diff --git a/crypto/fipsmodule/bn/asm/x86_64-mont5.pl b/crypto/fipsmodule/bn/asm/x86_64-mont5.pl
index 2d5f653032..7a6954d0ca 100755
--- a/crypto/fipsmodule/bn/asm/x86_64-mont5.pl
+++ b/crypto/fipsmodule/bn/asm/x86_64-mont5.pl
@@ -54,7 +54,7 @@
 $addx = 1;
 for (@ARGV) { $addx = 0 if (/-DMY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX/); }
 
-# int bn_mul_mont_gather5(
+# int bn_mul_mont_gather5_nohw(
 $rp="%rdi";	# BN_ULONG *rp,
 $ap="%rsi";	# const BN_ULONG *ap,
 $bp="%rdx";	# const BN_ULONG *bp,
@@ -76,31 +76,15 @@
 $code=<<___;
 .text
 
-.extern	OPENSSL_ia32cap_P
-
-.globl	bn_mul_mont_gather5
-.type	bn_mul_mont_gather5,\@function,6
+.globl	bn_mul_mont_gather5_nohw
+.type	bn_mul_mont_gather5_nohw,\@function,6
 .align	64
-bn_mul_mont_gather5:
+bn_mul_mont_gather5_nohw:
 .cfi_startproc
 	_CET_ENDBR
 	mov	${num}d,${num}d
 	mov	%rsp,%rax
 .cfi_def_cfa_register	%rax
-	test	\$7,${num}d
-	jnz	.Lmul_enter
-___
-$code.=<<___ if ($addx);
-#ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	leaq	OPENSSL_ia32cap_P(%rip),%r11
-	mov	8(%r11),%r11d
-#endif
-___
-$code.=<<___;
-	jmp	.Lmul4x_enter
-
-.align	16
-.Lmul_enter:
 	movd	`($win64?56:8)`(%rsp),%xmm5	# load 7th argument
 	push	%rbx
 .cfi_push	%rbx
@@ -460,29 +444,21 @@
 .Lmul_epilogue:
 	ret
 .cfi_endproc
-.size	bn_mul_mont_gather5,.-bn_mul_mont_gather5
+.size	bn_mul_mont_gather5_nohw,.-bn_mul_mont_gather5_nohw
 ___
 {{{
 my @A=("%r10","%r11");
 my @N=("%r13","%rdi");
 $code.=<<___;
+.global bn_mul4x_mont_gather5
 .type	bn_mul4x_mont_gather5,\@function,6
 .align	32
 bn_mul4x_mont_gather5:
 .cfi_startproc
+    _CET_ENDBR
 	.byte	0x67
 	mov	%rsp,%rax
 .cfi_def_cfa_register	%rax
-.Lmul4x_enter:
-___
-$code.=<<___ if ($addx);
-#ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	and	\$0x80108,%r11d
-	cmp	\$0x80108,%r11d		# check for AD*X+BMI2+BMI1
-	je	.Lmulx4x_enter
-#endif
-___
-$code.=<<___;
 	push	%rbx
 .cfi_push	%rbx
 	push	%rbp
@@ -1087,7 +1063,7 @@
 }}}
 {{{
 ######################################################################
-# void bn_power5(
+# void bn_power5_nohw(
 my $rptr="%rdi";	# BN_ULONG *rptr,
 my $aptr="%rsi";	# const BN_ULONG *aptr,
 my $bptr="%rdx";	# const BN_ULONG *table,
@@ -1102,25 +1078,14 @@
 my ($a0,$a1,$ai)=("%r14","%r15","%rbx");
 
 $code.=<<___;
-.globl	bn_power5
-.type	bn_power5,\@function,6
+.globl	bn_power5_nohw
+.type	bn_power5_nohw,\@function,6
 .align	32
-bn_power5:
+bn_power5_nohw:
 .cfi_startproc
 	_CET_ENDBR
 	mov	%rsp,%rax
 .cfi_def_cfa_register	%rax
-___
-$code.=<<___ if ($addx);
-#ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	leaq	OPENSSL_ia32cap_P(%rip),%r11
-	mov	8(%r11),%r11d
-	and	\$0x80108,%r11d
-	cmp	\$0x80108,%r11d		# check for AD*X+BMI2+BMI1
-	je	.Lpowerx5_enter
-#endif
-___
-$code.=<<___;
 	push	%rbx
 .cfi_push	%rbx
 	push	%rbp
@@ -1243,7 +1208,7 @@
 .Lpower5_epilogue:
 	ret
 .cfi_endproc
-.size	bn_power5,.-bn_power5
+.size	bn_power5_nohw,.-bn_power5_nohw
 
 .globl	bn_sqr8x_internal
 .hidden	bn_sqr8x_internal
@@ -2119,13 +2084,14 @@
 
 $code.=<<___;
 #ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
+.global bn_mulx4x_mont_gather5
 .type	bn_mulx4x_mont_gather5,\@function,6
 .align	32
 bn_mulx4x_mont_gather5:
 .cfi_startproc
+    _CET_ENDBR
 	mov	%rsp,%rax
 .cfi_def_cfa_register	%rax
-.Lmulx4x_enter:
 	push	%rbx
 .cfi_push	%rbx
 	push	%rbp
@@ -2594,7 +2560,7 @@
 ___
 }{
 ######################################################################
-# void bn_power5(
+# void bn_powerx5(
 my $rptr="%rdi";	# BN_ULONG *rptr,
 my $aptr="%rsi";	# const BN_ULONG *aptr,
 my $bptr="%rdx";	# const BN_ULONG *table,
@@ -2609,13 +2575,14 @@
 my ($a0,$a1,$ai)=("%r14","%r15","%rbx");
 
 $code.=<<___;
+.global bn_powerx5
 .type	bn_powerx5,\@function,6
 .align	32
 bn_powerx5:
 .cfi_startproc
+    _CET_ENDBR
 	mov	%rsp,%rax
 .cfi_def_cfa_register	%rax
-.Lpowerx5_enter:
 	push	%rbx
 .cfi_push	%rbx
 	push	%rbp
@@ -3717,17 +3684,17 @@
 
 .section	.pdata
 .align	4
-	.rva	.LSEH_begin_bn_mul_mont_gather5
-	.rva	.LSEH_end_bn_mul_mont_gather5
-	.rva	.LSEH_info_bn_mul_mont_gather5
+	.rva	.LSEH_begin_bn_mul_mont_gather5_nohw
+	.rva	.LSEH_end_bn_mul_mont_gather5_nohw
+	.rva	.LSEH_info_bn_mul_mont_gather5_nohw
 
 	.rva	.LSEH_begin_bn_mul4x_mont_gather5
 	.rva	.LSEH_end_bn_mul4x_mont_gather5
 	.rva	.LSEH_info_bn_mul4x_mont_gather5
 
