https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88760
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
So LLVM unrolls 4 times while GCC (always) unrolls 8 times. The unrolled body
for GCC (x86_64 this time) is
.L4:
movl (%rdx), %ecx
vmovsd (%rax), %xmm8
addq $32, %rdx
addq $64, %rax
vmovsd -56(%rax), %xmm9
vmovsd -48(%rax), %xmm10
vfmadd231sd (%rsi,%rcx,8), %xmm8, %xmm0
movl -28(%rdx), %ecx
vmovsd -40(%rax), %xmm11
vmovsd -32(%rax), %xmm12
vfmadd231sd (%rsi,%rcx,8), %xmm9, %xmm0
movl -24(%rdx), %ecx
vmovsd -24(%rax), %xmm13
vmovsd -16(%rax), %xmm14
vfmadd231sd (%rsi,%rcx,8), %xmm10, %xmm0
movl -20(%rdx), %ecx
vmovsd -8(%rax), %xmm15
vfmadd231sd (%rsi,%rcx,8), %xmm11, %xmm0
movl -16(%rdx), %ecx
vfmadd231sd (%rsi,%rcx,8), %xmm12, %xmm0
movl -12(%rdx), %ecx
vfmadd231sd (%rsi,%rcx,8), %xmm13, %xmm0
movl -8(%rdx), %ecx
vfmadd231sd (%rsi,%rcx,8), %xmm14, %xmm0
movl -4(%rdx), %ecx
vfmadd231sd (%rsi,%rcx,8), %xmm15, %xmm0
cmpq %rax, %r9
jne .L4
and what you quoted is the prologue. You didn't quote llvms prologue
but if I read my clangs outout correct it uses a loop there.
(is there sth like -fdump-tree-optimized for clang?)
Our RTL unroller cannot do a loopy prologue but it always has this
jump-into peeled copies thing. Using --param max-unroll-times=4
produces
.L4:
movl (%rdx), %ecx
vmovsd (%rax), %xmm2
addq $16, %rdx
addq $32, %rax
vmovsd -24(%rax), %xmm3
vmovsd -16(%rax), %xmm4
vfmadd231sd (%rsi,%rcx,8), %xmm2, %xmm0
movl -12(%rdx), %ecx
vmovsd -8(%rax), %xmm5
vfmadd231sd (%rsi,%rcx,8), %xmm3, %xmm0
movl -8(%rdx), %ecx
vfmadd231sd (%rsi,%rcx,8), %xmm4, %xmm0
movl -4(%rdx), %ecx
vfmadd231sd (%rsi,%rcx,8), %xmm5, %xmm0
cmpq %rax, %r8
jne .L4
which is nearly equivalent to clnags varaint?
