@@ -20,8 +20,9 @@ SIMD := 3dnow sse sse2 sse4 avx avx2 xop
FMA := fma4 fma
SG := avx2-sg avx512f-sg avx512vl-sg
AES := ssse3-aes avx-aes avx2-vaes avx512bw-vaes
+SHA := sse4-sha avx-sha avx512f-sha
GF := sse2-gf avx2-gf avx512bw-gf
-TESTCASES := blowfish $(SIMD) $(FMA) $(SG) $(AES) $(GF)
+TESTCASES := blowfish $(SIMD) $(FMA) $(SG) $(AES) $(SHA) $(GF)
OPMASK := avx512f avx512dq avx512bw
@@ -148,6 +149,10 @@ define simd-aes-defs
$(1)-cflags := $(foreach vec,$($(patsubst %-aes,sse,$(1))-vecs) $($(patsubst %-vaes,%,$(1))-vecs), \
"-D_$(vec) -maes $(addprefix -m,$(subst -,$(space),$(1))) $(call non-sse,$(1)) -Os -DVEC_SIZE=$(vec)")
endef
+define simd-sha-defs
+$(1)-cflags := $(foreach vec,$(sse-vecs), \
+ "-D_$(vec) $(addprefix -m,$(subst -,$(space),$(1))) -Os -DVEC_SIZE=$(vec)")
+endef
define simd-gf-defs
$(1)-cflags := $(foreach vec,$($(1:-gf=)-vecs), \
"-D_$(vec) -mgfni -m$(1:-gf=) $(call non-sse,$(1)) -Os -DVEC_SIZE=$(vec)")
@@ -159,6 +164,7 @@ endef
$(foreach flavor,$(SIMD) $(FMA),$(eval $(call simd-defs,$(flavor))))
$(foreach flavor,$(SG),$(eval $(call simd-sg-defs,$(flavor))))
$(foreach flavor,$(AES),$(eval $(call simd-aes-defs,$(flavor))))
+$(foreach flavor,$(SHA),$(eval $(call simd-sha-defs,$(flavor))))
$(foreach flavor,$(GF),$(eval $(call simd-gf-defs,$(flavor))))
$(foreach flavor,$(OPMASK),$(eval $(call opmask-defs,$(flavor))))
@@ -212,10 +218,13 @@ $(addsuffix .c,$(SG)):
$(addsuffix .c,$(AES)):
ln -sf simd-aes.c $@
+$(addsuffix .c,$(SHA)):
+ ln -sf simd-sha.c $@
+
$(addsuffix .c,$(GF)):
ln -sf simd-gf.c $@
-$(addsuffix .h,$(SIMD) $(FMA) $(SG) $(AES) $(GF)): simd.h
+$(addsuffix .h,$(SIMD) $(FMA) $(SG) $(AES) $(SHA) $(GF)): simd.h
xop.h avx512f.h: simd-fma.c
@@ -0,0 +1,392 @@
+#define INT_SIZE 4
+
+#include "simd.h"
+ENTRY(sha_test);
+
+#define SHA(op, a...) __builtin_ia32_sha ## op(a)
+
+#ifdef __AVX512F__
+# define ALL_TRUE (~0ULL >> (64 - ELEM_COUNT))
+# define eq(x, y) (B(pcmpeqd, _mask, x, y, -1) == ALL_TRUE)
+# define blend(x, y, sel) B(movdqa32_, _mask, y, x, sel)
+# define rot_c(f, r, x, n) B(pro ## f ## d, _mask, x, n, undef(), ~0)
+# define rot_s(f, r, x, n) ({ /* gcc does not support embedded broadcast */ \
+ vec_t r_; \
+ asm ( "vpro" #f "vd %2%{1to%c3%}, %1, %0" \
+ : "=v" (r_) \
+ : "v" (x), "m" (n), "i" (ELEM_COUNT) ); \
+ r_; \
+})
+# define rot_v(d, x, n) B(pro ## d ## vd, _mask, x, n, undef(), ~0)
+# define shift_s(d, x, n) ({ \
+ vec_t r_; \
+ asm ( "vps" #d "lvd %2%{1to%c3%}, %1, %0" \
+ : "=v" (r_) \
+ : "v" (x), "m" (n), "i" (ELEM_COUNT) ); \
+ r_; \
+})
+# define vshift(d, x, n) ({ /* gcc does not allow memory operands */ \
+ vec_t r_; \
