libbb/sha1: shrink and speed up unrolled x86-64 code

function                                             old     new   delta
sha1_process_block64                                3514    3482     -32

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
diff --git a/libbb/hash_md5_sha256_x86-32_shaNI.S b/libbb/hash_md5_sha256_x86-32_shaNI.S
index 417da37..39e2baf 100644
--- a/libbb/hash_md5_sha256_x86-32_shaNI.S
+++ b/libbb/hash_md5_sha256_x86-32_shaNI.S
@@ -257,8 +257,8 @@
 	ret
 	.size	sha256_process_block64_shaNI, .-sha256_process_block64_shaNI
 
-.section	.rodata.cst256.K256, "aM", @progbits, 256
-.balign 16
+	.section	.rodata.cst256.K256, "aM", @progbits, 256
+	.balign 16
 K256:
 	.long	0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
 	.long	0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
@@ -277,8 +277,8 @@
 	.long	0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
 	.long	0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
 
-.section	.rodata.cst16.PSHUFFLE_BSWAP32_FLIP_MASK, "aM", @progbits, 16
-.balign 16
+	.section	.rodata.cst16.PSHUFFLE_BSWAP32_FLIP_MASK, "aM", @progbits, 16
+	.balign 16
 PSHUFFLE_BSWAP32_FLIP_MASK:
 	.octa 0x0c0d0e0f08090a0b0405060700010203
 
diff --git a/libbb/hash_md5_sha256_x86-64_shaNI.S b/libbb/hash_md5_sha256_x86-64_shaNI.S
index dbf3911..c6c9313 100644
--- a/libbb/hash_md5_sha256_x86-64_shaNI.S
+++ b/libbb/hash_md5_sha256_x86-64_shaNI.S
@@ -253,8 +253,8 @@
 	ret
 	.size	sha256_process_block64_shaNI, .-sha256_process_block64_shaNI
 
-.section	.rodata.cst256.K256, "aM", @progbits, 256
-.balign 16
+	.section	.rodata.cst256.K256, "aM", @progbits, 256
+	.balign 16
 K256:
 	.long	0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
 	.long	0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
@@ -273,8 +273,8 @@
 	.long	0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
 	.long	0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
 
-.section	.rodata.cst16.PSHUFFLE_BSWAP32_FLIP_MASK, "aM", @progbits, 16
-.balign 16
+	.section	.rodata.cst16.PSHUFFLE_BSWAP32_FLIP_MASK, "aM", @progbits, 16
+	.balign 16
 PSHUFFLE_BSWAP32_FLIP_MASK:
 	.octa 0x0c0d0e0f08090a0b0405060700010203
 
diff --git a/libbb/hash_md5_sha_x86-32_shaNI.S b/libbb/hash_md5_sha_x86-32_shaNI.S
index 11b855e..5d082eb 100644
--- a/libbb/hash_md5_sha_x86-32_shaNI.S
+++ b/libbb/hash_md5_sha_x86-32_shaNI.S
@@ -223,8 +223,8 @@
 	ret
 	.size	sha1_process_block64_shaNI, .-sha1_process_block64_shaNI
 
-.section	.rodata.cst16.PSHUFFLE_BYTE_FLIP_MASK, "aM", @progbits, 16
-.balign 16
+	.section	.rodata.cst16.PSHUFFLE_BYTE_FLIP_MASK, "aM", @progbits, 16
+	.balign 16
 PSHUFFLE_BYTE_FLIP_MASK:
 	.octa 0x000102030405060708090a0b0c0d0e0f
 
