| Summary: | Missing simple optimization | ||
|---|---|---|---|
| Product: | gcc | Reporter: | matt |
| Component: | middle-end | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | NEW --- | ||
| Severity: | enhancement | CC: | gcc-bugs, rguenth |
| Priority: | P3 | Keywords: | missed-optimization |
| Version: | 4.3.0 | ||
| Target Milestone: | 4.7.0 | ||
| Host: | x86_64--netbsd | Target: | x86_64--netbsd |
| Build: | x86_64--netbsd | Known to work: | |
| Known to fail: | 4.6.4 | Last reconfirmed: | 2007-03-20 10:39:11 |
Confirmed. This is neither done at the tree nor at the rtl level. We produce in 4.7.0+
in_canforward(unsigned int):
.LFB0:
.cfi_startproc
andl $224, %edi
xorl %eax, %eax
cmpl $224, %edi
setne %al
ret
That is:
D.2201_1 = in_2(D) & 224;
D.2199_10 = D.2201_1 != 224;
I think we could do slightly better
((~in_2(D)) & 224) == 0
But only at exand time.
This gives:
notl %edi
xorl %eax, %eax
testb $-32, %dil
setne %al
Or for aarch64:
mov w8, #224
bics wzr, w8, w0
cset w0, ne
ret
(In reply to Andrew Pinski from comment #2) > > I think we could do slightly better > ((~in_2(D)) & 224) == 0 > > But only at exand time. > This gives: > notl %edi > xorl %eax, %eax > testb $-32, %dil > setne %al x86_64 produces that in GCC 13 with r13-792-g29ae455901ac71 . > > Or for aarch64: > mov w8, #224 > bics wzr, w8, w0 > cset w0, ne > ret For aarch64, it could define an instruction to catch: (set (reg:CC_NZV 66 cc) (compare:CC_NZV (and:SI (not:SI (reg:SI 100)) (const_int 224 [0xe0])) (const_int 0 [0]))) Anyways the original issue was fixed in GCC 4.7.0 and the small improvement for x86_64 is in GCC 13. The aarch64 code generation is currently: and w0, w0, 224 cmp w0, 224 cset w0, ne ret Which is only slightly worse than what I proposed too. |
The following shows a missing easy optimization for GCC: int in_canforward(unsigned int in) { if ((in & ~0xffffff0f) == 0xf0 || (in & ~0xffffff0f) == 0xe0) return 0; return 1; } results in (@ -O2): in_canforward: andl $240, %edi cmpl $240, %edi sete %al cmpl $224, %edi sete %dl orl %edx, %eax xorl $1, %eax movzbl %al, %eax ret given that 0xf0 and 0xe0 only differ by one bit, there is no reason to test for that bit so the comparision could be: (in & 0xffffff1f) == 0xe0. More generally the optimization is: given (x & m) == a0 || (x & m) == a1 where m, a0, and a1 are all constant let b = (a0 ^ a1) then if (b & (b - 1)) == 0 [b is a power of 2] rewrite to: (x & (m|b)) == (a0 & ~b)