+ asm ( "vps" #d "ldq %2, %1, %0" \
+ : "=v" (r_) : "m" (x), "i" ((n) * ELEM_SIZE) ); \
+ r_; \
+})
+#else
+# define to_bool(cmp) (__builtin_ia32_pmovmskb128(cmp) == 0xffff)
+# define eq(x, y) to_bool((x) == (y))
+# define blend(x, y, sel) \
+ ((vec_t)__builtin_ia32_pblendw128((vhi_t)(x), (vhi_t)(y), \
+ ((sel) & 1 ? 0x03 : 0) | \
+ ((sel) & 2 ? 0x0c : 0) | \
+ ((sel) & 4 ? 0x30 : 0) | \
+ ((sel) & 8 ? 0xc0 : 0)))
+# define rot_c(f, r, x, n) (sh ## f ## _c(x, n) | sh ## r ## _c(x, 32 - (n)))
+# define rot_s(f, r, x, n) ({ /* gcc does not allow memory operands */ \
+ vec_t r_, t_, n_ = (vec_t){ 32 } - (n); \
+ asm ( "ps" #f "ld %2, %0; ps" #r "ld %3, %1; por %1, %0" \
+ : "=&x" (r_), "=&x" (t_) \
+ : "m" (n), "m" (n_), "0" (x), "1" (x) ); \
+ r_; \
+})
+static inline unsigned int rotl(unsigned int x, unsigned int n)
+{
+ return (x << (n & 0x1f)) | (x >> ((32 - n) & 0x1f));
+}
+static inline unsigned int rotr(unsigned int x, unsigned int n)
+{
+ return (x >> (n & 0x1f)) | (x << ((32 - n) & 0x1f));
+}
+# define rot_v(d, x, n) ({ \
+ vec_t t_; \
+ unsigned int i_; \
+ for ( i_ = 0; i_ < ELEM_COUNT; ++i_ ) \
+ t_[i_] = rot ## d((x)[i_], (n)[i_]); \
+ t_; \
+})
+# define shift_s(d, x, n) ({ \
+ vec_t r_; \
+ asm ( "ps" #d "ld %1, %0" : "=&x" (r_) : "m" (n), "0" (x) ); \
+ r_; \
+})
+# define vshift(d, x, n) \
+ (vec_t)(__builtin_ia32_ps ## d ## ldqi128((vdi_t)(x), (n) * ELEM_SIZE * 8))
+#endif
+
+#define alignr(x, y, n) ((vec_t)__builtin_ia32_palignr128((vdi_t)(x), (vdi_t)(y), (n) * 8))
+#define hadd(x, y) __builtin_ia32_phaddd128(x, y)
+#define rol_c(x, n) rot_c(l, r, x, n)
+#define rol_s(x, n) rot_s(l, r, x, n)
+#define rol_v(x, n...) rot_v(l, x, n)
+#define ror_c(x, n) rot_c(r, l, x, n)
+#define ror_s(x, n) rot_s(r, l, x, n)
+#define ror_v(x, n...) rot_v(r, x, n)
+#define shl_c(x, n) __builtin_ia32_pslldi128(x, n)
+#define shl_s(x, n) shift_s(l, x, n)
+#define shr_c(x, n) __builtin_ia32_psrldi128(x, n)
+#define shr_s(x, n) shift_s(r, x, n)
+#define shuf(x, s) __builtin_ia32_pshufd(x, s)
+#define swap(x) shuf(x, 0b00011011)
+#define vshl(x, n) vshift(l, x, n)
+#define vshr(x, n) vshift(r, x, n)
+
+static inline vec_t sha256_sigma0(vec_t w)
+{
+ vec_t res;
+
+ touch(w);
+ res = ror_c(w, 7);
+ touch(w);
+ res ^= rol_c(w, 14);
+ touch(w);
+ res ^= shr_c(w, 3);
+ touch(w);
+
+ return res;
+}
+
+static inline vec_t sha256_sigma1(vec_t w)
+{
+ vec_t _17 = { 17 }, _19 = { 19 }, _10 = { 10 };
+
+ return ror_s(w, _17) ^ ror_s(w, _19) ^ shr_s(w, _10);
+}
+
+static inline vec_t sha256_Sigma0(vec_t w)
+{
+ vec_t res, n1 = { 0, 0, 2, 2 }, n2 = { 0, 0, 13, 13 }, n3 = { 0, 0, 10, 10 };
