Manual code generation with binaryen.js, very slow program vs emcc

[ddan-io]

Hi,
I am using binaryen.js to generate code for a simple C program. My problem is that manually generated code is 10X slower then code generated with emcc.

Could anyone please help me in figuring out my mistakes? Are the load/store or global instructions slow?

Following I attach the C program, the manually generated code binaryen.js instructions, and the disassembled emcc program. Thanks for any help.

The C program:
1:void main()
2:{
3: int i;
4: i = 1;
5: while(i<1000000000) {
6: ++i;
7: }
8:}

My manually generated code - javascript binaryen.js instructions:
(Legend:
/ * source code line:column * /,
/ * x86 "equivalent" instructions, after block of binaryen.js instructions * /,
bp - base pointer, sp - stack pointer, ax - ax register )

/* INIT * /mod.global.set('sp',mod.i32.const(11291432)),
mod.global.set('sp',mod.i32.sub(mod.global.get('sp',binaryen.i32),mod.i32.const(4))),
mod.i32.store(0,4,mod.global.get('sp',binaryen.i32),mod.i32.const(36)),
mod.global.set('sp',mod.i32.sub(mod.global.get('sp',binaryen.i32),mod.i32.const(4))),
mod.i32.store(0,4,mod.global.get('sp',binaryen.i32),mod.i32.const(13)),
mod.global.set('sp',mod.i32.sub(mod.global.get('sp',binaryen.i32),mod.i32.const(4))),
mod.i32.store(0,4,mod.global.get('sp',binaryen.i32),mod.i32.const(0)),
mod.global.set('sp',mod.i32.sub(mod.global.get('sp',binaryen.i32),mod.i32.const(4))),
mod.i32.store(0,4,mod.global.get('sp',binaryen.i32),mod.i32.const(0)),
mod.global.set('sp',mod.i32.sub(mod.global.get('sp',binaryen.i32),mod.i32.const(4))),
mod.i32.store(0,4,mod.global.get('sp',binaryen.i32),mod.i32.const(11291424)),
/1:0/
/2:0/
/3:0/
/3:4:2/mod.global.set('sp',mod.i32.sub(mod.global.get('sp',binaryen.i32),mod.i32.const(4)))/ENT/,
/3:4:2/mod.i32.store(0,4,mod.global.get('sp',binaryen.i32),mod.global.get('bp',binaryen.i32)),
/3:4:2/mod.global.set('bp',mod.global.get('sp',binaryen.i32)),
/3:4:2/mod.global.set('sp',mod.i32.sub(mod.global.get('sp',binaryen.i32),mod.i32.const(4))),
/4:2:4/mod.global.set('ax',mod.i32.add(mod.global.get('bp',binaryen.i32),mod.i32.const(-4)))/LEA/,
/4:2:4/mod.global.set('sp',mod.i32.sub(mod.global.get('sp',binaryen.i32),mod.i32.const(4)))/ASGN,P/,
mod.i32.store(0,4,mod.global.get('sp',binaryen.i32),mod.global.get('ax',binaryen.i32)),
/4:6:5/mod.global.set('ax',mod.i32.const(1))/II/,
/4:6:5/mod.i32.store(0,4,mod.i32.load(0,4,mod.global.get('sp',binaryen.i32)),mod.global.get('ax',binaryen.i32))/ASGN,SI/,
mod.global.set('sp',mod.i32.add(mod.global.get('sp',binaryen.i32),mod.i32.const(4))),
/4:0/
/* ENT 1*/
/* LEA -1*/
/* PUSH*/
/* IMM 1*/
/* SI /
/5:6:3/mod.block('while_5_7',[
mod.loop('loop_5_7',mod.block(null,[
/5:8:4/mod.global.set('ax',mod.i32.add(mod.global.get('bp',binaryen.i32),mod.i32.const(-4)))/LEA/,
