OV - attention-avx512 - Loop 39

	Loop Id: 39	Module: attention-avx512	Source: attention.cpp:55-56	Coverage: 0.11%

0x69f0 VMOVSS	(%RCX,%R13,4),%XMM0    [3]

0x69f6 VSUBSS	%XMM1,%XMM0,%XMM0

0x69fa VZEROUPPER

0x69fd CALL	7170 <@plt_start@+0x20>

0x6a02 VMOVAPS	0x80(%RSP),%XMM2    [1]

0x6a0b VMOVAPS	0x160(%RSP),%XMM1    [1]

0x6a14 MOV	0x1a0(%RSP),%RDX    [1]

0x6a1c MOV	0x18(%RSP),%RCX    [1]

0x6a21 VDIVSS	%XMM2,%XMM0,%XMM0

0x6a25 VMOVSS	%XMM0,(%RDX,%R13,4)    [2]

0x6a2b INC	%R13

0x6a2e CMP	%R13,%R12

0x6a31 JNE	69f0

/home/eoseret/llm-attention/attention.cpp: 55 - 56

--------------------------------------------------------------------------------

55:       for (int idx = 0; idx <= row; ++idx)  //vectorised

56:          P[row * N + idx] = expf(S_row[idx] - max_val) / sum;

Coverage (%)	Name	Source Location	Module

min	med	avg	max

Percentile Index	10	20	30	40	50	60	70	80	90	100
Value

min	med	avg	max

Percentile Index	10	20	30	40	50	60	70	80	90	100
Value

Path /

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	2.39
CQA speedup if fully vectorized	12.31
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.43
Bottlenecks	P10, P11,
Function	main
Source	attention.cpp:55-56
Source loop unroll info	unrolled by 8
Source loop unroll confidence level	max
Unroll/vectorization loop type	peel/tail
Unroll factor	1
CQA cycles	2.50
CQA cycles if no scalar integer	2.50
CQA cycles if FP arith vectorized	1.05
CQA cycles if fully vectorized	0.20
Front-end cycles	1.63
P0 cycles	0.33
P1 cycles	0.33
P2 cycles	0.33
P3 cycles	0.67
P4 cycles	0.67
P5 cycles	0.67
P6 cycles	1.75
P7 cycles	1.75
P8 cycles	1.75
P9 cycles	1.75
P10 cycles	0.50
P11 cycles	0.50
P12 cycles	0.50
P13 cycles	0.50
P14 cycles	0.50
P15 cycles	0.50
DIV/SQRT cycles	2.50
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	13.00
Nb uops	13.00
Nb loads	5.00
Nb stores	1.00
Nb stack references	4.00
FLOP/cycle	0.80
Nb FLOP add-sub	1.00
Nb FLOP mul	0.00
Nb FLOP fma	0.00
Nb FLOP div	1.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	22.40
Bytes prefetched	0.00
Bytes loaded	52.00
Bytes stored	4.00
Stride 0	1.00
Stride 1	2.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	42.86
Vectorization ratio load	66.67
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	0.00
Vectorization ratio other	100.00
Vector-efficiency ratio all	14.29
Vector-efficiency ratio load	18.75
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	6.25
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	6.25
Vector-efficiency ratio other	25.00