-	.rva	.LSEH_begin_bn_power5
-	.rva	.LSEH_end_bn_power5
-	.rva	.LSEH_info_bn_power5
+	.rva	.LSEH_begin_bn_power5_nohw
+	.rva	.LSEH_end_bn_power5_nohw
+	.rva	.LSEH_info_bn_power5_nohw
 ___
 $code.=<<___ if ($addx);
 #ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
@@ -3747,7 +3714,7 @@
 
 .section	.xdata
 .align	4
-.LSEH_info_bn_mul_mont_gather5:
+.LSEH_info_bn_mul_mont_gather5_nohw:
 	.byte	9,0,0,0
 	.rva	mul_handler
 	.rva	.Lmul_body,.Lmul_body,.Lmul_epilogue		# HandlerData[]
@@ -3757,7 +3724,7 @@
 	.rva	mul_handler
 	.rva	.Lmul4x_prologue,.Lmul4x_body,.Lmul4x_epilogue		# HandlerData[]
 .align	4
-.LSEH_info_bn_power5:
+.LSEH_info_bn_power5_nohw:
 	.byte	9,0,0,0
 	.rva	mul_handler
 	.rva	.Lpower5_prologue,.Lpower5_body,.Lpower5_epilogue	# HandlerData[]
diff --git a/crypto/fipsmodule/bn/bn_test.cc b/crypto/fipsmodule/bn/bn_test.cc
index ea6ba285f8..5f8f400f0a 100644
--- a/crypto/fipsmodule/bn/bn_test.cc
+++ b/crypto/fipsmodule/bn/bn_test.cc
@@ -2959,14 +2959,12 @@ TEST_F(BNTest, BNMulMontABI) {
     b[0] = 42;
 
 #if defined(OPENSSL_X86_64)
-#if !defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
     if (bn_mulx4x_mont_capable(words)) {
       CHECK_ABI(bn_mulx4x_mont, r.data(), a.data(), b.data(), mont->N.d,
                 mont->n0, words);
       CHECK_ABI(bn_mulx4x_mont, r.data(), a.data(), a.data(), mont->N.d,
                 mont->n0, words);
     }
-#endif // !defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
     if (bn_mul4x_mont_capable(words)) {
       CHECK_ABI(bn_mul4x_mont, r.data(), a.data(), b.data(), mont->N.d,
                 mont->n0, words);
@@ -2977,12 +2975,10 @@ TEST_F(BNTest, BNMulMontABI) {
               mont->n0, words);
     CHECK_ABI(bn_mul_mont_nohw, r.data(), a.data(), a.data(), mont->N.d,
               mont->n0, words);
-#if !defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
     if (bn_sqr8x_mont_capable(words)) {
       CHECK_ABI(bn_sqr8x_mont, r.data(), a.data(), bn_mulx_adx_capable(),
                 mont->N.d, mont->n0, words);
     }
-#endif // !defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
 #elif defined(OPENSSL_ARM)
     if (bn_mul8x_mont_neon_capable(words)) {
       CHECK_ABI(bn_mul8x_mont_neon, r.data(), a.data(), b.data(), mont->N.d,
@@ -3028,18 +3024,35 @@ TEST_F(BNTest, BNMulMont5ABI) {
       bn_scatter5(r.data(), words, table.data(), i);
     }
     CHECK_ABI(bn_gather5, r.data(), words, table.data(), 13);
-
-    CHECK_ABI(bn_mul_mont_gather5, r.data(), r.data(), table.data(), m->d,
+    if (bn_mulx4x_mont_gather5_capable(words)) {
+      CHECK_ABI(bn_mulx4x_mont_gather5, r.data(), r.data(), table.data(), m->d,
+                mont->n0, words, 13);
+      CHECK_ABI(bn_mulx4x_mont_gather5, r.data(), a.data(), table.data(), m->d,
+                mont->n0, words, 13);
+    }
+    if (bn_mul4x_mont_gather5_capable(words)) {
+      CHECK_ABI(bn_mul4x_mont_gather5, r.data(), r.data(), table.data(), m->d,
+                mont->n0, words, 13);
+      CHECK_ABI(bn_mul4x_mont_gather5, r.data(), a.data(), table.data(), m->d,
+                mont->n0, words, 13);
+    }
+    CHECK_ABI(bn_mul_mont_gather5_nohw, r.data(), r.data(), table.data(), m->d,
               mont->n0, words, 13);
-    CHECK_ABI(bn_mul_mont_gather5, r.data(), a.data(), table.data(), m->d,
+    CHECK_ABI(bn_mul_mont_gather5_nohw, r.data(), a.data(), table.data(), m->d,
               mont->n0, words, 13);
 
-    if (words % 8 == 0) {
-      CHECK_ABI(bn_power5, r.data(), r.data(), table.data(), m->d, mont->n0,
+    if (bn_powerx5_capable(words)) {
+      CHECK_ABI(bn_powerx5, r.data(), r.data(), table.data(), m->d, mont->n0,
                 words, 13);
-      CHECK_ABI(bn_power5, r.data(), a.data(), table.data(), m->d, mont->n0,
+      CHECK_ABI(bn_powerx5, r.data(), a.data(), table.data(), m->d, mont->n0,
                 words, 13);
     }
+    if (bn_power5_capable(words)) {
+      CHECK_ABI(bn_power5_nohw, r.data(), r.data(), table.data(), m->d,
+                mont->n0, words, 13);
+      CHECK_ABI(bn_power5_nohw, r.data(), a.data(), table.data(), m->d,
+                mont->n0, words, 13);
+    }
   }
 }
 #endif  // OPENSSL_BN_ASM_MONT5 && SUPPORTS_ABI_TEST
diff --git a/crypto/fipsmodule/bn/exponentiation.c b/crypto/fipsmodule/bn/exponentiation.c
index 8713715b03..4f08008481 100644
--- a/crypto/fipsmodule/bn/exponentiation.c
+++ b/crypto/fipsmodule/bn/exponentiation.c
@@ -107,6 +107,7 @@
  * Hudson (tjh@cryptsoft.com). */
 
 #include <openssl/bn.h>
+#include <openssl/cpu.h>
 
 #include <assert.h>
 #include <limits.h>
@@ -163,6 +164,56 @@ static void exponentiation_s2n_bignum_copy_from_prebuf(BN_ULONG *dest, int width
 #endif
 }
 
