Re: Is float8 a reference type?

On Fri, Sep 22, 2017 at 8:38 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> "Premature optimization is the root of all evil". Do you have good reason
> to think that it's worth your time to write unsafe/unportable code? Do
> you know that your compiler doesn't turn Float8GetDatum into a no-op
> already? (Mine does, on a 64-bit machine.)

Ha ha, thank you for keeping me honest! But can you explain what is
unsafe about the cast? For a little more context: I've loaded a float8
array from a file, but I need to pass a Datum array to
construct_md_array. With an 8-byte Datum, I can just pass the original
float array, right? But with smaller Datums I need to go through the
array and convert each element. (I'm not really worried about these
files being moved between machines, so I'm willing to make the on-disk
format the same as the in-memory format.)

Since I'm expecting ~10 million elements per array, it seems like
skipping the conversion will have a real effect. I checked the
assembly and do see a difference (on both Mac+clang and Linux+gcc).
Here is the Mac command line:

platter:floatfile paul$ clang -Wall -Wmissing-prototypes
-Wpointer-arith -Wdeclaration-after-statement -Wendif-labels
-Wmissing-format-attribute -Wformat-security -fno-strict-aliasing
-fwrapv -Wno-unused-command-line-argument -O2 -I. -I./
-I/usr/local/Cellar/postgresql(at)9(dot)6/9.6.3/include/server
-I/usr/local/Cellar/postgresql(at)9(dot)6/9.6.3/include/internal
-I/usr/local/opt/gettext/include -I/usr/local/opt/openldap/include
-I/usr/local/opt/openssl/include -I/usr/local/opt/readline/include
-I/usr/local/opt/tcl-tk/include -g -S -o floatfile.s floatfile.c

Here is the assembly for the cast:

.loc 2 391 23 is_stmt 1 ## floatfile.c:391:23
movq -48(%rbp), %r15
Ltmp176:
##DEBUG_VALUE: load_floatfile:datums <- %R15

Here is the assembly for the loop (after just changing the code to `if
(FLOAT8PASSBYVAL && false)`):

.loc 2 393 21 is_stmt 1 ## floatfile.c:393:21
movslq %r15d, %r13
.loc 2 393 28 is_stmt 0 ## floatfile.c:393:28
leaq (,%r13,8), %rdi
.loc 2 393 14 ## floatfile.c:393:14
callq _palloc
movq %rax, %r12
Ltmp177:
##DEBUG_VALUE: load_floatfile:i <- 0
.loc 2 394 19 is_stmt 1 discriminator 1 ## floatfile.c:394:19
testl %r15d, %r15d
Ltmp178:
.loc 2 394 5 is_stmt 0 discriminator 1 ## floatfile.c:394:5
je LBB7_11
Ltmp179:
## BB#9:
##DEBUG_VALUE: load_floatfile:arrlen <- %R15D
##DEBUG_VALUE: load_floatfile:nulls <- [%RBP+-80]
##DEBUG_VALUE: load_floatfile:floats <- [%RBP+-72]
##DEBUG_VALUE: load_floatfile:filename <- %RBX
.loc 2 0 5 discriminator 1 ## floatfile.c:0:5
movq -72(%rbp), %rbx
Ltmp180:
##DEBUG_VALUE: load_floatfile:floats <- %RBX
xorl %r14d, %r14d
Ltmp181:
.p2align 4, 0x90
LBB7_10: ## =>This Inner Loop Header: Depth=1
##DEBUG_VALUE: load_floatfile:floats <- %RBX
##DEBUG_VALUE: load_floatfile:arrlen <- %R15D
##DEBUG_VALUE: load_floatfile:nulls <- [%RBP+-80]
.loc 2 395 34 is_stmt 1 ## floatfile.c:395:34
movsd (%rbx,%r14,8), %xmm0 ## xmm0 = mem[0],zero
.loc 2 395 19 is_stmt 0 ## floatfile.c:395:19
callq _Float8GetDatum
.loc 2 395 17 ## floatfile.c:395:17
movq %rax, (%r12,%r14,8)
Ltmp182:
.loc 2 394 30 is_stmt 1 discriminator 2 ## floatfile.c:394:30
incq %r14
.loc 2 394 19 is_stmt 0 discriminator 1 ## floatfile.c:394:19
cmpq %r13, %r14
Ltmp183:
.loc 2 394 5 discriminator 1 ## floatfile.c:394:5
jl LBB7_10
Ltmp184:
LBB7_11:
##DEBUG_VALUE: load_floatfile:arrlen <- %R15D
##DEBUG_VALUE: load_floatfile:nulls <- [%RBP+-80]

I get the same results on gcc too: the palloc, the loop, and even
`call Float8GetDatum(at)PLT`(dot)

I'll do some timing of each version too, but it doesn't look like a
pointless optimization. I'd still like to know what is unsafe about it
though.

Thanks!
Paul