diff --git a/libbb/hash_md5_sha_x86-64.S b/libbb/hash_md5_sha_x86-64.S
index 47ace60..e26c46f 100644
--- a/libbb/hash_md5_sha_x86-64.S
+++ b/libbb/hash_md5_sha_x86-64.S
@@ -180,8 +180,13 @@
 # PREP %xmm0 %xmm1 %xmm2 %xmm3 -64+16*0(%rsp)
 	movaps	%xmm3, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm0, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm1, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm0, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm1, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm0, %xmm5
+	shufps	$0x4e, %xmm1, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm2, %xmm0	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm0	# ^
@@ -252,8 +257,13 @@
 # PREP %xmm1 %xmm2 %xmm3 %xmm0 -64+16*1(%rsp)
 	movaps	%xmm0, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm1, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm2, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm1, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm2, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm1, %xmm5
+	shufps	$0x4e, %xmm2, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm3, %xmm1	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm1	# ^
@@ -323,8 +333,13 @@
 # PREP %xmm2 %xmm3 %xmm0 %xmm1 -64+16*2(%rsp)
 	movaps	%xmm1, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm2, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm3, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm2, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm3, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm2, %xmm5
+	shufps	$0x4e, %xmm3, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm0, %xmm2	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm2	# ^
@@ -392,8 +407,13 @@
 # PREP %xmm3 %xmm0 %xmm1 %xmm2 -64+16*3(%rsp)
 	movaps	%xmm2, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm3, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm0, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm3, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm0, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm3, %xmm5
+	shufps	$0x4e, %xmm0, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm1, %xmm3	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm3	# ^
@@ -457,8 +477,13 @@
 # PREP %xmm0 %xmm1 %xmm2 %xmm3 -64+16*0(%rsp)
 	movaps	%xmm3, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm0, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm1, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm0, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm1, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm0, %xmm5
+	shufps	$0x4e, %xmm1, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm2, %xmm0	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm0	# ^
@@ -522,8 +547,13 @@
 # PREP %xmm1 %xmm2 %xmm3 %xmm0 -64+16*1(%rsp)
 	movaps	%xmm0, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm1, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm2, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm1, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm2, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm1, %xmm5
+	shufps	$0x4e, %xmm2, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm3, %xmm1	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm1	# ^
@@ -588,8 +618,13 @@
 # PREP %xmm2 %xmm3 %xmm0 %xmm1 -64+16*2(%rsp)
 	movaps	%xmm1, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm2, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm3, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm2, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm3, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm2, %xmm5
+	shufps	$0x4e, %xmm3, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm0, %xmm2	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm2	# ^
@@ -653,8 +688,13 @@
 # PREP %xmm3 %xmm0 %xmm1 %xmm2 -64+16*3(%rsp)
 	movaps	%xmm2, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm3, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm0, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm3, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm0, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm3, %xmm5
+	shufps	$0x4e, %xmm0, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm1, %xmm3	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm3	# ^
@@ -718,8 +758,13 @@
 # PREP %xmm0 %xmm1 %xmm2 %xmm3 -64+16*0(%rsp)
 	movaps	%xmm3, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm0, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm1, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm0, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm1, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm0, %xmm5
+	shufps	$0x4e, %xmm1, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm2, %xmm0	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm0	# ^
@@ -795,8 +840,13 @@
 # PREP %xmm1 %xmm2 %xmm3 %xmm0 -64+16*1(%rsp)
 	movaps	%xmm0, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm1, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm2, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm1, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm2, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm1, %xmm5
+	shufps	$0x4e, %xmm2, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm3, %xmm1	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm1	# ^
@@ -872,8 +922,13 @@
 # PREP %xmm2 %xmm3 %xmm0 %xmm1 -64+16*2(%rsp)
 	movaps	%xmm1, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm2, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm3, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm2, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm3, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm2, %xmm5
+	shufps	$0x4e, %xmm3, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm0, %xmm2	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm2	# ^
@@ -950,8 +1005,13 @@
 # PREP %xmm3 %xmm0 %xmm1 %xmm2 -64+16*3(%rsp)
 	movaps	%xmm2, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm3, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm0, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm3, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm0, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm3, %xmm5
+	shufps	$0x4e, %xmm0, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm1, %xmm3	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm3	# ^
@@ -1027,8 +1087,13 @@
 # PREP %xmm0 %xmm1 %xmm2 %xmm3 -64+16*0(%rsp)
 	movaps	%xmm3, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm0, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm1, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm0, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm1, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm0, %xmm5
+	shufps	$0x4e, %xmm1, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm2, %xmm0	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm0	# ^
@@ -1104,8 +1169,13 @@
 # PREP %xmm1 %xmm2 %xmm3 %xmm0 -64+16*1(%rsp)
 	movaps	%xmm0, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm1, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm2, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm1, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm2, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm1, %xmm5
+	shufps	$0x4e, %xmm2, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm3, %xmm1	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm1	# ^
@@ -1169,8 +1239,13 @@
 # PREP %xmm2 %xmm3 %xmm0 %xmm1 -64+16*2(%rsp)
 	movaps	%xmm1, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm2, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm3, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm2, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm3, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm2, %xmm5
+	shufps	$0x4e, %xmm3, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm0, %xmm2	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm2	# ^
@@ -1234,8 +1309,13 @@
 # PREP %xmm3 %xmm0 %xmm1 %xmm2 -64+16*3(%rsp)
 	movaps	%xmm2, %xmm4
 	psrldq	$4, %xmm4	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
-	pshufd	$0x4e, %xmm3, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq %xmm0, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	$0x4e, %xmm3, %xmm5	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq %xmm0, %xmm5	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	%xmm3, %xmm5
+	shufps	$0x4e, %xmm0, %xmm5	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 	xorps	%xmm1, %xmm3	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	%xmm4, %xmm5	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
 	xorps	%xmm5, %xmm3	# ^
diff --git a/libbb/hash_md5_sha_x86-64.S.sh b/libbb/hash_md5_sha_x86-64.S.sh
index 656fb54..fb1e4b5 100755
--- a/libbb/hash_md5_sha_x86-64.S.sh
+++ b/libbb/hash_md5_sha_x86-64.S.sh
@@ -203,8 +203,13 @@
 	movaps	$xmmW12, $xmmT1
 	psrldq	\$4, $xmmT1	# rshift by 4 bytes: T1 = ([13],[14],[15],0)
 