Metric	Value
CQA speedup if no scalar integer	1.00
CQA speedup if FP arith vectorized	2.39
CQA speedup if fully vectorized	12.31
CQA speedup if no inter-iteration dependency	NA
CQA speedup if next bottleneck killed	1.43
Bottlenecks	P10, P11,
Function	main
Source	attention.cpp:55-56
Source loop unroll info	unrolled by 8
Source loop unroll confidence level	max
Unroll/vectorization loop type	peel/tail
Unroll factor	1
CQA cycles	2.50
CQA cycles if no scalar integer	2.50
CQA cycles if FP arith vectorized	1.05
CQA cycles if fully vectorized	0.20
Front-end cycles	1.63
P0 cycles	0.33
P1 cycles	0.33
P2 cycles	0.33
P3 cycles	0.67
P4 cycles	0.67
P5 cycles	0.67
P6 cycles	1.75
P7 cycles	1.75
P8 cycles	1.75
P9 cycles	1.75
P10 cycles	0.50
P11 cycles	0.50
P12 cycles	0.50
P13 cycles	0.50
P14 cycles	0.50
P15 cycles	0.50
DIV/SQRT cycles	2.50
Inter-iter dependencies cycles	1
FE+BE cycles (UFS)	NA
Stall cycles (UFS)	NA
Nb insns	13.00
Nb uops	13.00
Nb loads	5.00
Nb stores	1.00
Nb stack references	4.00
FLOP/cycle	0.80
Nb FLOP add-sub	1.00
Nb FLOP mul	0.00
Nb FLOP fma	0.00
Nb FLOP div	1.00
Nb FLOP rcp	0.00
Nb FLOP sqrt	0.00
Nb FLOP rsqrt	0.00
Bytes/cycle	22.40
Bytes prefetched	0.00
Bytes loaded	52.00
Bytes stored	4.00
Stride 0	1.00
Stride 1	2.00
Stride n	0.00
Stride unknown	0.00
Stride indirect	0.00
Vectorization ratio all	42.86
Vectorization ratio load	66.67
Vectorization ratio store	0.00
Vectorization ratio mul	NA
Vectorization ratio add_sub	0.00
Vectorization ratio fma	NA
Vectorization ratio div_sqrt	0.00
Vectorization ratio other	100.00
Vector-efficiency ratio all	14.29
Vector-efficiency ratio load	18.75
Vector-efficiency ratio store	6.25
Vector-efficiency ratio mul	NA
Vector-efficiency ratio add_sub	6.25
Vector-efficiency ratio fma	NA
Vector-efficiency ratio div_sqrt	6.25
Vector-efficiency ratio other	25.00

Path /

Average path: Display a virtual path defined by average values of all real paths

Function	main
Source file and lines	attention.cpp:55-56
Module	attention-avx512

The loop is defined in /home/eoseret/llm-attention/attention.cpp:55-56.

It is peel/tail loop of related source loop which is unrolled by 8 (including vectorization).

gain
potential
hint
expert

Unrolling/vectorization cost

This loop is peel/tail of a unrolled/vectorized loop. If its cost is not negligible compared to the main (unrolled/vectorized) loop, unrolling/vectorization is counterproductive due to low trip count.

Details

The more iterations the main loop is processing, the higher the trip count must be to amortize peel/tail overhead.

Workaround

recompile with -fprofile-instr-generate, execute ,merge raw profiles with 'llvm-profdata merge -o default.profdata default.profraw' and recompile with -fprofile-instr-use (profile-guided optimization)
hardcode most frequent values of loop bounds by adding specialized paths.:
- For instance, replace for (i=0; i<n; i++) foo(i) with: switch (n) { case (4): for (i=0; i<4; i++) foo(i); break; case (6): for (i=0; i<6; i++) foo(i); break; default : for (i=0; i<n; i++) foo(i); break; }

Vectorization

Your loop is probably not vectorized. Only 14% of vector register length is used (average across all SSE/AVX instructions).

Details

Store and arithmetical SSE/AVX instructions are used in scalar version (process only one data element in vector registers).

Execution units bottlenecks

Performance is limited by execution of divide and square root operations (the divide/square root unit is a bottleneck). By removing all these bottlenecks, you can lower the cost of an iteration from 2.50 to 1.75 cycles (1.43x speedup).

Workaround

Reduce the number of division or square root instructions:

If denominator is constant over iterations, use reciprocal (replace x/y with x*(1/y)). Check precision impact. This will be done by your compiler with ffast-math or Ofast

Expensive FP math instructions/calls

Detected performance impact from expensive FP math instructions/calls. By removing/reexpressing them, you can lower the cost of an iteration from 2.50 to 1.75 cycles (1.43x speedup).