+
+ touch(n1);
+ res = ror_v(w, n1);
+ touch(n2);
+ res ^= ror_v(w, n2);
+ touch(n3);
+
+ return res ^ rol_v(w, n3);
+}
+
+static inline vec_t sha256_Sigma1(vec_t w)
+{
+ return ror_c(w, 6) ^ ror_c(w, 11) ^ rol_c(w, 7);
+}
+
+int sha_test(void)
+{
+ unsigned int i;
+ vec_t src, one = { 1 };
+ vqi_t raw = {};
+
+ for ( i = 1; i < VEC_SIZE; ++i )
+ raw[i] = i;
+ src = (vec_t)raw;
+
+ for ( i = 0; i < 256; i += VEC_SIZE )
+ {
+ vec_t x, y, tmp, hash = -src;
+ vec_t a, b, c, d, e, g, h;
+ unsigned int k, r;
+
+ touch(src);
+ x = SHA(1msg1, hash, src);
+ touch(src);
+ y = hash ^ alignr(hash, src, 8);
+ touch(src);
+
+ if ( !eq(x, y) ) return __LINE__;
+
+ touch(src);
+ x = SHA(1msg2, hash, src);
+ touch(src);
+ tmp = hash ^ alignr(src, hash, 12);
+ touch(tmp);
+ y = rol_c(tmp, 1);
+ tmp = hash ^ alignr(src, y, 12);
+ touch(tmp);
+ y = rol_c(tmp, 1);
+
+ if ( !eq(x, y) ) return __LINE__;
+
+ touch(src);
+ x = SHA(1msg2, hash, src);
+ touch(src);
+ tmp = rol_s(hash ^ alignr(src, hash, 12), one);
+ y = rol_s(hash ^ alignr(src, tmp, 12), one);
+
+ if ( !eq(x, y) ) return __LINE__;
+
+ touch(src);
+ x = SHA(1nexte, hash, src);
+ touch(src);
+ touch(hash);
+ tmp = rol_c(hash, 30);
+ tmp[2] = tmp[1] = tmp[0] = 0;
+
+ if ( !eq(x, src + tmp) ) return __LINE__;
+
+ /*
+ * SHA1RNDS4
+ *
+ * SRC1 = { A0, B0, C0, D0 }
+ * SRC2 = W' = { W[0]E0, W[1], W[2], W[3] }
+ *
+ * (NB that the notation is not C-like, i.e. elements are listed
+ * high-to-low everywhere in this comment.)
+ *
+ * In order to pick a simple rounds function, an immediate value of
+ * 1 is used; 3 would also be a possibility.
+ *
+ * Applying
+ *
+ * A1 = ROL5(A0) + (B0 ^ C0 ^ D0) + W'[0] + K
+ * E1 = D0
+ * D1 = C0
+ * C1 = ROL30(B0)
+ * B1 = A0
+ *
+ * iteratively four times and resolving round variable values to
+ * A<n> and B0, C0, and D0 we get
+ *
+ * A4 = ROL5(A3) + (A2 ^ ROL30(A1) ^ ROL30(A0)) + W'[3] + ROL30(B0) + K
+ * A3 = ROL5(A2) + (A1 ^ ROL30(A0) ^ ROL30(B0)) + W'[2] + C0 + K
+ * A2 = ROL5(A1) + (A0 ^ ROL30(B0) ^ C0 ) + W'[1] + D0 + K
+ * A1 = ROL5(A0) + (B0 ^ C0 ^ D0 ) + W'[0] + K
+ *
+ * (respective per-column variable names:
+ * y a b c d src e k
+ * )
+ *
+ * with
+ *
+ * B4 = A3
+ * C4 = ROL30(A2)
+ * D4 = ROL30(A1)
+ * E4 = ROL30(A0)
+ *
+ * and hence
+ *
+ * DST = { A4, A3, ROL30(A2), ROL30(A1) }
+ */
+
+ touch(src);
+ x = SHA(1rnds4, hash, src, 1);
+ touch(src);
+
+ a = vshr(hash, 3);
+ b = vshr(hash, 2);
+ touch(hash);
+ d = rol_c(hash, 30);
+ touch(hash);
+ d = blend(d, hash, 0b0011);
+ c = vshr(d, 1);
+ e = vshl(d, 1);
+ tmp = (vec_t){};
+ k = rol_c(SHA(1rnds4, tmp, tmp, 1), 2)[0];
+
