002126 func[20] <cos>:
00239f func[21] <cosf>:
00257b func[22] <exp>:
0027f6 func[23] <expf>:
0029fb func[24] <exp2>:
002be6 func[25] <exp2f>:
002cff func[26] <log>:
002e86 func[27] <logf>:
002f96 func[28] <log2>:
003160 func[29] <log2f>:
00328a func[30] <log10>:
00346d func[31] <log10f>:
0035a7 func[32] <fabs>:
0035ba func[33] <fabsf>:
0035c8 func[34] <trunc>:
003625 func[35] <truncf>:
00367b func[36] <round>:
00373c func[37] <roundf>:
0037e1 func[38] <sqrt>:
003a07 func[39] <sqrtf>:
003f00 func[40] <__clzsi2>:
003f6e func[41] <__letf2>:
00404f func[42] <__gesf2>:
0040c0 func[43] <__gedf2>:
004149 func[44] <__getf2>:
004222 func[45] <__eqsf2>:
00429b func[46] <__eqdf2>:
00432c func[47] <__gnu_h2f_ieee>:
0043c5 func[48] <__gnu_f2h_ieee>:
0044d7 func[49] <__unordsf2>:
0044fd func[50] <__unorddf2>:
004537 func[51] <__unordtf2>:
00459b func[52] <__addsf3>:
0045a9 func[53] <compiler_rt.addXf3.addXf3>:
004826 func[54] <__adddf3>:
004834 func[55] <compiler_rt.addXf3.addXf3.26>:
004b3c func[56] <__addtf3>:
004b82 func[57] <compiler_rt.addXf3.addXf3.27>:
005129 func[58] <__ashlti3>:
005180 func[59] <__lshrti3>:
0051d7 func[60] <__subsf3>:
0051ed func[61] <__subdf3>:
005208 func[62] <__subtf3>:
00525a func[63] <__mulsf3>:
0054ba func[64] <__muldf3>:
0057f7 func[65] <__multf3>:
005e0b func[66] <__divsf3>:
0060ad func[67] <__divdf3>:
006459 func[68] <__divtf3>:
006e62 func[69] <__multi3>:
006ed8 func[70] <__ashldi3>:
006f38 func[71] <__ashrdi3>:
006f9b func[72] <__ashrti3>:
006ff5 func[73] <__lshrdi3>:
007056 func[74] <__floatsidf>:
007107 func[75] <__floatsisf>:
00719a func[76] <__floatdidf>:
0071e4 func[77] <__floatsitf>:
007292 func[78] <__floatunsisf>:
007308 func[79] <__floatundisf>:
0073ca func[80] <__floatunsidf>:
00740c func[81] <__floatundidf>:
007457 func[82] <__floatditf>:
007510 func[83] <__floattitf>:
0076e2 func[84] <__floattidf>:
0078c2 func[85] <__floattisf>:
007a31 func[86] <__floatdisf>:
007b0e func[87] <__floatunditf>:
007ba4 func[88] <__floatunsitf>:
007c3a func[89] <__floatuntitf>:
007ddd func[90] <__floatuntidf>:
007f87 func[91] <__floatuntisf>:
0080c6 func[92] <__extenddftf2>:
0081db func[93] <__extendsftf2>:
0082d7 func[94] <__truncdfhf2>:
008432 func[95] <__trunctfdf2>:
008631 func[96] <__trunctfsf2>:
00880e func[97] <__truncdfsf2>:
00896a func[98] <__extendsfdf2>:
008a17 func[99] <__fixunssfsi>:
008a85 func[100] <__fixunssfdi>:
008af9 func[101] <__fixunssfti>:
008bd1 func[102] <__fixunsdfsi>:
008c32 func[103] <__fixunsdfdi>:
008cac func[104] <__fixunsdfti>:
008d8a func[105] <__fixunstfsi>:
008e12 func[106] <__fixunstfdi>:
008e9e func[107] <__fixunstfti>:
008f84 func[108] <__fixdfdi>:
009059 func[109] <__fixdfsi>:
0090fb func[110] <__fixdfti>:
00927b func[111] <__fixsfdi>:
00934a func[112] <__fixsfsi>:
0093fa func[113] <__fixsfti>:
009573 func[114] <__fixtfdi>:
009676 func[115] <__fixtfsi>:
009763 func[116] <__fixtfti>:
0098e1 func[117] <__udivmoddi4>:
009bef func[118] <__popcountdi2>:
009c56 func[119] <__mulsi3>:
009c8c func[120] <__muldi3>:
009d05 func[121] <__divmoddi4>:
009d49 func[122] <__divdi3>:
009d7e func[123] <__divsi3>:
009dad func[124] <__udivsi3>:
009e55 func[125] <__udivdi3>:
009e64 func[126] <__modsi3>:
009e98 func[127] <__moddi3>:
009ef1 func[128] <__umodsi3>:
009f9d func[129] <__umoddi3>:
009fd7 func[130] <__divmodsi4>:
00a015 func[131] <__udivmodsi4>:
00a0c8 func[132] <__negsf2>:
00a0d6 func[133] <__negdf2>:
00a0ea func[134] <__divti3>:
00a189 func[135] <compiler_rt.udivmod.udivmod.64>:
00a501 func[136] <__modti3>:
00a59a func[137] <__udivti3>:
00a5e8 func[138] <__udivmodti4>:
00a636 func[139] <__umodti3>:
00a681 func[140] <__muloti4>:
00a831 func[141] <__mulodi4>:
00a907 func[142] <__atomic_load>:
00a947 func[143] <__atomic_store>:
00a987 func[144] <__atomic_exchange>:
00a9d5 func[145] <__atomic_compare_exchange>:
00aa61 func[146] <__atomic_load_1>:
00aa69 func[147] <__atomic_load_2>:
00aa71 func[148] <__atomic_load_4>:
00aa79 func[149] <__atomic_load_8>:
00aab7 func[150] <__atomic_store_1>:
00aac1 func[151] <__atomic_store_2>:
00aacb func[152] <__atomic_store_4>:
00aad5 func[153] <__atomic_store_8>:
00ab0f func[154] <__atomic_exchange_1>:
00ab24 func[155] <__atomic_exchange_2>:
00ab39 func[156] <__atomic_exchange_4>:
00ab4e func[157] <__atomic_exchange_8>:
00ab93 func[158] <__atomic_compare_exchange_1>:
00abc4 func[159] <__atomic_compare_exchange_2>:
00abf5 func[160] <__atomic_compare_exchange_4>:
00ac26 func[161] <__atomic_compare_exchange_8>:
00ac87 func[162] <__atomic_fetch_add_1>:
00ac9d func[163] <__atomic_fetch_add_2>:
00acb3 func[164] <__atomic_fetch_add_4>:
00acc9 func[165] <__atomic_fetch_add_8>:
00ad0f func[166] <__atomic_fetch_sub_1>:
00ad25 func[167] <__atomic_fetch_sub_2>:
00ad3b func[168] <__atomic_fetch_sub_4>:
00ad51 func[169] <__atomic_fetch_sub_8>:
00ad97 func[170] <__atomic_fetch_and_1>:
00adad func[171] <__atomic_fetch_and_2>:
00adc3 func[172] <__atomic_fetch_and_4>:
00add9 func[173] <__atomic_fetch_and_8>:
00ae1f func[174] <__atomic_fetch_or_1>:
00ae35 func[175] <__atomic_fetch_or_2>:
00ae4b func[176] <__atomic_fetch_or_4>:
00ae61 func[177] <__atomic_fetch_or_8>:
00aea7 func[178] <__atomic_fetch_xor_1>:
00aebd func[179] <__atomic_fetch_xor_2>:
00aed3 func[180] <__atomic_fetch_xor_4>:
00aee9 func[181] <__atomic_fetch_xor_8>:
00af2f func[182] <__atomic_fetch_nand_1>:
00af48 func[183] <__atomic_fetch_nand_2>:
00af61 func[184] <__atomic_fetch_nand_4>:
00af7a func[185] <__atomic_fetch_nand_8>:
00afc3 func[186] <__lesf2>:
00afd0 func[187] <__ledf2>:
00afdd func[188] <__ltsf2>:
00afea func[189] <__ltdf2>:
00aff7 func[190] <__nesf2>:
00b004 func[191] <__nedf2>:
00b011 func[192] <__gtsf2>:
00b01e func[193] <__gtdf2>:

