bextr_screwup.asm

// for this code:
// (see cbloom rant 10-17-15 http://www.cbloom.com/rants.html for context)
//
// this is completely untested, it's just a sketch

#include <immintrin.h>
#include <stdint.h>

static const uint32_t CODELEN_LIMIT = 12;

static inline uint64_t read_be_u64(uint8_t const *decode_in)
{
    return __builtin_bswap64(*((uint64_t const *) decode_in));
}

void huffdec(uint8_t * decodeptr, uint8_t const * decode_in, uint8_t const * codelens, uint8_t const * symbols, size_t count)
{
    // NB peekpos can get negative on last DECONE!
    // (since it's CODELEN_LIMIT bits *past* the actual end)
    // not a problem, just something to be aware of
    int32_t peekpos = 64 - CODELEN_LIMIT;

    while (count--)
    {
        // refill
        // note that the number of bits remaining in the buffer is
        //   bitcount = peekpos + CODELEN_LIMIT;
        uint64_t bits = read_be_u64(decode_in);
        uint64_t bytes_advance = ((64 - CODELEN_LIMIT) - peekpos) >> 3;
        decode_in += bytes_advance;
        peekpos += bytes_advance << 3;

        uint64_t peek; 
        int32_t cl, sym;
#define DECONE() \
        peek = _bextr_u64(bits, peekpos, CODELEN_LIMIT); \
        cl = codelens[peek]; sym = symbols[peek]; \
        peekpos -= cl; \
        *decodeptr++ = (uint8_t)sym;

        DECONE();
        DECONE();
        DECONE();
        DECONE();

#undef DECONE
    }
}

; what I got was this:	
	
	mov	r11d, 52
	.align	16, 0x90
.LBB0_2:                                # %.lr.ph
                                        # =>This Inner Loop Header: Depth=1
	movbe	r9, qword ptr [rsi]
	mov	r10d, 52
	sub	r10d, r11d
	shr	r10d, 3
	add	rsi, r10
	lea	r11d, dword ptr [r11 + 8*r10]
	movzx	eax, r11b
	or	eax, 3072
	bextr	rax, r9, rax
	movzx	r10d, byte ptr [rdx + rax]
	mov	al, byte ptr [rcx + rax]
	sub	r11d, r10d
	mov	byte ptr [rdi], al
	movzx	eax, r11b
	or	eax, 3072
	bextr	rax, r9, rax
	movzx	r10d, byte ptr [rdx + rax]
	mov	al, byte ptr [rcx + rax]
	sub	r11d, r10d
	mov	byte ptr [rdi + 1], al
	movzx	eax, r11b
	or	eax, 3072
	bextr	rax, r9, rax
	movzx	r10d, byte ptr [rdx + rax]
	mov	al, byte ptr [rcx + rax]
	sub	r11d, r10d
	mov	byte ptr [rdi + 2], al
	movzx	eax, r11b
	or	eax, 3072
	bextr	r9, r9, rax
	movzx	eax, byte ptr [rdx + r9]
	sub	r11d, eax
	mov	al, byte ptr [rcx + r9]
	mov	byte ptr [rdi + 3], al
	lea	rdi, qword ptr [rdi + 4]
	dec	r8
	jne	.LBB0_2


; when what I wanted is this:
	
	mov	r11d, 52 + 3072
	.align	16, 0x90
.LBB0_2:                                # %.lr.ph
                                        # =>This Inner Loop Header: Depth=1
	movbe	r9, qword ptr [rsi]
	mov	r10d, 52
	sub	r10d, r11d
	shr	r10d, 3
	add	rsi, r10
	lea	r11d, dword ptr [r11 + 8*r10]
	bextr	rax, r9, r11
	movzx	r10d, byte ptr [rdx + rax]
	mov	al, byte ptr [rcx + rax]
	sub	r11d, r10d
	mov	byte ptr [rdi], al
	bextr	rax, r9, r11
	movzx	r10d, byte ptr [rdx + rax]
	mov	al, byte ptr [rcx + rax]
	sub	r11d, r10d
	mov	byte ptr [rdi + 1], al
	bextr	rax, r9, r11
	movzx	r10d, byte ptr [rdx + rax]
	mov	al, byte ptr [rcx + rax]
	sub	r11d, r10d
	mov	byte ptr [rdi + 2], al
	bextr	r9, r9, r11
	movzx	eax, byte ptr [rdx + r9]
	sub	r11d, eax
	mov	al, byte ptr [rcx + r9]
	mov	byte ptr [rdi + 3], al
	lea	rdi, qword ptr [rdi + 4]
	dec	r8
	jne	.LBB0_2