https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101139
Bug ID: 101139
Summary: Unable to remove double byteswap in fast path
Product: gcc
Version: 10.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: steinar+gcc at gunderson dot no
Target Milestone: ---
The following code is reduced from a real interpreter:
extern void (*a[])();
int d, e, h, l;
typedef struct {
char ab;
} f;
f g;
short i();
short m68ki_read_imm_16() {
short j, k;
int b = d;
f f = g;
if (b < h)
return __builtin_bswap16((&f.ab)[0]);
k = i();
short c = k;
j = __builtin_bswap16(c);
return j;
}
int b() {
short m;
do {
m = m68ki_read_imm_16();
short c = m;
l = __builtin_bswap16(c);
a[l]();
} while (e);
return e;
}
Compiling with arm-linux-gnueabihf-gcc-10 -O2 yields this interesting sequence
in the function:
b .L11
.L15:
ldrb r3, [r5, #8] @ zero_extendqisi2
rev16 r3, r3
uxth r3, r3
.L10:
rev16 r3, r3
uxth r3, r3
The original code intention was to have a reusable function that returned in
big-endian, but that a specific use of it would be able to ignore endianness
into a table lookup, removing the double-swap entirely. GCC can normally do
that, but it seems that the branch in m68ki_read_imm_16() somehow gets in the
way. Just to be clear, I expect zero rev16 instructions altogether in b() when
m68ki_read_imm_16() is inlined.
The problem is not ARM-specific; x86 shows a similar problematic sequence:
leaq a(%rip), %rbx
jmp .L11
.p2align 4,,10
.p2align 3
.L15:
movsbw g(%rip), %ax
rolw $8, %ax
.L10:
rolw $8, %ax
movzwl %ax, %edx
Also verified with
gcc version 12.0.0 20210527 (experimental) [master revision
262e75d22c3:7bb6b9b2f47:9d3a953ec4d2695e9a6bfa5f22655e2aea47a973] (Debian
20210527-1)