+#if defined(OPENSSL_BN_ASM_MONT5)
+
+// bn_mul_mont_gather5 multiples loads index |power| of |table|, multiplies it
+// by |ap| modulo |np|, and stores the result in |rp|. The values are |num|
+// words long and represented in Montgomery form. |n0| is a pointer to the
+// corresponding field in |BN_MONT_CTX|. |table| must be aligned to at least
+// 16 bytes. |power| must be less than 32 and is treated as secret.
+//
+// WARNING: This function implements Almost Montgomery Multiplication from
+// https://eprint.iacr.org/2011/239. The inputs do not need to be fully reduced.
+// However, even if they are fully reduced, the output may not be.
+static void bn_mul_mont_gather5(
+    BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *table, const BN_ULONG *np,
+    const BN_ULONG *n0, int num, int power) {
+  if (bn_mulx4x_mont_gather5_capable(num)) {
+    log_dispatch(15);
+    bn_mulx4x_mont_gather5(rp, ap, table, np, n0, num, power);
+  } else if (bn_mul4x_mont_gather5_capable(num)) {
+    log_dispatch(16);
+    bn_mul4x_mont_gather5(rp, ap, table, np, n0, num, power);
+  } else {
+    log_dispatch(17);
+    bn_mul_mont_gather5_nohw(rp, ap, table, np, n0, num, power);
+  }
+}
+
+// bn_power5 squares |ap| five times and multiplies it by the value stored at
+// index |power| of |table|, modulo |np|. It stores the result in |rp|. The
+// values are |num| words long and represented in Montgomery form. |n0| is a
+// pointer to the corresponding field in |BN_MONT_CTX|. |num| must be divisible
+// by 8. |power| must be less than 32 and is treated as secret.
+//
+// WARNING: This function implements Almost Montgomery Multiplication from
+// https://eprint.iacr.org/2011/239. The inputs do not need to be fully reduced.
+// However, even if they are fully reduced, the output may not be.
+static void bn_power5(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *table,
+                      const BN_ULONG *np, const BN_ULONG *n0, int num,
+                      int power)
+{
+  assert(bn_power5_capable(num));
+  if (bn_powerx5_capable(num)) {
+    log_dispatch(18);
+    bn_powerx5(rp, ap, table, np, n0, num, power);
+  } else {
+    log_dispatch(19);
+    bn_power5_nohw(rp, ap, table, np, n0, num, power);
+  }
+}
+
+#endif // defined(OPENSSL_BN_ASM_MONT5)
 
 int BN_exp(BIGNUM *r, const BIGNUM *a, const BIGNUM *p, BN_CTX *ctx) {
   int i, bits, ret = 0;
@@ -1122,7 +1173,7 @@ int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p,
 
     // Scan the exponent one window at a time starting from the most
     // significant bits.
-    if (top & 7) {
+    if (!bn_power5_capable(top)) {
       while (bits >= 0) {
         for (wvalue = 0, i = 0; i < 5; i++, bits--) {
           wvalue = (wvalue << 1) + BN_is_bit_set(p, bits);
diff --git a/crypto/fipsmodule/bn/generic.c b/crypto/fipsmodule/bn/generic.c
index 247398fddb..5a5f61d0db 100644
--- a/crypto/fipsmodule/bn/generic.c
+++ b/crypto/fipsmodule/bn/generic.c
@@ -622,3 +622,38 @@ BN_ULONG bn_sub_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b,
 }
 
 #endif  // !BN_ADD_ASM
+
+// Default implementations of hardware accelerated functions that throw errors if reached.
+// Used to reduce the number of guards needed in code.
+#if defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
+
+#if defined(OPENSSL_X86_64) && defined(OPENSSL_BN_ASM_MONT)
+int bn_mulx4x_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
+                   const BN_ULONG *np, const BN_ULONG *n0, size_t num) {
+  perror("bn_mulx4x_mont");
+  abort();
+}
+#endif
+
+#if defined(OPENSSL_BN_ASM_MONT5)
+void bn_mulx4x_mont_gather5(BN_ULONG *rp, const BN_ULONG *ap,
+                             const BN_ULONG *table, const BN_ULONG *np,
+                             const BN_ULONG *n0, int num, int power) {
+  perror("bn_mulx4x_mont_gather5");
+  abort();
+}
+
+void bn_powerx5(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *table,
+                const BN_ULONG *np, const BN_ULONG *n0, int num, int power) {
+  perror("bn_powerx5");
+  abort();
+}
+
+int bn_sqr8x_mont(BN_ULONG *rp, const BN_ULONG *ap, BN_ULONG mulx_adx_capable,
+                  const BN_ULONG *np, const BN_ULONG *n0, size_t num) {
+  perror("bn_sqr8x_mont");
+  abort();
+}
+#endif
+
+#endif // defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
diff --git a/crypto/fipsmodule/bn/internal.h b/crypto/fipsmodule/bn/internal.h
index a52bb5782f..b837aca823 100644
--- a/crypto/fipsmodule/bn/internal.h
+++ b/crypto/fipsmodule/bn/internal.h
@@ -419,18 +419,23 @@ OPENSSL_INLINE int bn_mul4x_mont_capable(size_t num) {
 }
 int bn_mul4x_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
                   const BN_ULONG *np, const BN_ULONG *n0, size_t num);
-#if !defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
+
 OPENSSL_INLINE int bn_mulx4x_mont_capable(size_t num) {
+#if defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
+  return 0;
+#endif // defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
   return bn_mul4x_mont_capable(num) && bn_mulx_adx_capable();
 }
 int bn_mulx4x_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
                    const BN_ULONG *np, const BN_ULONG *n0, size_t num);
 OPENSSL_INLINE int bn_sqr8x_mont_capable(size_t num) {
+#if defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
+  return 0;
+#endif // defined(MY_ASSEMBLER_IS_TOO_OOLD_FOR_512AVX)
   return (num >= 8) && ((num & 7) == 0);
 }
 int bn_sqr8x_mont(BN_ULONG *rp, const BN_ULONG *ap, BN_ULONG mulx_adx_capable,
                   const BN_ULONG *np, const BN_ULONG *n0, size_t num);
-#endif // !defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
 #elif defined(OPENSSL_ARM)
   OPENSSL_INLINE int bn_mul8x_mont_neon_capable(size_t num) {
     return (num & 7) == 0 && CRYPTO_is_NEON_capable();
@@ -446,18 +451,27 @@ int bn_sqr8x_mont(BN_ULONG *rp, const BN_ULONG *ap, BN_ULONG mulx_adx_capable,
 #if !defined(OPENSSL_NO_ASM) && defined(OPENSSL_X86_64)
 #define OPENSSL_BN_ASM_MONT5
 
