On Tue, Apr 12, 2022 at 05:28:12PM +0000, Nathan Huckleberry wrote: > diff --git a/arch/x86/crypto/aes_ctrby8_avx-x86_64.S b/arch/x86/crypto/aes_ctrby8_avx-x86_64.S > index 43852ba6e19c..9e20d7d3d6da 100644 > --- a/arch/x86/crypto/aes_ctrby8_avx-x86_64.S > +++ b/arch/x86/crypto/aes_ctrby8_avx-x86_64.S > @@ -53,6 +53,10 @@ > #define KEY_192 2 > #define KEY_256 3 > > +// XCTR mode only > +#define counter %r9 > +#define xiv %xmm8 > + It would be helpful if the registers were listed in order, and if the CTR-specific ones were marked as being specific to CTR. This would make it easy to verify that there are no collisions in register allocation. I.e.: [...] #define xdata7 %xmm7 #define xcounter %xmm8 // CTR mode only #define xiv %xmm8 // XCTR mode only #define xbyteswap %xmm9 // CTR mode only #define xkey0 %xmm10 [...] #define num_bytes %r8 #define counter %r9 // XCTR mode only #define tmp %r10 [...] I'm also not a fan of the naming, with "xcounter" being used by CTR only and "counter" being used by XCTR only... I see why you did it, though, as the existing code uses the "x" prefix to mean "this is an xmm register". It could at least use a comment that makes this super clear, though: // Note: the "x" prefix in these aliases means "this is an xmm register". // No relation to XCTR where the "X" prefix means "XOR counter". #define xdata0 %xmm0 > + .if (\xctr == 1) As \xctr is either 0 or 1, this can be written as simply '.if \xctr' > + .set i, 0 > + .rept (by) > + club XDATA, i > + movq counter, var_xdata > + .set i, (i +1) > + .endr > + .endif > + Since the 3-operand add instruction (vpaddq) is available here, and in fact is being used already, it isn't necessary to move 'counter' into all (up to 8) of the var_xdata registers. Just move it into the last var_xdata register, or into a temporary register, and use it as a source operand for all the additions. > - vpshufb xbyteswap, xcounter, xdata0 > - > - .set i, 1 > - .rept (by - 1) > - club XDATA, i > - vpaddq (ddq_add_1 + 16 * (i - 1))(%rip), xcounter, var_xdata > - vptest ddq_low_msk(%rip), var_xdata > - jnz 1f > - vpaddq ddq_high_add_1(%rip), var_xdata, var_xdata > - vpaddq ddq_high_add_1(%rip), xcounter, xcounter > - 1: > - vpshufb xbyteswap, var_xdata, var_xdata > - .set i, (i +1) > - .endr > + .if (\xctr == 0) > + vpshufb xbyteswap, xcounter, xdata0 > + .set i, 1 > + .rept (by - 1) > + club XDATA, i > + vpaddq (ddq_add_1 + 16 * (i - 1))(%rip), xcounter, var_xdata > + vptest ddq_low_msk(%rip), var_xdata > + jnz 1f > + vpaddq ddq_high_add_1(%rip), var_xdata, var_xdata > + vpaddq ddq_high_add_1(%rip), xcounter, xcounter > + 1: > + vpshufb xbyteswap, var_xdata, var_xdata > + .set i, (i +1) > + .endr > + .endif > + .if (\xctr == 1) > + .set i, 0 > + .rept (by) > + club XDATA, i > + vpaddq (ddq_add_1 + 16 * i)(%rip), var_xdata, var_xdata > + .set i, (i +1) > + .endr > + .set i, 0 > + .rept (by) > + club XDATA, i > + vpxor xiv, var_xdata, var_xdata > + .set i, (i +1) > + .endr > + .endif This can be written as: .if \xctr [second part above] .else [first part above] .endif > - vpaddq (ddq_add_1 + 16 * (by - 1))(%rip), xcounter, xcounter > - vptest ddq_low_msk(%rip), xcounter > - jnz 1f > - vpaddq ddq_high_add_1(%rip), xcounter, xcounter > - 1: > + .if (\xctr == 0) > + vpaddq (ddq_add_1 + 16 * (by - 1))(%rip), xcounter, xcounter > + vptest ddq_low_msk(%rip), xcounter > + jnz 1f > + vpaddq ddq_high_add_1(%rip), xcounter, xcounter > + 1: > + .endif > + .if (\xctr == 1) > + add $by, counter > + .endif Likewise here. > +.macro do_aes_ctrmain key_len, xctr > cmp $16, num_bytes > - jb .Ldo_return2\key_len > + jb .Ldo_return2\xctr\key_len > > vmovdqa byteswap_const(%rip), xbyteswap > - vmovdqu (p_iv), xcounter > - vpshufb xbyteswap, xcounter, xcounter > + .if (\xctr == 0) > + vmovdqu (p_iv), xcounter > + vpshufb xbyteswap, xcounter, xcounter > + .endif > + .if (\xctr == 1) > + andq $(~0xf), num_bytes > + shr $4, counter > + vmovdqu (p_iv), xiv > + .endif And likewise here. Also, the load of byteswap_const can be moved into the !\xctr block. - Eric