-	pshufd	\$0x4e, $xmmW0, $xmmT2	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
-	punpcklqdq $xmmW4, $xmmT2	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+#	pshufd	\$0x4e, $xmmW0, $xmmT2	# 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
+#	punpcklqdq $xmmW4, $xmmT2	# T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
+# same result as above, but shorter and faster:
+# pshufd/shufps are subtly different: pshufd takes all dwords from source operand,
+# shufps takes dwords 0,1 from *2nd* operand, and dwords 2,3 from 1st one!
+	movaps	$xmmW0, $xmmT2
+	shufps	\$0x4e, $xmmW4, $xmmT2	# 01001110=(T2.dw[2], T2.dw[3], W4.dw[0], W4.dw[1]) = ([2],[3],[4],[5])
 
 	xorps	$xmmW8, $xmmW0	# ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
 	xorps	$xmmT1, $xmmT2	# ([13],[14],[15],0) ^ ([2],[3],[4],[5])
diff --git a/libbb/hash_md5_sha_x86-64_shaNI.S b/libbb/hash_md5_sha_x86-64_shaNI.S
index ba92f09..8ddec87 100644
--- a/libbb/hash_md5_sha_x86-64_shaNI.S
+++ b/libbb/hash_md5_sha_x86-64_shaNI.S
@@ -217,8 +217,8 @@
 	ret
 	.size	sha1_process_block64_shaNI, .-sha1_process_block64_shaNI
 
-.section	.rodata.cst16.PSHUFFLE_BYTE_FLIP_MASK, "aM", @progbits, 16
-.balign 16
+	.section	.rodata.cst16.PSHUFFLE_BYTE_FLIP_MASK, "aM", @progbits, 16
+	.balign 16
 PSHUFFLE_BYTE_FLIP_MASK:
 	.octa 0x000102030405060708090a0b0c0d0e0f