-// bn_mul_mont_gather5 multiples loads index |power| of |table|, multiplies it
-// by |ap| modulo |np|, and stores the result in |rp|. The values are |num|
-// words long and represented in Montgomery form. |n0| is a pointer to the
-// corresponding field in |BN_MONT_CTX|. |table| must be aligned to at least
-// 16 bytes. |power| must be less than 32 and is treated as secret.
-//
-// WARNING: This function implements Almost Montgomery Multiplication from
-// https://eprint.iacr.org/2011/239. The inputs do not need to be fully reduced.
-// However, even if they are fully reduced, the output may not be.
-void bn_mul_mont_gather5(BN_ULONG *rp, const BN_ULONG *ap,
-                         const BN_ULONG *table, const BN_ULONG *np,
-                         const BN_ULONG *n0, int num, int power);
+  OPENSSL_INLINE int bn_mul4x_mont_gather5_capable(int num) {
+    return (num & 7) == 0;
+  }
+  void bn_mul4x_mont_gather5(BN_ULONG *rp, const BN_ULONG *ap,
+                             const BN_ULONG *table, const BN_ULONG *np,
+                             const BN_ULONG *n0, int num, int power);
+
+  OPENSSL_INLINE int bn_mulx4x_mont_gather5_capable(int num) {
+  #if defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
+    return 0;
+  #endif // defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
+    return bn_mul4x_mont_gather5_capable(num) && CRYPTO_is_ADX_capable() &&
+        CRYPTO_is_BMI1_capable() && CRYPTO_is_BMI2_capable();
+  }
+  void bn_mulx4x_mont_gather5(BN_ULONG *rp, const BN_ULONG *ap,
+                              const BN_ULONG *table, const BN_ULONG *np,
+                              const BN_ULONG *n0, int num, int power);
+
+  void bn_mul_mont_gather5_nohw(BN_ULONG *rp, const BN_ULONG *ap,
+                                const BN_ULONG *table, const BN_ULONG *np,
+                                const BN_ULONG *n0, int num, int power);
 
 // bn_scatter5 stores |inp| to index |power| of |table|. |inp| and each entry of
 // |table| are |num| words long. |power| must be less than 32 and is treated as
@@ -471,17 +485,23 @@ void bn_scatter5(const BN_ULONG *inp, size_t num, BN_ULONG *table,
 // is treated as secret. |table| must be aligned to at least 16 bytes.
 void bn_gather5(BN_ULONG *out, size_t num, const BN_ULONG *table, size_t power);
 
-// bn_power5 squares |ap| five times and multiplies it by the value stored at
-// index |power| of |table|, modulo |np|. It stores the result in |rp|. The
-// values are |num| words long and represented in Montgomery form. |n0| is a
-// pointer to the corresponding field in |BN_MONT_CTX|. |num| must be divisible
-// by 8. |power| must be less than 32 and is treated as secret.
-//
-// WARNING: This function implements Almost Montgomery Multiplication from
-// https://eprint.iacr.org/2011/239. The inputs do not need to be fully reduced.
-// However, even if they are fully reduced, the output may not be.
-void bn_power5(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *table,
-               const BN_ULONG *np, const BN_ULONG *n0, int num, int power);
+  void bn_power5_nohw(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *table,
+                      const BN_ULONG *np, const BN_ULONG *n0, int num, int power);
+
+  OPENSSL_INLINE int bn_power5_capable(int num) {
+    return (num & 7) == 0;
+  }
+
+  OPENSSL_INLINE int bn_powerx5_capable(int num) {
+  #if defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
+    return 0;
+  #endif // defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
+    return bn_power5_capable(num) && CRYPTO_is_ADX_capable() &&
+           CRYPTO_is_BMI1_capable() && CRYPTO_is_BMI2_capable();
+  }
+  void bn_powerx5(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *table,
+                  const BN_ULONG *np, const BN_ULONG *n0, int num, int power);
+
 #endif  // !OPENSSL_NO_ASM && OPENSSL_X86_64
 
 uint64_t bn_mont_n0(const BIGNUM *n);
diff --git a/crypto/fipsmodule/bn/montgomery.c b/crypto/fipsmodule/bn/montgomery.c
index c7ac15c18d..0a4432b7ae 100644
--- a/crypto/fipsmodule/bn/montgomery.c
+++ b/crypto/fipsmodule/bn/montgomery.c
@@ -627,18 +627,15 @@ void bn_mod_mul_montgomery_small(BN_ULONG *r, const BN_ULONG *a,
 int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
                 const BN_ULONG *np, const BN_ULONG *n0, size_t num)
 {
-#if !defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
   if (ap == bp && bn_sqr8x_mont_capable(num)) {
     return bn_sqr8x_mont(rp, ap, bn_mulx_adx_capable(), np, n0, num);
-  }
-  if (bn_mulx4x_mont_capable(num)) {
+  } else if (bn_mulx4x_mont_capable(num)) {
     return bn_mulx4x_mont(rp, ap, bp, np, n0, num);
-  }
-#endif // !defined(MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX)
-  if (bn_mul4x_mont_capable(num)) {
+  } else if (bn_mul4x_mont_capable(num)) {
     return bn_mul4x_mont(rp, ap, bp, np, n0, num);
+  } else {
+    return bn_mul_mont_nohw(rp, ap, bp, np, n0, num);
   }
-  return bn_mul_mont_nohw(rp, ap, bp, np, n0, num);
 }
 #endif
 
diff --git a/crypto/fipsmodule/cpucap/cpucap.c b/crypto/fipsmodule/cpucap/cpucap.c
index 07874c7788..014096a56e 100644
--- a/crypto/fipsmodule/cpucap/cpucap.c
+++ b/crypto/fipsmodule/cpucap/cpucap.c
@@ -94,6 +94,12 @@ HIDDEN uint32_t OPENSSL_armcap_P = 0;
 HIDDEN uint8_t BORINGSSL_function_hit[15] = {0};
 #endif // BORINGSSL_DISPATCH_TEST
 
+void log_dispatch(size_t id) {
+#if BORINGSSL_DISPATCH_TEST
+  BORINGSSL_function_hit[id] = 1;
+#endif
+}
+
 // This variable is used only for testing purposes to ensure that the library
 // constructor is executed and the capability variable is initialized.
 HIDDEN uint8_t OPENSSL_cpucap_initialized = 0;
diff --git a/crypto/fipsmodule/sha/keccak1600.c b/crypto/fipsmodule/sha/keccak1600.c
index f97d03a6a9..bbe8682e76 100644
--- a/crypto/fipsmodule/sha/keccak1600.c
+++ b/crypto/fipsmodule/sha/keccak1600.c
@@ -8,6 +8,7 @@
  */
 
 #include <assert.h>
+#include <openssl/cpu.h>
 #include "internal.h"
 #include "../../internal.h"
 #include "../cpucap/internal.h"
@@ -321,14 +322,6 @@ void Keccak1600_Squeeze(uint64_t A[KECCAK1600_ROWS][KECCAK1600_ROWS], uint8_t *o
 // Scalar implementation from OpenSSL provided by keccak1600-armv8.pl
 extern void KeccakF1600_hw(uint64_t state[25]);
 
