https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91202
Bug ID: 91202
Summary: Unnecessary promotion of shift operands
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: ubizjak at gmail dot com
Target Milestone: ---
Following testcase:
--cut here--
unsigned char
foo (unsigned char a, unsigned char b)
{
return a >> b;
}
--cut here--
does not need its operand to be promoted to int on targets that have shift
instruction in QI and HI mode.
Currently, gcc compiles the testcase (-O2) via _.original:
--cut here--
;; Function foo (null)
;; enabled by -tree-original
{
return (unsigned char) ((int) a >> (int) b);
}
--cut here--
and ._optimized:
;; Function foo (foo, funcdef_no=0, decl_uid=1907, cgraph_uid=1,
symbol_order=0)
--cut here--
foo (unsigned char a, unsigned char b)
{
int _1;
int _2;
int _3;
unsigned char _6;
<bb 2> [local count: 1073741824]:
_1 = (int) a_4(D);
_2 = (int) b_5(D);
_3 = _1 >> _2;
_6 = (unsigned char) _3;
return _6;
}
--cut here--
to:
movzbl %dil, %eax
movl %esi, %ecx
sarl %cl, %eax
ret
However, on targets that provide QImode shift, the above code could be emitted
as:
movzbl %dil, %eax
movl %esi, %ecx
shrb %cl, %al
ret
This optimization would help the following testcase:
--cut here--
struct S1
{
unsigned char val;
unsigned char pad1;
unsigned short pad2;
};
struct S1
test_shrb (struct S1 a, unsigned char b)
{
a.val >>= b;
return a;
}
--cut here--
that currently compiles to:
movzbl %dil, %edx
movl %esi, %ecx
movl %edi, %eax
sarl %cl, %edx
movb %dl, %al
ret
but after PR91188 was fixed, the testcase could be compiled with the proposed
optimization to:
movl %edi, %eax
movl %esi, %ecx
shrb %cl, %al
ret
Please note that MSVC v19.21 implements the proposed optimization and
generates:
movzx eax, cl
movzx ecx, dl
shr al, cl
ret 0
