From: "Jason A. Donenfeld" Jason@zx2c4.com
commit 2122e2a4efc2cd139474079e11939b6e07adfacd upstream.
Currently the clamp algorithm does:
if (val > hi) val = hi; if (val < lo) val = lo;
But since hi > lo by definition, this can be made more efficient with:
if (val > hi) val = hi; else if (val < lo) val = lo;
So fix up the clamp and clamp_t functions to do this, adding the same argument checking as for min and min_t.
For simple cases, code generation on x86_64 and aarch64 stay about the same:
before: cmp edi, edx mov eax, esi cmova edi, edx cmp edi, esi cmovnb eax, edi ret after: cmp edi, esi mov eax, edx cmovnb esi, edi cmp edi, edx cmovb eax, esi ret
before: cmp w0, w2 csel w8, w0, w2, lo cmp w8, w1 csel w0, w8, w1, hi ret after: cmp w0, w1 csel w8, w0, w1, hi cmp w0, w2 csel w0, w8, w2, lo ret
On MIPS64, however, code generation improves, by removing arithmetic in the second branch:
before: sltu $3,$6,$4 bne $3,$0,.L2 move $2,$6
move $2,$4 .L2: sltu $3,$2,$5 bnel $3,$0,.L7 move $2,$5
.L7: jr $31 nop after: sltu $3,$4,$6 beq $3,$0,.L13 move $2,$6
sltu $3,$4,$5 bne $3,$0,.L12 move $2,$4
.L13: jr $31 nop
.L12: jr $31 move $2,$5
For more complex cases with surrounding code, the effects are a bit more complicated. For example, consider this simplified version of timestamp_truncate() from fs/inode.c on x86_64:
struct timespec64 timestamp_truncate(struct timespec64 t, struct inode *inode) { struct super_block *sb = inode->i_sb; unsigned int gran = sb->s_time_gran;
t.tv_sec = clamp(t.tv_sec, sb->s_time_min, sb->s_time_max); if (t.tv_sec == sb->s_time_max || t.tv_sec == sb->s_time_min) t.tv_nsec = 0; return t; }
before: mov r8, rdx mov rdx, rsi mov rcx, QWORD PTR [r8] mov rax, QWORD PTR [rcx+8] mov rcx, QWORD PTR [rcx+16] cmp rax, rdi mov r8, rcx cmovge rdi, rax cmp rdi, rcx cmovle r8, rdi cmp rax, r8 je .L4 cmp rdi, rcx jge .L4 mov rax, r8 ret .L4: xor edx, edx mov rax, r8 ret
after: mov rax, QWORD PTR [rdx] mov rdx, QWORD PTR [rax+8] mov rax, QWORD PTR [rax+16] cmp rax, rdi jg .L6 mov r8, rax xor edx, edx .L2: mov rax, r8 ret .L6: cmp rdx, rdi mov r8, rdi cmovge r8, rdx cmp rax, r8 je .L4 xor eax, eax cmp rdx, rdi cmovl rax, rsi mov rdx, rax mov rax, r8 ret .L4: xor edx, edx jmp .L2
In this case, we actually gain a branch, unfortunately, because the compiler's replacement axioms no longer as cleanly apply.
So all and all, this change is a bit of a mixed bag.
Link: https://lkml.kernel.org/r/20220926133435.1333846-2-Jason@zx2c4.com Signed-off-by: Jason A. Donenfeld Jason@zx2c4.com Cc: Andy Shevchenko andriy.shevchenko@linux.intel.com Cc: Kees Cook keescook@chromium.org Signed-off-by: Andrew Morton akpm@linux-foundation.org (cherry picked from commit 2122e2a4efc2cd139474079e11939b6e07adfacd) Signed-off-by: SeongJae Park sj@kernel.org --- include/linux/minmax.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/include/linux/minmax.h b/include/linux/minmax.h index 8b092c66c5aa..abdeae409dad 100644 --- a/include/linux/minmax.h +++ b/include/linux/minmax.h @@ -38,7 +38,7 @@ __cmp_once(x, y, __UNIQUE_ID(__x), __UNIQUE_ID(__y), op))
#define __clamp(val, lo, hi) \ - __cmp(__cmp(val, lo, >), hi, <) + ((val) >= (hi) ? (hi) : ((val) <= (lo) ? (lo) : (val)))
#define __clamp_once(val, lo, hi, unique_val, unique_lo, unique_hi) ({ \ typeof(val) unique_val = (val); \