-#if defined(OPENSSL_AARCH64)
-static void keccak_log_dispatch(size_t id) {
-#if BORINGSSL_DISPATCH_TEST
-    BORINGSSL_function_hit[id] = 1;
-#endif
-}
-#endif
-
 void KeccakF1600(uint64_t A[KECCAK1600_ROWS][KECCAK1600_ROWS]) {
     // Dispatch logic for Keccak-x1 on AArch64:
     //
@@ -354,21 +347,21 @@ void KeccakF1600(uint64_t A[KECCAK1600_ROWS][KECCAK1600_ROWS]) {
 #if defined(OPENSSL_AARCH64)
 #if defined(KECCAK1600_S2N_BIGNUM_ASM)
     if (CRYPTO_is_Neoverse_N1() || CRYPTO_is_Neoverse_V1() || CRYPTO_is_Neoverse_V2()) {
-        keccak_log_dispatch(10); // kFlag_sha3_keccak_f1600
+        log_dispatch(10); // kFlag_sha3_keccak_f1600
         sha3_keccak_f1600((uint64_t *)A, iotas);
         return;
     }
 
 #if defined(MY_ASSEMBLER_SUPPORTS_NEON_SHA3_EXTENSION)
     if (CRYPTO_is_ARMv8_SHA3_capable()) {
-        keccak_log_dispatch(11); // kFlag_sha3_keccak_f1600_alt
+        log_dispatch(11); // kFlag_sha3_keccak_f1600_alt
         sha3_keccak_f1600_alt((uint64_t *)A, iotas);
         return;
     }
 #endif
 #endif
 
-    keccak_log_dispatch(9); // kFlag_KeccakF1600_hw
+    log_dispatch(9); // kFlag_KeccakF1600_hw
     KeccakF1600_hw((uint64_t *) A);
 
 #elif defined(OPENSSL_X86_64)
@@ -427,20 +420,20 @@ static void Keccak1600_x4(uint64_t A[4][KECCAK1600_ROWS][KECCAK1600_ROWS]) {
     //   (which has its own dispatch logic).
 #if defined(KECCAK1600_S2N_BIGNUM_ASM) && defined(OPENSSL_AARCH64)
     if (CRYPTO_is_Neoverse_N1()) {
-        keccak_log_dispatch(13); // kFlag_sha3_keccak4_f1600_alt
+        log_dispatch(13); // kFlag_sha3_keccak4_f1600_alt
         sha3_keccak4_f1600_alt((uint64_t *)A, iotas);
         return;
     }
 
 #if defined(MY_ASSEMBLER_SUPPORTS_NEON_SHA3_EXTENSION)
     if (CRYPTO_is_Neoverse_V1() || CRYPTO_is_Neoverse_V2()) {
-        keccak_log_dispatch(14); // kFlag_sha3_keccak4_f1600_alt2
+        log_dispatch(14); // kFlag_sha3_keccak4_f1600_alt2
         sha3_keccak4_f1600_alt2((uint64_t *)A, iotas);
         return;
     }
 
     if (CRYPTO_is_ARMv8_SHA3_capable()) {
-        keccak_log_dispatch(12); // kFlag_sha3_keccak2_f1600
+        log_dispatch(12); // kFlag_sha3_keccak2_f1600
         // Use 2-fold function twice: A[0:1] and A[2:3]
         sha3_keccak2_f1600((uint64_t *)&A[0], iotas);
         sha3_keccak2_f1600((uint64_t *)&A[2], iotas);
diff --git a/crypto/impl_dispatch_test.cc b/crypto/impl_dispatch_test.cc
index c971419b83..87c3a193f8 100644
--- a/crypto/impl_dispatch_test.cc
+++ b/crypto/impl_dispatch_test.cc
@@ -31,8 +31,8 @@
 #include "fipsmodule/cpucap/internal.h"
 #include "fipsmodule/modes/internal.h"
 #include "fipsmodule/bn/rsaz_exp.h"
+#include "fipsmodule/bn/internal.h"
 #include "fipsmodule/sha/internal.h"
-
 #include "test/file_test.h"
 
 class ImplDispatchTest : public ::testing::Test {
@@ -140,6 +140,11 @@ class ImplDispatchTest : public ::testing::Test {
   bool is_assembler_too_old = false;
   bool is_assembler_too_old_avx512 = false;
   bool ifma_avx512 = false;
+  bool is_bn_mulx4x_mont_gather5 = false;
+  bool is_bn_mul4x_mont_gather5 = false;
+  bool is_bn_mul_mont_gather5_nohw = false;
+  bool is_power5 = false;
+  bool is_powerx5 = false;
 #else // AARCH64
   bool aes_gcm_pmull_ = false;
   bool aes_gcm_8x_ = false;
@@ -167,6 +172,16 @@ constexpr size_t kFlag_sha256_hw = 6;
 constexpr size_t kFlag_aesni_gcm_encrypt = 2;
 constexpr size_t kFlag_aes_gcm_encrypt_avx512 = 7;
 constexpr size_t kFlag_RSAZ_mod_exp_avx512_x2 = 8;
+__attribute__((unused))
+constexpr size_t kFlag_bn_mulx4x_mont_gather5 = 15;
+__attribute__((unused))
+constexpr size_t kFlag_bn_mul4x_mont_gather5 = 16;
+__attribute__((unused))
+constexpr size_t kFlag_bn_mul_mont_gather5_nohw = 17;
+__attribute__((unused))
+constexpr size_t kFlag_bn_powerx5 = 18;
+__attribute__((unused))
+constexpr size_t kFlag_bn_power5_nohw = 19;
 #else // AARCH64
 constexpr size_t kFlag_aes_gcm_enc_kernel = 2;
 constexpr size_t kFlag_aesv8_gcm_8x_enc_128 = 7;
@@ -179,6 +194,91 @@ constexpr size_t kFlag_sha3_keccak4_f1600_alt = 13;
 constexpr size_t kFlag_sha3_keccak4_f1600_alt2 = 14;
 #endif
 
