[Haskell-cafe] redundant loads and saves in code generated for recursive functions?
Jyotirmoy Bhattacharya
jyotirmoy at jyotirmoy.net
Wed Jul 30 08:24:37 UTC 2014
Dear All,
I am new to Haskell so please forgive me if I am asking about something
already well-understood.
I was trying to understand the performance of my Haskell program compiled
with the LLVM backend. I used -ddump-llvm to dump the LLVM assembly and
then ran llc -O3 on the resulting file to look at the native assembly.
One of the generated function starts off with
s5BH_info: # @s5BH_info
# BB#0:
subq $208, %rsp
movq %r13, 200(%rsp)
movq %rbp, 192(%rsp)
movq %r12, 184(%rsp)
movq %rbx, 176(%rsp)
movq %r14, 168(%rsp)
movq %rsi, 160(%rsp)
movq %rdi, 152(%rsp)
movq %r8, 144(%rsp)
movq %r9, 136(%rsp)
movq %r15, 128(%rsp)
movss %xmm1, 124(%rsp)
movss %xmm2, 120(%rsp)
movss %xmm3, 116(%rsp)
movss %xmm4, 112(%rsp)
movsd %xmm5, 104(%rsp)
movsd %xmm6, 96(%rsp)
At some point down the line the function makes a tail call to itself and
this is the code generated
movq %r14, 168(%rsp)
movq 200(%rsp), %r13
movq 192(%rsp), %rbp
movq 184(%rsp), %r12
movq 176(%rsp), %rbx
movq 128(%rsp), %r15
movsd 104(%rsp), %xmm5
addq $208, %rsp
jmp s5BH_info
So it looks like some values are being moved from registers to the stack
only to be immediately moved from the stack to the register on entry to the
function. It should be possible to eliminate both the load and the stores.
Is this behaviour due to LLVM or GHC? If it is GHC, it this an optimization
a newcomer can attempt to implement or are there deep issues here?
Jyotirmoy Bhattacharya
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/haskell-cafe/attachments/20140730/a4b59519/attachment.html>
More information about the Haskell-Cafe
mailing list