/5:8:4/mod.global.set('ax',mod.i32.load(0,4,mod.global.get('ax',binaryen.i32)))/LI/,
/5:8:4/mod.global.set('sp',mod.i32.sub(mod.global.get('sp',binaryen.i32),mod.i32.const(4)))/Lt,P/,
mod.i32.store(0,4,mod.global.get('sp',binaryen.i32),mod.global.get('ax',binaryen.i32)),
/5:19:5/mod.global.set('ax',mod.i32.const(1000000000))/II/,
/5:19:5/mod.global.set('ax',mod.i32.lt_s(mod.i32.load(0,4,mod.global.get('sp',binaryen.i32)),mod.global.get('ax',binaryen.i32))),
mod.global.set('sp',mod.i32.add(mod.global.get('sp',binaryen.i32),mod.i32.const(4))),
mod.break('while_5_7',mod.i32.eq(mod.global.get('ax',binaryen.i32),mod.i32.const(0))),
/5:0/
/ LEA -1*/
/* LI /
/ PUSH*/
/* IMM 1000000000*/
/* LT /
/ JZ 0*/
/6:7:7/mod.global.set('ax',mod.i32.add(mod.global.get('bp',binaryen.i32),mod.i32.const(-4)))/LEA/,
/6:7:7/mod.global.set('sp',mod.i32.sub(mod.global.get('sp',binaryen.i32),mod.i32.const(4)))/INC/DEC,P/,
mod.i32.store(0,4,mod.global.get('sp',binaryen.i32),mod.global.get('ax',binaryen.i32)),
mod.global.set('ax',mod.i32.load(0,4,mod.global.get('ax',binaryen.i32))),/LI/
/6:7:7/mod.global.set('sp',mod.i32.sub(mod.global.get('sp',binaryen.i32),mod.i32.const(4)))/P/,
mod.i32.store(0,4,mod.global.get('sp',binaryen.i32),mod.global.get('ax',binaryen.i32)),
mod.global.set('ax',mod.i32.const(1))/II/,
/6:7:7/mod.global.set('ax',mod.i32.add(mod.global.get('ax',binaryen.i32),mod.i32.load(0,4,mod.global.get('sp',binaryen.i32))))/ADD/,
/6:7:7/mod.global.set('sp',mod.i32.add(mod.global.get('sp',binaryen.i32),mod.i32.const(4))),
mod.i32.store(0,4,mod.i32.load(0,4,mod.global.get('sp',binaryen.i32)),mod.global.get('ax',binaryen.i32))/SI/,
mod.global.set('sp',mod.i32.add(mod.global.get('sp',binaryen.i32),mod.i32.const(4))),
/6:0/
/* LEA -1*/
/* PUSH*/
/* LI /
/ PUSH*/
/* IMM 1*/
/* ADD /
/ SI */
/7:0/
mod.break('loop_5_7')
]))]),

emcc generated program (only the while block):
101 (block $label$1
102 (loop $label$2
103 (local.set $5
104 (i32.load offset=8
105 (local.get $2)
106 )
107 )
108 (local.set $6
109 (i32.const 1000000000)
110 )
111 (local.set $7
112 (local.get $5)
113 )
114 (local.set $8
115 (local.get $6)
116 )
117 (local.set $9
118 (i32.lt_s
119 (local.get $7)
120 (local.get $8)
121 )
122 )
123 (local.set $10
124 (i32.const 1)
125 )
126 (local.set $11
127 (i32.and
128 (local.get $9)
129 (local.get $10)
130 )
131 )
132 (br_if $label$1
133 (i32.eqz
134 (local.get $11)
135 )
136 )
137 (local.set $12
138 (i32.load offset=8
139 (local.get $2)
140 )
141 )
142 (local.set $13
143 (i32.const 1)
144 )
145 (local.set $14
146 (i32.add
147 (local.get $12)
148 (local.get $13)
149 )
150 )
151 (i32.store offset=8
152 (local.get $2)
153 (local.get $14)
154 )
155 (br $label$2)
156 )
157 )

[ddan-io]

I managed to profile the jitted code in d8 with perf. The stores/loads are the problem. I will close this question.