+#if defined(OPENSSL_BN_ASM_MONT5)
+/**
+TEST_F(ImplDispatchTest, BN_mul_mont_gather5) {
+  for (size_t words : {4, 5, 6, 7, 8, 16, 32}) {
+    SCOPED_TRACE(words);
+
+    bssl::UniquePtr<BIGNUM> m(BN_new());
+    ASSERT_TRUE(m);
+    bssl::UniquePtr<BN_CTX> ctx(BN_CTX_new());
+    BN_CTX_start(ctx.get());
+    ASSERT_TRUE(BN_set_bit(m.get(), 0));
+    ASSERT_TRUE(BN_set_bit(m.get(), words * BN_BITS2 - 1));
+    bssl::UniquePtr<BN_MONT_CTX> mont(
+        BN_MONT_CTX_new_for_modulus(m.get(), ctx.get()));
+    ASSERT_TRUE(mont);
+
+    std::vector<BN_ULONG> r(words), a(words), b(words), table(words * 32);
+    a[0] = 1;
+    b[0] = 42;
+
+    bn_mul_mont(r.data(), a.data(), b.data(), mont->N.d, mont->n0, words);
+
+    is_bn_mulx4x_mont_gather5 = bn_mulx4x_mont_gather5_capable(words);
+    if (is_bn_mulx4x_mont_gather5) {
+      is_bn_mul4x_mont_gather5 = false;
+      is_bn_mul_mont_gather5_nohw = false;
+    } else {
+      is_bn_mul4x_mont_gather5 = bn_mul4x_mont_gather5_capable(words);
+      if (is_bn_mul4x_mont_gather5) {
+        is_bn_mul_mont_gather5_nohw = false;
+      } else {
+        is_bn_mul_mont_gather5_nohw = true;
+      }
+    }
+
+    AssertFunctionsHit(
+      {
+          {kFlag_bn_mulx4x_mont_gather5, is_bn_mulx4x_mont_gather5},
+          {kFlag_bn_mul4x_mont_gather5, is_bn_mul4x_mont_gather5},
+          {kFlag_bn_mul_mont_gather5_nohw, is_bn_mul_mont_gather5_nohw},
+      },
+      [r, table, m, mont, words] {
+        bn_mul_mont_gather5(r.data(), r.data(), table.data(), m->d, mont->n0, words, 13);
+        bn_mul_mont_gather5(r.data(), a.data(), table.data(), m->d, mont->n0, words, 13);
+      });
+  }
+}
+*/
+
+/**
+TEST_F(ImplDispatchTest, BN_power5) {
+  for (size_t words : {4, 5, 6, 7, 8, 16, 32}) {
+    SCOPED_TRACE(words);
+
+    bssl::UniquePtr<BIGNUM> m(BN_new());
+    ASSERT_TRUE(m);
+    ASSERT_TRUE(BN_set_bit(m.get(), 0));
+    ASSERT_TRUE(BN_set_bit(m.get(), words * BN_BITS2 - 1));
+    bssl::UniquePtr<BN_MONT_CTX> mont(
+        BN_MONT_CTX_new_for_modulus(m.get(), ctx()));
+    ASSERT_TRUE(mont);
+
+    std::vector<BN_ULONG> r(words), a(words), b(words), table(words * 32);
+    a[0] = 1;
+    b[0] = 42;
+
+    bn_mul_mont(r.data(), a.data(), b.data(), mont->N.d, mont->n0, words);
+
+    is_power5 = bn_power5_capable(words);
+    is_powerx5 = bn_powerx5_capable(words);
+
+    AssertFunctionsHit(
+      {
+          {kFlag_bn_powerx5, is_powerx5},
+          {kFlag_bn_power5_nohw, !is_powerx5 && is_power5},
+      },
+      [] {
+        bn_power5(r.data(), r.data(), table.data(), m->d, mont->n0, words, 13);
+        bn_power5(r.data(), a.data(), table.data(), m->d, mont->n0, words, 13);
+      });
+  }
+}
+*/
+#endif // defined(OPENSSL_BN_ASM_MONT5)
+
 TEST_F(ImplDispatchTest, AEAD_AES_GCM) {
   AssertFunctionsHit(
       {
diff --git a/generated-src/linux-x86_64/crypto/fipsmodule/x86_64-mont5.S b/generated-src/linux-x86_64/crypto/fipsmodule/x86_64-mont5.S
index 4450c3ef72..e42d7cffbb 100644
--- a/generated-src/linux-x86_64/crypto/fipsmodule/x86_64-mont5.S
+++ b/generated-src/linux-x86_64/crypto/fipsmodule/x86_64-mont5.S
@@ -6,29 +6,16 @@
 #if !defined(OPENSSL_NO_ASM) && defined(OPENSSL_X86_64) && defined(__ELF__)
 .text	
 
-.extern	OPENSSL_ia32cap_P
-.hidden OPENSSL_ia32cap_P
-
-.globl	bn_mul_mont_gather5
-.hidden bn_mul_mont_gather5
-.type	bn_mul_mont_gather5,@function
+.globl	bn_mul_mont_gather5_nohw
+.hidden bn_mul_mont_gather5_nohw
+.type	bn_mul_mont_gather5_nohw,@function
 .align	64
-bn_mul_mont_gather5:
+bn_mul_mont_gather5_nohw:
 .cfi_startproc	
 _CET_ENDBR
 	movl	%r9d,%r9d
 	movq	%rsp,%rax
 .cfi_def_cfa_register	%rax
-	testl	$7,%r9d
-	jnz	.Lmul_enter
-#ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	leaq	OPENSSL_ia32cap_P(%rip),%r11
-	movl	8(%r11),%r11d
-#endif
-	jmp	.Lmul4x_enter
-
-.align	16
-.Lmul_enter:
 	movd	8(%rsp),%xmm5
 	pushq	%rbx
 .cfi_offset	%rbx,-16
@@ -454,20 +441,17 @@ _CET_ENDBR
 .Lmul_epilogue:
 	.byte	0xf3,0xc3
 .cfi_endproc	
-.size	bn_mul_mont_gather5,.-bn_mul_mont_gather5
+.size	bn_mul_mont_gather5_nohw,.-bn_mul_mont_gather5_nohw
+.global	bn_mul4x_mont_gather5
+.hidden bn_mul4x_mont_gather5
 .type	bn_mul4x_mont_gather5,@function
 .align	32
 bn_mul4x_mont_gather5:
 .cfi_startproc	
+_CET_ENDBR
 .byte	0x67
 	movq	%rsp,%rax
 .cfi_def_cfa_register	%rax
-.Lmul4x_enter:
-#ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	andl	$0x80108,%r11d
-	cmpl	$0x80108,%r11d
-	je	.Lmulx4x_enter
-#endif
 	pushq	%rbx
 .cfi_offset	%rbx,-16
 	pushq	%rbp
@@ -1092,22 +1076,15 @@ mul4x_internal:
 	jmp	.Lsqr4x_sub_entry
 .cfi_endproc	
 .size	mul4x_internal,.-mul4x_internal
-.globl	bn_power5
-.hidden bn_power5
-.type	bn_power5,@function
+.globl	bn_power5_nohw
+.hidden bn_power5_nohw
+.type	bn_power5_nohw,@function
 .align	32
-bn_power5:
+bn_power5_nohw:
 .cfi_startproc	
 _CET_ENDBR
 	movq	%rsp,%rax
 .cfi_def_cfa_register	%rax
-#ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	leaq	OPENSSL_ia32cap_P(%rip),%r11
-	movl	8(%r11),%r11d
-	andl	$0x80108,%r11d
-	cmpl	$0x80108,%r11d
-	je	.Lpowerx5_enter
-#endif
 	pushq	%rbx
 .cfi_offset	%rbx,-16
 	pushq	%rbp
@@ -1230,7 +1207,7 @@ _CET_ENDBR
 .Lpower5_epilogue:
 	.byte	0xf3,0xc3
 .cfi_endproc	
-.size	bn_power5,.-bn_power5
+.size	bn_power5_nohw,.-bn_power5_nohw
 
