https://llvm.org/bugs/show_bug.cgi?id=27631
Bug ID: 27631
Summary: Poor code generation from LLVM.
Product: new-bugs
Version: trunk
Hardware: PC
OS: Linux
Status: NEW
Severity: normal
Priority: P
Component: new bugs
Assignee: unassignedb...@nondot.org
Reporter: co...@google.com
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
LLVM generates very poor code for the following source code:
struct A {
int t[16] = {0};
};
const int N = 10000;
A a[N];
void foo() {
for (int i = 0; i < N; ++i) {
a[i] = A();
a[i].t[0] = a[i].t[1] = a[i].t[2] = a[i].t[3] =
a[i].t[4] = a[i].t[5] = a[i].t[6] = a[i].t[7] = i;
}
}
When disabling loop unrolling, the assembly of the loop body generated on x86
is shown below (from clang++ t.C -O2 -fno-unroll-loops -S):
xorps %xmm0, %xmm0
LOOP:
movaps %xmm0, -24(%rsp)
movaps %xmm0, -40(%rsp)
movaps %xmm0, -56(%rsp)
movaps %xmm0, -72(%rsp)
movaps -72(%rsp), %xmm1
movaps -56(%rsp), %xmm2
movaps -40(%rsp), %xmm3
movaps -24(%rsp), %xmm4
movups %xmm4, 48(%rax)
movups %xmm3, 32(%rax)
movups %xmm2, 16(%rax)
movups %xmm1, (%rax)
movd %ecx, %xmm1
pshufd $0, %xmm1, %xmm1 # xmm1 = xmm1[0,0,0,0]
movdqa %xmm1, 16(%rax)
movdqa %xmm1, (%rax)
incq %rcx
addq $64, %rax
cmpq $10000, %rcx # imm = 0x2710
jne .LBB0_1
Note that how the values in %xmm0 (all zeros) are copied to the stack multiple
times, which are then copied back to other SIMD registers, which are copied to
memory. Furthermore, the first 32 bytes of each object are written twice.
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
