Compiler does not insert vec/loop or uses stack temporary and unrolls far too much

[tkoenig1]

The code

void memc (unsigned char * restrict dest, unsigned char * const restrict src,
           unsigned long int n)
{
  for (unsigned long i=0; i<n; i++)
    dest[i] = src[i] + 1;
}

when translated with

clang -O1 -emit-llvm -S cpy.c
clang -emit-llvm -c cpy.ll
opt -O2 -disable-loop-unrolling -o cpy-O.bc cpy.bc
llc -O2  --march=my66000 --frame-pointer=none --enable-vvm cpy-O.bc

yields an interesting translation, where the function starts with

        enter   r16,r0,0,1432

and the loop later has

.LBB0_4:                                ; %vector.body
                                        ; =>This Inner Loop Header: Depth=1
        vec     r1,{}
        std     r1,[sp,1424]
        ldub    r1,[r6,-161]
        std     r1,[sp,1416]
        ldub    r1,[r6,-162]
        std     r1,[sp,1312]
        ldub    r1,[r6,-163]
        std     r1,[sp,1208]
        ldub    r1,[r6,-164]
        std     r1,[sp,1104]
        ldub    r1,[r6,-165]
        std     r1,[sp,1000]
        ldub    r1,[r6,-166]
        std     r1,[sp,896]
        ldub    r1,[r6,-167]
        std     r1,[sp,792]

Two issues: Why is there a stack intermediate at all, and why the unrolling?

When compiling with

clang -O0 -emit-llvm -S cpy.c
clang -emit-llvm -c cpy.ll
opt -O2 -disable-loop-unrolling -o cpy-O.bc cpy.bc
llc -O2  --march=my66000 --frame-pointer=none --enable-vvm cpy-O.bc

it does not vectorize at all.

[bagel99]

Try compiling with the llc flag "-disable-lsr".

[tkoenig1]

Try compiling with the llc flag "-disable-lsr".

I've now modified my compile script to

#! /bin/bash
a=${1%%.c}
b=${a}_opt
clang -c -O3 -emit-llvm -fno-unroll-loops $a.c
opt  -disable-loop-unrolling -O3  --march=my66000 --frame-pointer=none --enable-vvm $a.bc  > $b.bc
llc --disable-lsr --enable-predication --enable-predication2 --enable-carry-generation --early-carry-coalesce --enable-vvm -march=my66000 $b.bc

and still get the same issue:

LBB0_4:                                ; %vector.body
                                        ; =>This Inner Loop Header: Depth=1
        vec     r6,{}
        ldub    r7,[r2,r5,15]
        ldub    r8,[r2,r5,14]
        ldub    r9,[r2,r5,13]
        ldub    r10,[r2,r5,12]
        ldub    r11,[r2,r5,11]
        ldub    r12,[r2,r5,10]
        ldub    r13,[r2,r5,9]

and so on.

[tkoenig1]

Update: This is not vectorized with

clang -c --target=my66000 -O3 -fno-vectorize -fno-slp-vectorize  -emit-llvm -fno-unroll-loops $1
opt  -disable-loop-unrolling -O3  --march=my66000 --frame-pointer=none --enable-vvm $a.bc  > $b.bc
llc --disable-lsr --enable-predication --enable-predication2 --enable-carry-generation --early-carry-coalesce --enable-vvm -march=my66000 $b.bc

the generated code is

        beq0    r3,.LBB0_3
; %bb.1:                                ; %.preheader
        ldub    r4,[r2]
        add     r4,r4,#1
        stb     r4,[r1]
        cmp     r4,r3,#1
        bne     r4,.LBB0_2
.LBB0_3:                                ; %.loopexit
        ret
.LBB0_2:                                ; %.preheader..lr.ph_crit_edge
        add     r6,r2,#1
        add     r5,r1,#1
        mov     r4,#2
.LBB0_4:                                ; %.lr.ph
                                        ; =>This Inner Loop Header: Depth=1
        ldub    r6,[r6]
        add     r6,r6,#1
        stb     r6,[r5]
        cmp     r5,r4,r3
        beq     r5,.LBB0_3
; %bb.5:                                ; %.lr.ph..lr.ph_crit_edge
                                        ;   in Loop: Header=BB0_4 Depth=1
        add     r6,r2,r4
        add     r5,r1,r4
        add     r4,r4,#1
        br      .LBB0_

[tkoenig1]

I just tried the most recent compiler, and it seems to be OK now:

memc:                                   ; @memc
; %bb.0:                                ; %entry
        beq0    r3,.LBB0_2
; %bb.1:                                ; %for.body.preheader
        mov     r4,#0
.LBB0_3:                                ; %for.body
                                        ; =>This Inner Loop Header: Depth=1
        vec     r5,{}
        ldub    r6,[r2,r4,0]
        add     r6,r6,#1
        stb     r6,[r1,r4,0]
        loop1   eq,r4,r3,#1
.LBB0_2:                                ; %for.cond.cleanup
        ret

This is with

#! /bin/bash
a=${1%%.[ci]}
b=${a}_opt
clang -fverbose-asm -c --target=my66000 -O3 -fno-vectorize -fno-slp-vectorize  -emit-llvm -fno-unroll-loops -fomit-frame-pointer $1
opt  -disable-loop-unrolling -O3  --march=my66000 --frame-pointer=none --enable-vvm $a.bc  > $b.bc
llc -O2 -enable-remove-range-check --enable-predication --enable-predication2 --enable-carry-generation --early-carry-coalesce --enable-vvm -march=my66000 $b.bc

so I guess the issue has been fixed in the meantime.

[tkoenig1]

Closing.