 .globl	bn_sqr8x_internal
 .hidden bn_sqr8x_internal
@@ -2074,13 +2051,15 @@ __bn_post4x_internal:
 .cfi_endproc	
 .size	__bn_post4x_internal,.-__bn_post4x_internal
 #ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
+.global	bn_mulx4x_mont_gather5
+.hidden bn_mulx4x_mont_gather5
 .type	bn_mulx4x_mont_gather5,@function
 .align	32
 bn_mulx4x_mont_gather5:
 .cfi_startproc	
+_CET_ENDBR
 	movq	%rsp,%rax
 .cfi_def_cfa_register	%rax
-.Lmulx4x_enter:
 	pushq	%rbx
 .cfi_offset	%rbx,-16
 	pushq	%rbp
@@ -2611,13 +2590,15 @@ mulx4x_internal:
 	jmp	.Lsqrx4x_sub_entry
 .cfi_endproc	
 .size	mulx4x_internal,.-mulx4x_internal
+.global	bn_powerx5
+.hidden bn_powerx5
 .type	bn_powerx5,@function
 .align	32
 bn_powerx5:
 .cfi_startproc	
+_CET_ENDBR
 	movq	%rsp,%rax
 .cfi_def_cfa_register	%rax
-.Lpowerx5_enter:
 	pushq	%rbx
 .cfi_offset	%rbx,-16
 	pushq	%rbp
diff --git a/generated-src/mac-x86_64/crypto/fipsmodule/x86_64-mont5.S b/generated-src/mac-x86_64/crypto/fipsmodule/x86_64-mont5.S
index cd7c7dadf1..31e468a4f4 100644
--- a/generated-src/mac-x86_64/crypto/fipsmodule/x86_64-mont5.S
+++ b/generated-src/mac-x86_64/crypto/fipsmodule/x86_64-mont5.S
@@ -6,28 +6,16 @@
 #if !defined(OPENSSL_NO_ASM) && defined(OPENSSL_X86_64) && defined(__APPLE__)
 .text	
 
-
-
-.globl	_bn_mul_mont_gather5
-.private_extern _bn_mul_mont_gather5
+.globl	_bn_mul_mont_gather5_nohw
+.private_extern _bn_mul_mont_gather5_nohw
 
 .p2align	6
-_bn_mul_mont_gather5:
+_bn_mul_mont_gather5_nohw:
 
 _CET_ENDBR
 	movl	%r9d,%r9d
 	movq	%rsp,%rax
 
-	testl	$7,%r9d
-	jnz	L$mul_enter
-#ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	leaq	_OPENSSL_ia32cap_P(%rip),%r11
-	movl	8(%r11),%r11d
-#endif
-	jmp	L$mul4x_enter
-
-.p2align	4
-L$mul_enter:
 	movd	8(%rsp),%xmm5
 	pushq	%rbx
 
@@ -454,19 +442,16 @@ L$mul_epilogue:
 	.byte	0xf3,0xc3
 
 
+.global	_bn_mul4x_mont_gather5
+.private_extern _bn_mul4x_mont_gather5
 
 .p2align	5
-bn_mul4x_mont_gather5:
+_bn_mul4x_mont_gather5:
 
+_CET_ENDBR
 .byte	0x67
 	movq	%rsp,%rax
 
-L$mul4x_enter:
-#ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	andl	$0x80108,%r11d
-	cmpl	$0x80108,%r11d
-	je	L$mulx4x_enter
-#endif
 	pushq	%rbx
 
 	pushq	%rbp
@@ -1091,22 +1076,15 @@ L$inner4x:
 	jmp	L$sqr4x_sub_entry
 
 
-.globl	_bn_power5
-.private_extern _bn_power5
+.globl	_bn_power5_nohw
+.private_extern _bn_power5_nohw
 
 .p2align	5
-_bn_power5:
+_bn_power5_nohw:
 
 _CET_ENDBR
 	movq	%rsp,%rax
 
-#ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	leaq	_OPENSSL_ia32cap_P(%rip),%r11
-	movl	8(%r11),%r11d
-	andl	$0x80108,%r11d
-	cmpl	$0x80108,%r11d
-	je	L$powerx5_enter
-#endif
 	pushq	%rbx
 
 	pushq	%rbp
@@ -2073,13 +2051,15 @@ L$sqr4x_sub_entry:
 
 
 #ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
+.global	_bn_mulx4x_mont_gather5
+.private_extern _bn_mulx4x_mont_gather5
 
 .p2align	5
-bn_mulx4x_mont_gather5:
+_bn_mulx4x_mont_gather5:
 
+_CET_ENDBR
 	movq	%rsp,%rax
 
-L$mulx4x_enter:
 	pushq	%rbx
 
 	pushq	%rbp
@@ -2610,13 +2590,15 @@ L$mulx4x_inner:
 	jmp	L$sqrx4x_sub_entry
 
 
+.global	_bn_powerx5
+.private_extern _bn_powerx5
 
 .p2align	5
-bn_powerx5:
+_bn_powerx5:
 
+_CET_ENDBR
 	movq	%rsp,%rax
 
-L$powerx5_enter:
 	pushq	%rbx
 
 	pushq	%rbp
diff --git a/generated-src/win-x86_64/crypto/fipsmodule/x86_64-mont5.asm b/generated-src/win-x86_64/crypto/fipsmodule/x86_64-mont5.asm
index a3bb9883fb..6b868427db 100644
--- a/generated-src/win-x86_64/crypto/fipsmodule/x86_64-mont5.asm
+++ b/generated-src/win-x86_64/crypto/fipsmodule/x86_64-mont5.asm
@@ -12,16 +12,14 @@ default	rel
 section	.text code align=64
 
 
-EXTERN	OPENSSL_ia32cap_P
-
-global	bn_mul_mont_gather5
+global	bn_mul_mont_gather5_nohw
 
 ALIGN	64
-bn_mul_mont_gather5:
+bn_mul_mont_gather5_nohw:
 	mov	QWORD[8+rsp],rdi	;WIN64 prologue
 	mov	QWORD[16+rsp],rsi
 	mov	rax,rsp
-$L$SEH_begin_bn_mul_mont_gather5:
+$L$SEH_begin_bn_mul_mont_gather5_nohw:
 	mov	rdi,rcx
 	mov	rsi,rdx
 	mov	rdx,r8
@@ -35,16 +33,6 @@ _CET_ENDBR
 	mov	r9d,r9d
 	mov	rax,rsp
 
-	test	r9d,7
-	jnz	NEAR $L$mul_enter
-%ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	lea	r11,[OPENSSL_ia32cap_P]
-	mov	r11d,DWORD[8+r11]
-%endif
-	jmp	NEAR $L$mul4x_enter
-
-ALIGN	16
-$L$mul_enter:
 	movd	xmm5,DWORD[56+rsp]
 	push	rbx
 