CALL instructions

Detected function call instructions.

Details

Calling (and then returning from) a function prevents many compiler optimizations (like vectorization), breaks control flow (which reduces pipeline performance) and executes extra instructions to save/restore the registers used inside it, which is very expensive (dozens of cycles). Consider to inline small functions.

unknown: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:56

Type of elements and instruction set

2 SSE or AVX instructions are processing arithmetic or math operations on single precision FP elements in scalar mode (one at a time).

Matching between your loop (in the source code) and the binary loop

The binary loop is composed of 2 FP arithmetical operations:

1: addition or subtraction
1: divide

The binary loop is loading 52 bytes (13 single precision FP elements). The binary loop is storing 4 bytes (1 single precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.04 FP operations per loaded or stored byte.

General properties

nb instructions	13
nb uops	13
loop length	67
used x86 registers	5
used mmx registers	0
used xmm registers	3
used ymm registers	0
used zmm registers	0
nb stack references	4

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	1.63 cycles
front end	1.63 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15
uops	0.33	0.33	0.33	0.67	0.67	0.67	1.75	1.75	1.75	1.75	0.50	0.50	0.50	0.50	0.50	0.50
cycles	0.33	0.33	0.33	0.67	0.67	0.67	1.75	1.75	1.75	1.75	0.50	0.50	0.50	0.50	0.50	0.50

Execution ports to units layout:

P0: ALU
P1: ALU
P2: ALU
P3: ALU, BRU
P4: ALU, BRU
P5: ALU, BRU
P6 (512 bits): store address, load
P7 (512 bits): store address, load
P8 (512 bits): store address, load
P9 (512 bits): store address, load
P10 (512 bits): VPU, DIV/SQRT
P11 (512 bits): VPU, DIV/SQRT
P12 (512 bits): VPU
P13 (512 bits): VPU
P14 (512 bits): FP store data
P15 (512 bits): FP store data

Cycles executing div or sqrt instructions	2.50
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	1.63
Dispatch	1.75
DIV/SQRT	2.50
Data deps.	1.00
Overall L1	2.50

Vectorization ratios

INT

all	100%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	NA (no fma vectorizable/vectorized instructions)
other	100%

all	33%
load	66%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	0%
other	NA (no other vectorizable/vectorized instructions)

INT+FP

all	42%
load	66%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	0%
other	100%

Vector efficiency ratios

INT

all	25%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	NA (no fma vectorizable/vectorized instructions)
other	25%

all	12%
load	18%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	6%
other	NA (no other vectorizable/vectorized instructions)

INT+FP

all	14%
load	18%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	6%
other	25%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 2.50 cycles. At this rate:

8% of peak load performance is reached (20.80 out of 256.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (1.60 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 69f0

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	Latency	Recip. throughput	Vectorization
VMOVSS (%RCX,%R13,4),%XMM0	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.50	scal (6.3%)
VSUBSS %XMM1,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (6.3%)
VZEROUPPER	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1	vect (25.0%)
CALL 7170 <@plt_start@+0x20>	2	0	0	0	0.33	0.33	0.33	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0.50	N/A
VMOVAPS 0x80(%RSP),%XMM2	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.50	vect (25.0%)
VMOVAPS 0x160(%RSP),%XMM1	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.50	vect (25.0%)
MOV 0x1a0(%RSP),%RDX	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
MOV 0x18(%RSP),%RCX	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
VDIVSS %XMM2,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	10	2.50	scal (6.3%)
VMOVSS %XMM0,(%RDX,%R13,4)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0.50	0.50	1	0.50	scal (6.3%)
INC %R13	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %R13,%R12	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
JNE 69f0 <main+0x2d50>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-1	N/A

Function	main
Source file and lines	attention.cpp:55-56
Module	attention-avx512

The loop is defined in /home/eoseret/llm-attention/attention.cpp:55-56.

