ivopts: Improve code generated for very simple loops.
This patch tidies up the code that GCC generates for simple loops,
by selecting/generating a simpler loop bound expression in ivopts.
The original motivation came from looking at the following loop (from
gcc.target/i386/pr90178.c)
int *find_ptr (int* mem, int sz, int val)
{
for (int i = 0; i < sz; i++)
if (mem[i] == val)
return &mem[i];
return 0;
}
which GCC currently compiles to:
find_ptr:
movq %rdi, %rax
testl %esi, %esi
jle .L4
leal -1(%rsi), %ecx
leaq 4(%rdi,%rcx,4), %rcx
jmp .L3
.L7: addq $4, %rax
cmpq %rcx, %rax
je .L4
.L3: cmpl %edx, (%rax)
jne .L7
ret
.L4: xorl %eax, %eax
ret
Notice the relatively complex leal/leaq instructions, that result
from ivopts using the following expression for the loop bound:
inv_expr 2: ((unsigned long) ((unsigned int) sz_8(D) +
4294967295)
* 4 + (unsigned long) mem_9(D)) + 4
which results from NITERS being (unsigned int) sz_8(D) +
4294967295,
i.e. (sz - 1), and the logic in cand_value_at determining the bound
as BASE + NITERS*STEP at the start of the final iteration and as
BASE + NITERS*STEP + STEP at the end of the final iteration.
Ideally, we'd like the middle-end optimizers to simplify
BASE + NITERS*STEP + STEP as BASE + (NITERS+1)*STEP, especially
when NITERS already has the form BOUND-1, but with type conversions
and possible overflow to worry about, the above "inv_expr 2" is the
best that can be done by fold (without additional context information).
This patch improves ivopts' cand_value_at by instead of using just
the tree expression for NITERS, passing the data structure that
explains how that expression was derived. This allows us to peek
under the surface to check that NITERS+1 doesn't overflow, and in
this patch to use the SSA_NAME already holding the required value.
In the motivating loop above, inv_expr 2 now becomes:
(unsigned long) sz_8(D) * 4 + (unsigned long) mem_9(D)
And as a result, on x86_64 we now generate:
find_ptr:
movq %rdi, %rax
testl %esi, %esi
jle .L4
movslq %esi, %rsi
leaq (%rdi,%rsi,4), %rcx
jmp .L3
.L7: addq $4, %rax
cmpq %rcx, %rax
je .L4
.L3: cmpl %edx, (%rax)
jne .L7
ret
.L4: xorl %eax, %eax
ret
This improvement required one minor tweak to GCC's testsuite for
gcc.dg/wrapped-binop-simplify.c, where we again generate better
code, and therefore no longer find as many optimization opportunities
in later passes (vrp2).
Previously:
void v1 (unsigned long *in, unsigned long *out, unsigned int n)
{
int i;
for (i = 0; i < n; i++) {
out[i] = in[i];
}
}
on x86_64 generated:
v1: testl %edx, %edx
je .L1
movl %edx, %edx
xorl %eax, %eax
.L3: movq (%rdi,%rax,8), %rcx
movq %rcx, (%rsi,%rax,8)
addq $1, %rax
cmpq %rax, %rdx
jne .L3
.L1: ret
and now instead generates:
v1: testl %edx, %edx
je .L1
movl %edx, %edx
xorl %eax, %eax
leaq 0(,%rdx,8), %rcx
.L3: movq (%rdi,%rax), %rdx
movq %rdx, (%rsi,%rax)
addq $8, %rax
cmpq %rax, %rcx
jne .L3
.L1: ret
2021-11-26 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* tree-ssa-loop-ivopts.c (cand_value_at): Take a class
tree_niter_desc* argument instead of just a tree for NITER.
If we require the iv candidate value at the end of the final
loop iteration, try using the original loop bound as the
NITER for sufficiently simple loops.
(may_eliminate_iv): Update (only) call to cand_value_at.
gcc/testsuite/ChangeLog
* gcc.dg/wrapped-binop-simplify.c: Update expected test result.
* gcc.dg/tree-ssa/ivopts-5.c: New test case.
* gcc.dg/tree-ssa/ivopts-6.c: New test case.
* gcc.dg/tree-ssa/ivopts-7.c: New test case.
* gcc.dg/tree-ssa/ivopts-8.c: New test case.
* gcc.dg/tree-ssa/ivopts-9.c: New test case.