[FireFox317]

This suggest we are not passing the --gc-sections flag to the linker anymore. Well actually that only happends when linking as pointed out above (with build-exe)

[s-ol]

I'm not sure what --gc-sections does exactly, but if it is the equivalent of wabt's wasm-strip then that's not all thats wrong; running the example math.wasm through wasm-strip reduces the filesize from 51k to 48k only.

By adding a dummy pub fn main() that calls add and using build-exe, the filesize goes down to just 150 bytes, but the add function is inlined and no longer exported as a symbol, so this is not a viable workaround for compiling wasm libraries.

[jedisct1]

Build an executable instead.

If you are building a reactor, still add a main() function, containing only references to your entry points:

std.mem.doNotOptimizeAway(entrypoint1);
std.mem.doNotOptimizeAway(entrypoint2);
...

Unused functions will not be present in your module.

[s-ol]

thanks @jedisct1 for the workaround, this brings the binary size down to 300 bytes, but it is clearly neither comfortable nor how things are intended to be used.

That's correct. Unused functions are not garbage collected until linking. Linking happens with build_exe.

This does not appear to be fully correct, compiling the WASM example without an added main entrypoint or any other workarounds for a native target (in this case x86_64-macos.10.15.3...10.15.3-gnu) and with build-lib, only the symbols marked extern or export are exported:

% objdump -t libmath.a 

libmath.a(zig-cache/o/7b9f6554b15377ff30a5ec8eb751292c/math.o):	file format Mach-O 64-bit x86-64

SYMBOL TABLE:
0000000000000008 gw    O __DATA,__data __mh_execute_header
0000000000000000 g     F __TEXT,__text _add
0000000000000000         *UND* _print

I would expect an equivalent result once the wasm target is added - a very small wasm file that contains only the symbols that are marked as exported, and whichever symbols these require to operate correctly.