+ for ( r = 0; r < 4; ++r )
+ {
+ y = rol_c(a, 5) + (b ^ c ^ d) + swap(src) + e + k;
+
+ switch ( r )
+ {
+ case 0:
+ c[3] = rol_c(y, 30)[0];
+ /* fall through */
+ case 1:
+ b[r + 2] = y[r];
+ /* fall through */
+ case 2:
+ a[r + 1] = y[r];
+ break;
+ }
+
+ switch ( r )
+ {
+ case 3:
+ if ( a[3] != y[2] ) return __LINE__;
+ /* fall through */
+ case 2:
+ if ( a[2] != y[1] ) return __LINE__;
+ if ( b[3] != y[1] ) return __LINE__;
+ /* fall through */
+ case 1:
+ if ( a[1] != y[0] ) return __LINE__;
+ if ( b[2] != y[0] ) return __LINE__;
+ if ( c[3] != rol_c(y, 30)[0] ) return __LINE__;
+ break;
+ }
+ }
+
+ a = blend(rol_c(y, 30), y, 0b1100);
+
+ if ( !eq(x, a) ) return __LINE__;
+
+ touch(src);
+ x = SHA(256msg1, hash, src);
+ touch(src);
+ y = hash + sha256_sigma0(alignr(src, hash, 4));
+
+ if ( !eq(x, y) ) return __LINE__;
+
+ touch(src);
+ x = SHA(256msg2, hash, src);
+ touch(src);
+ tmp = hash + sha256_sigma1(alignr(hash, src, 8));
+ y = hash + sha256_sigma1(alignr(tmp, src, 8));
+
+ if ( !eq(x, y) ) return __LINE__;
+
+ /*
+ * SHA256RNDS2
+ *
+ * SRC1 = { C0, D0, G0, H0 }
+ * SRC2 = { A0, B0, E0, F0 }
+ * XMM0 = W' = { ?, ?, WK1, WK0 }
+ *
+ * (NB that the notation again is not C-like, i.e. elements are listed
+ * high-to-low everywhere in this comment.)
+ *
+ * Ch(E,F,G) = (E & F) ^ (~E & G)
+ * Maj(A,B,C) = (A & B) ^ (A & C) ^ (B & C)
+ *
+ * Σ0(A) = ROR2(A) ^ ROR13(A) ^ ROR22(A)
+ * Σ1(E) = ROR6(E) ^ ROR11(E) ^ ROR25(E)
+ *
+ * Applying
+ *
+ * A1 = Ch(E0, F0, G0) + Σ1(E0) + WK0 + H0 + Maj(A0, B0, C0) + Σ0(A0)
+ * B1 = A0
+ * C1 = B0
+ * D1 = C0
+ * E1 = Ch(E0, F0, G0) + Σ1(E0) + WK0 + H0 + D0
+ * F1 = E0
+ * G1 = F0
+ * H1 = G0
+ *
+ * iteratively four times and resolving round variable values to
+ * A<n> / E<n> and B0, C0, D0, F0, G0, and H0 we get
+ *
+ * A2 = Ch(E1, E0, F0) + Σ1(E1) + WK1 + G0 + Maj(A1, A0, B0) + Σ0(A1)
+ * A1 = Ch(E0, F0, G0) + Σ1(E0) + WK0 + H0 + Maj(A0, B0, C0) + Σ0(A0)
+ * E2 = Ch(E1, E0, F0) + Σ1(E1) + WK1 + G0 + C0
+ * E1 = Ch(E0, F0, G0) + Σ1(E0) + WK0 + H0 + D0
+ *
+ * with
+ *
+ * B2 = A1
+ * F2 = E1
+ *
+ * and hence
+ *
+ * DST = { A2, A1, E2, E1 }
+ *
+ * which we can simplify a little, by letting A0, B0, and E0 be zero
+ * and F0 = ~G0, and by then utilizing
+ *
+ * Ch(0, 0, x) = x
+ * Ch(x, 0, y) = ~x & y
+ * Maj(x, 0, 0) = Maj(0, x, 0) = Maj(0, 0, x) = 0
+ *
+ * A2 = (~E1 & F0) + Σ1(E1) + WK1 + G0 + Σ0(A1)
+ * A1 = (~E0 & G0) + Σ1(E0) + WK0 + H0 + Σ0(A0)
+ * E2 = (~E1 & F0) + Σ1(E1) + WK1 + G0 + C0
+ * E1 = (~E0 & G0) + Σ1(E0) + WK0 + H0 + D0
+ *
+ * (respective per-column variable names:
+ * y e g e src h d
+ * )
+ */
+
