http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57477
Bug ID: 57477
Summary: gcc generates suboptimal code for a simple
and-shift-zeroextend combination on x86_64
Product: gcc
Version: 4.9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: mikpe at it dot uu.se
Consider the following set of trivial functions:
> cat q.c
unsigned int g(unsigned int x) { return x & 0x1f; }
unsigned long f(unsigned int x) { return x << 4; }
unsigned long h(unsigned int x) { return (x & 0x1f) << 4; }
h(x) == f(g(x)).
The code generated for f and g is good (not much choice there), but the code
for h contains some (suboptimal) surprises:
> gcc -O3 -c q.c ; objdump -d q.o
q.o: file format elf64-x86-64
Disassembly of section .text:
0000000000000000 <g>:
0: 89 f8 mov %edi,%eax
2: 83 e0 1f and $0x1f,%eax
5: c3 retq
6: 66 2e 0f 1f 84 00 00 nopw %cs:0x0(%rax,%rax,1)
d: 00 00 00
0000000000000010 <f>:
10: c1 e7 04 shl $0x4,%edi
13: 89 f8 mov %edi,%eax
15: c3 retq
16: 66 2e 0f 1f 84 00 00 nopw %cs:0x0(%rax,%rax,1)
1d: 00 00 00
0000000000000020 <h>:
20: 48 89 f8 mov %rdi,%rax
23: 48 c1 e0 04 shl $0x4,%rax
27: 25 f0 01 00 00 and $0x1f0,%eax
2c: c3 retq
1. In h gcc exchanged the order of the '&' and the '<<', forcing it to use a
larger 4-byte immediate where g could use a 1-byte immediate, resulting in a
2-byte larger instruction encoding.
2. In h the '<<' is done in 64-bit precision, even though the input clearly is
32-bit ('unsigned int'), and the result clearly also is 32-bit (notice the
absence of a REX.W on the 'and'), resulting in 1-byte larger instruction
encoding.
3. In h the move from rdi to rax is unavoidable (due to the ABI), but it too is
redundantly done in 64-bit precision where 32-bit precision would have
sufficed, resulting in a 1-byte larger instruction encoding.
In short, h compiles to 13 bytes but could have compiled to 9 bytes.
This is with gcc 4.9, but 4.8 and 4.7 generate identical code.