@@ -472,7 +460,8 @@ $L$mul_epilogue:
 	mov	rsi,QWORD[16+rsp]
 	DB	0F3h,0C3h		;repret
 
-$L$SEH_end_bn_mul_mont_gather5:
+$L$SEH_end_bn_mul_mont_gather5_nohw:
+global	bn_mul4x_mont_gather5
 
 ALIGN	32
 bn_mul4x_mont_gather5:
@@ -489,15 +478,10 @@ $L$SEH_begin_bn_mul4x_mont_gather5:
 
 
 
+_CET_ENDBR
 	DB	0x67
 	mov	rax,rsp
 
-$L$mul4x_enter:
-%ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	and	r11d,0x80108
-	cmp	r11d,0x80108
-	je	NEAR $L$mulx4x_enter
-%endif
 	push	rbx
 
 	push	rbp
@@ -1124,14 +1108,14 @@ $L$inner4x:
 	jmp	NEAR $L$sqr4x_sub_entry
 
 
-global	bn_power5
+global	bn_power5_nohw
 
 ALIGN	32
-bn_power5:
+bn_power5_nohw:
 	mov	QWORD[8+rsp],rdi	;WIN64 prologue
 	mov	QWORD[16+rsp],rsi
 	mov	rax,rsp
-$L$SEH_begin_bn_power5:
+$L$SEH_begin_bn_power5_nohw:
 	mov	rdi,rcx
 	mov	rsi,rdx
 	mov	rdx,r8
@@ -1144,13 +1128,6 @@ $L$SEH_begin_bn_power5:
 _CET_ENDBR
 	mov	rax,rsp
 
-%ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
-	lea	r11,[OPENSSL_ia32cap_P]
-	mov	r11d,DWORD[8+r11]
-	and	r11d,0x80108
-	cmp	r11d,0x80108
-	je	NEAR $L$powerx5_enter
-%endif
 	push	rbx
 
 	push	rbp
@@ -1275,7 +1252,7 @@ $L$power5_epilogue:
 	mov	rsi,QWORD[16+rsp]
 	DB	0F3h,0C3h		;repret
 
-$L$SEH_end_bn_power5:
+$L$SEH_end_bn_power5_nohw:
 
 global	bn_sqr8x_internal
 
@@ -2118,6 +2095,7 @@ $L$sqr4x_sub_entry:
 
 
 %ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
+global	bn_mulx4x_mont_gather5
 
 ALIGN	32
 bn_mulx4x_mont_gather5:
@@ -2134,9 +2112,9 @@ $L$SEH_begin_bn_mulx4x_mont_gather5:
 
 
 
+_CET_ENDBR
 	mov	rax,rsp
 
-$L$mulx4x_enter:
 	push	rbx
 
 	push	rbp
@@ -2669,6 +2647,7 @@ $L$mulx4x_inner:
 	jmp	NEAR $L$sqrx4x_sub_entry
 
 
+global	bn_powerx5
 
 ALIGN	32
 bn_powerx5:
@@ -2685,9 +2664,9 @@ $L$SEH_begin_bn_powerx5:
 
 
 
+_CET_ENDBR
 	mov	rax,rsp
 
-$L$powerx5_enter:
 	push	rbx
 
 	push	rbp
@@ -3810,17 +3789,17 @@ $L$common_seh_tail:
 
 section	.pdata rdata align=4
 ALIGN	4
-	DD	$L$SEH_begin_bn_mul_mont_gather5 wrt ..imagebase
-	DD	$L$SEH_end_bn_mul_mont_gather5 wrt ..imagebase
-	DD	$L$SEH_info_bn_mul_mont_gather5 wrt ..imagebase
+	DD	$L$SEH_begin_bn_mul_mont_gather5_nohw wrt ..imagebase
+	DD	$L$SEH_end_bn_mul_mont_gather5_nohw wrt ..imagebase
+	DD	$L$SEH_info_bn_mul_mont_gather5_nohw wrt ..imagebase
 
 	DD	$L$SEH_begin_bn_mul4x_mont_gather5 wrt ..imagebase
 	DD	$L$SEH_end_bn_mul4x_mont_gather5 wrt ..imagebase
 	DD	$L$SEH_info_bn_mul4x_mont_gather5 wrt ..imagebase
 
-	DD	$L$SEH_begin_bn_power5 wrt ..imagebase
-	DD	$L$SEH_end_bn_power5 wrt ..imagebase
-	DD	$L$SEH_info_bn_power5 wrt ..imagebase
+	DD	$L$SEH_begin_bn_power5_nohw wrt ..imagebase
+	DD	$L$SEH_end_bn_power5_nohw wrt ..imagebase
+	DD	$L$SEH_info_bn_power5_nohw wrt ..imagebase
 %ifndef MY_ASSEMBLER_IS_TOO_OLD_FOR_512AVX
 	DD	$L$SEH_begin_bn_mulx4x_mont_gather5 wrt ..imagebase
 	DD	$L$SEH_end_bn_mulx4x_mont_gather5 wrt ..imagebase
@@ -3836,7 +3815,7 @@ ALIGN	4
 
 section	.xdata rdata align=4
 ALIGN	4
-$L$SEH_info_bn_mul_mont_gather5:
+$L$SEH_info_bn_mul_mont_gather5_nohw:
 	DB	9,0,0,0
 	DD	mul_handler wrt ..imagebase
 	DD	$L$mul_body wrt ..imagebase,$L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase
@@ -3846,7 +3825,7 @@ $L$SEH_info_bn_mul4x_mont_gather5:
 	DD	mul_handler wrt ..imagebase
 	DD	$L$mul4x_prologue wrt ..imagebase,$L$mul4x_body wrt ..imagebase,$L$mul4x_epilogue wrt ..imagebase
 ALIGN	4
-$L$SEH_info_bn_power5:
+$L$SEH_info_bn_power5_nohw:
 	DB	9,0,0,0
 	DD	mul_handler wrt ..imagebase
 	DD	$L$power5_prologue wrt ..imagebase,$L$power5_body wrt ..imagebase,$L$power5_epilogue wrt ..imagebase
diff --git a/include/openssl/cpu.h b/include/openssl/cpu.h
index d865020c1c..742a7a1d3a 100644
--- a/include/openssl/cpu.h
+++ b/include/openssl/cpu.h
@@ -14,5 +14,12 @@
 
 // This header is provided for compatibility with older revisions of BoringSSL.
 // TODO(davidben): Remove this header.
+#ifndef OPENSSL_HEADER_CPU_H
+#define OPENSSL_HEADER_CPU_H
 
 #include "crypto.h"
+
+__attribute__((unused))
+void log_dispatch(size_t id);
+
+#endif // OPENSSL_HEADER_CPU_H