+ tmp = (vec_t){ ~hash[1] };
+ touch(tmp);
+ x = SHA(256rnds2, hash, tmp, src);
+ touch(tmp);
+
+ e = y = (vec_t){};
+ d = alignr(y, hash, 8);
+ g = (vec_t){ hash[1], tmp[0], hash[1], tmp[0] };
+ h = shuf(hash, 0b01000100);
+
+ for ( r = 0; r < 2; ++r )
+ {
+ y = (~e & g) + sha256_Sigma1(e) + shuf(src, 0b01000100) +
+ h + sha256_Sigma0(d);
+
+ if ( !r )
+ {
+ d[3] = y[2];
+ e[3] = e[1] = y[0];
+ }
+ else if ( d[3] != y[2] )
+ return __LINE__;
+ else if ( e[1] != y[0] )
+ return __LINE__;
+ else if ( e[3] != y[0] )
+ return __LINE__;
+ }
+
+ if ( !eq(x, y) ) return __LINE__;
+
+ src += 0x01010101 * VEC_SIZE;
+ }
+
+ return 0;
+}
@@ -14,8 +14,10 @@ asm ( ".pushsection .test, \"ax\", @prog
#include "sse2-gf.h"
#include "ssse3-aes.h"
#include "sse4.h"
+#include "sse4-sha.h"
#include "avx.h"
#include "avx-aes.h"
+#include "avx-sha.h"
#include "fma4.h"
#include "fma.h"
#include "avx2.h"
@@ -28,6 +30,7 @@ asm ( ".pushsection .test, \"ax\", @prog
#include "avx512bw-opmask.h"
#include "avx512f.h"
#include "avx512f-sg.h"
+#include "avx512f-sha.h"
#include "avx512vl-sg.h"
#include "avx512bw.h"
#include "avx512bw-vaes.h"
@@ -155,6 +158,21 @@ static bool simd_check_avx512vbmi_vl(voi
return cpu_has_avx512_vbmi && cpu_has_avx512vl;
}
+static bool simd_check_sse4_sha(void)
+{
+ return cpu_has_sha && cpu_has_sse4_2;
+}
+
+static bool simd_check_avx_sha(void)
+{
+ return cpu_has_sha && cpu_has_avx;
+}
+
+static bool simd_check_avx512f_sha_vl(void)
+{
+ return cpu_has_sha && cpu_has_avx512vl;
+}
+
static bool simd_check_avx2_vaes(void)
{
return cpu_has_aesni && cpu_has_vaes && cpu_has_avx2;
@@ -450,6 +468,9 @@ static const struct {
AVX512VL(_VBMI+VL u16x8, avx512vbmi, 16u2),
AVX512VL(_VBMI+VL s16x16, avx512vbmi, 32i2),
AVX512VL(_VBMI+VL u16x16, avx512vbmi, 32u2),
+ SIMD(SHA, sse4_sha, 16),
+ SIMD(AVX+SHA, avx_sha, 16),
+ AVX512VL(VL+SHA, avx512f_sha, 16),
SIMD(VAES (VEX/x32), avx2_vaes, 32),
SIMD(VAES (EVEX/x64), avx512bw_vaes, 64),
AVX512VL(VL+VAES (x16), avx512bw_vaes, 16),
@@ -142,6 +142,7 @@ static inline bool xcr0_mask(uint64_t ma
#define cpu_has_avx512_ifma (cp.feat.avx512_ifma && xcr0_mask(0xe6))
#define cpu_has_avx512er (cp.feat.avx512er && xcr0_mask(0xe6))
#define cpu_has_avx512cd (cp.feat.avx512cd && xcr0_mask(0xe6))
+#define cpu_has_sha cp.feat.sha
#define cpu_has_avx512bw (cp.feat.avx512bw && xcr0_mask(0xe6))
#define cpu_has_avx512vl (cp.feat.avx512vl && xcr0_mask(0xe6))
#define cpu_has_avx512_vbmi (cp.feat.avx512_vbmi && xcr0_mask(0xe6))
Also use this for AVX512VL VPRO{L,R}{,V}D as well as some further shifts testing. Signed-off-by: Jan Beulich <jbeulich@suse.com> --- v8: New.