It is peel/tail loop of related source loop which is unrolled by 8 (including vectorization).

gain
potential
hint
expert

Unrolling/vectorization cost

Details

The more iterations the main loop is processing, the higher the trip count must be to amortize peel/tail overhead.

Workaround

recompile with -fprofile-instr-generate, execute ,merge raw profiles with 'llvm-profdata merge -o default.profdata default.profraw' and recompile with -fprofile-instr-use (profile-guided optimization)
hardcode most frequent values of loop bounds by adding specialized paths.:
- For instance, replace for (i=0; i<n; i++) foo(i) with: switch (n) { case (4): for (i=0; i<4; i++) foo(i); break; case (6): for (i=0; i<6; i++) foo(i); break; default : for (i=0; i<n; i++) foo(i); break; }

Vectorization

Your loop is probably not vectorized. Only 14% of vector register length is used (average across all SSE/AVX instructions).

Details

Store and arithmetical SSE/AVX instructions are used in scalar version (process only one data element in vector registers).

Execution units bottlenecks

Workaround

Reduce the number of division or square root instructions:

If denominator is constant over iterations, use reciprocal (replace x/y with x*(1/y)). Check precision impact. This will be done by your compiler with ffast-math or Ofast

Expensive FP math instructions/calls

Detected performance impact from expensive FP math instructions/calls. By removing/reexpressing them, you can lower the cost of an iteration from 2.50 to 1.75 cycles (1.43x speedup).

CALL instructions

Detected function call instructions.

Details

unknown: 1 occurrences►
- /home/eoseret/llm-attention/attention.cpp:56

Type of elements and instruction set

2 SSE or AVX instructions are processing arithmetic or math operations on single precision FP elements in scalar mode (one at a time).

Matching between your loop (in the source code) and the binary loop

The binary loop is composed of 2 FP arithmetical operations:

1: addition or subtraction
1: divide

The binary loop is loading 52 bytes (13 single precision FP elements). The binary loop is storing 4 bytes (1 single precision FP elements).

Arithmetic intensity

Arithmetic intensity is 0.04 FP operations per loaded or stored byte.

General properties

nb instructions	13
nb uops	13
loop length	67
used x86 registers	5
used mmx registers	0
used xmm registers	3
used ymm registers	0
used zmm registers	0
nb stack references	4

Front-end

ASSUMED MACRO FUSION FIT IN UOP CACHE

micro-operation queue	1.63 cycles
front end	1.63 cycles

Back-end

	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15
uops	0.33	0.33	0.33	0.67	0.67	0.67	1.75	1.75	1.75	1.75	0.50	0.50	0.50	0.50	0.50	0.50
cycles	0.33	0.33	0.33	0.67	0.67	0.67	1.75	1.75	1.75	1.75	0.50	0.50	0.50	0.50	0.50	0.50

Execution ports to units layout:

P0: ALU
P1: ALU
P2: ALU
P3: ALU, BRU
P4: ALU, BRU
P5: ALU, BRU
P6 (512 bits): store address, load
P7 (512 bits): store address, load
P8 (512 bits): store address, load
P9 (512 bits): store address, load
P10 (512 bits): VPU, DIV/SQRT
P11 (512 bits): VPU, DIV/SQRT
P12 (512 bits): VPU
P13 (512 bits): VPU
P14 (512 bits): FP store data
P15 (512 bits): FP store data

Cycles executing div or sqrt instructions	2.50
Longest recurrence chain latency (RecMII)	1.00

Cycles summary

Front-end	1.63
Dispatch	1.75
DIV/SQRT	2.50
Data deps.	1.00
Overall L1	2.50

Vectorization ratios

INT

all	100%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	NA (no fma vectorizable/vectorized instructions)
other	100%

all	33%
load	66%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	0%
other	NA (no other vectorizable/vectorized instructions)

INT+FP

all	42%
load	66%
store	0%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	0%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	0%
other	100%

Vector efficiency ratios

INT

all	25%
load	NA (no load vectorizable/vectorized instructions)
store	NA (no store vectorizable/vectorized instructions)
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	NA (no add-sub vectorizable/vectorized instructions)
fma	NA (no fma vectorizable/vectorized instructions)
other	25%

all	12%
load	18%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	6%
other	NA (no other vectorizable/vectorized instructions)

INT+FP

all	14%
load	18%
store	6%
mul	NA (no mul vectorizable/vectorized instructions)
add-sub	6%
fma	NA (no fma vectorizable/vectorized instructions)
div/sqrt	6%
other	25%

Cycles and memory resources usage

Assuming all data fit into the L1 cache, each iteration of the binary loop takes 2.50 cycles. At this rate:

8% of peak load performance is reached (20.80 out of 256.00 bytes loaded per cycle (GB/s @ 1GHz))
1% of peak store performance is reached (1.60 out of 128.00 bytes stored per cycle (GB/s @ 1GHz))

Front-end bottlenecks

Found no such bottlenecks.

ASM code

In the binary file, the address of the loop is: 69f0

Instruction	Nb FU	P0	P1	P2	P3	P4	P5	P6	P7	P8	P9	P10	P11	P12	P13	P14	P15	Latency	Recip. throughput	Vectorization
VMOVSS (%RCX,%R13,4),%XMM0	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	1	0.50	scal (6.3%)
VSUBSS %XMM1,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	2	0.50	scal (6.3%)
VZEROUPPER	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1	vect (25.0%)
CALL 7170 <@plt_start@+0x20>	2	0	0	0	0.33	0.33	0.33	0.25	0.25	0.25	0.25	0	0	0	0	0	0	0	0.50	N/A
VMOVAPS 0x80(%RSP),%XMM2	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.50	vect (25.0%)
VMOVAPS 0x160(%RSP),%XMM1	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.50	vect (25.0%)
MOV 0x1a0(%RSP),%RDX	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
MOV 0x18(%RSP),%RCX	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0	0	3	0.25	N/A
VDIVSS %XMM2,%XMM0,%XMM0	1	0	0	0	0	0	0	0	0	0	0	0.50	0.50	0	0	0	0	10	2.50	scal (6.3%)
VMOVSS %XMM0,(%RDX,%R13,4)	1	0	0	0	0	0	0	0.25	0.25	0.25	0.25	0	0	0	0	0.50	0.50	1	0.50	scal (6.3%)
INC %R13	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
CMP %R13,%R12	1	0.17	0.17	0.17	0.17	0.17	0.17	0	0	0	0	0	0	0	0	0	0	1	0.17	N/A
JNE 69f0 <main+0x2d50>	1	0	0	0	0.33	0.33	0.33	0	0	0	0	0	0	0	0	0	0	0	0.33-1	N/A

Report Configuration

Unrolling/vectorization cost

Details

Workaround

Vectorization

Details

Execution units bottlenecks

Workaround

Expensive FP math instructions/calls

CALL instructions

Details

Type of elements and instruction set

Matching between your loop (in the source code) and the binary loop

Arithmetic intensity

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Vector efficiency ratios

Cycles and memory resources usage

Front-end bottlenecks

ASM code

Unrolling/vectorization cost

Details

Workaround

Vectorization

Details

Execution units bottlenecks

Workaround

Expensive FP math instructions/calls

CALL instructions

Details

Type of elements and instruction set

Matching between your loop (in the source code) and the binary loop

Arithmetic intensity

General properties

Front-end

Back-end

Cycles summary

Vectorization ratios

Vector efficiency ratios

Cycles and memory resources usage

Front-end bottlenecks

ASM code