diff --git a/libjpeg-turbo-CET.patch b/libjpeg-turbo-CET.patch
index 1905fde..75e0cea 100644
--- a/libjpeg-turbo-CET.patch
+++ b/libjpeg-turbo-CET.patch
@@ -1,6 +1,6 @@
-From ec7dcec4d423b8af21ef32381de5a4c1f7a541ec Mon Sep 17 00:00:00 2001
+From db1afe88c361ceea22c6ac8abdb9b0ff41a39aa2 Mon Sep 17 00:00:00 2001
 From: =?UTF-8?q?Nikola=20Forr=C3=B3?= <nforro@redhat.com>
-Date: Thu, 3 Dec 2020 18:19:33 +0100
+Date: Mon, 15 Apr 2019 16:56:45 +0200
 Subject: [PATCH] x86 SIMD: Add endbr32/endbr64 instructions
 MIME-Version: 1.0
 Content-Type: text/plain; charset=UTF-8
@@ -92,7 +92,7 @@ index c46d684..7dc6e08 100644
 +++ b/simd/i386/jccolext-avx2.asm
 @@ -42,6 +42,7 @@
      GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_avx2)
- 
+
  EXTN(jsimd_rgb_ycc_convert_avx2):
 +    _endbr32
      push        ebp
@@ -104,7 +104,7 @@ index 6357a42..8048abb 100644
 +++ b/simd/i386/jccolext-mmx.asm
 @@ -42,6 +42,7 @@
      GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_mmx)
- 
+
  EXTN(jsimd_rgb_ycc_convert_mmx):
 +    _endbr32
      push        ebp
@@ -116,7 +116,7 @@ index c6c8085..5307ddc 100644
 +++ b/simd/i386/jccolext-sse2.asm
 @@ -41,6 +41,7 @@
      GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_sse2)
- 
+
  EXTN(jsimd_rgb_ycc_convert_sse2):
 +    _endbr32
      push        ebp
@@ -128,7 +128,7 @@ index 3fa7973..27a0e11 100644
 +++ b/simd/i386/jcgryext-avx2.asm
 @@ -42,6 +42,7 @@
      GLOBAL_FUNCTION(jsimd_rgb_gray_convert_avx2)
- 
+
  EXTN(jsimd_rgb_gray_convert_avx2):
 +    _endbr32
      push        ebp
@@ -140,7 +140,7 @@ index 8af42e5..dda0e05 100644
 +++ b/simd/i386/jcgryext-mmx.asm
 @@ -42,6 +42,7 @@
      GLOBAL_FUNCTION(jsimd_rgb_gray_convert_mmx)
- 
+
  EXTN(jsimd_rgb_gray_convert_mmx):
 +    _endbr32
      push        ebp
@@ -152,22 +152,22 @@ index c9d6ff1..f8835bb 100644
 +++ b/simd/i386/jcgryext-sse2.asm
 @@ -41,6 +41,7 @@
      GLOBAL_FUNCTION(jsimd_rgb_gray_convert_sse2)
- 
+
  EXTN(jsimd_rgb_gray_convert_sse2):
 +    _endbr32
      push        ebp
      mov         eax, esp                     ; eax = original ebp
      sub         esp, byte 4
 diff --git a/simd/i386/jchuff-sse2.asm b/simd/i386/jchuff-sse2.asm
-index 278cf5e..b16b974 100644
+index 76cc85f..0217480 100644
 --- a/simd/i386/jchuff-sse2.asm
 +++ b/simd/i386/jchuff-sse2.asm
-@@ -349,6 +349,7 @@ times 1 << 14 db 15
+@@ -350,6 +350,7 @@ times 1 << 14 db 15
      GLOBAL_FUNCTION(jsimd_huff_encode_one_block_sse2)
- 
+
  EXTN(jsimd_huff_encode_one_block_sse2):
 +    _endbr32
- 
+
  %assign stack_offset      0
  %define arg_state         4 + stack_offset
 diff --git a/simd/i386/jcphuff-sse2.asm b/simd/i386/jcphuff-sse2.asm
@@ -176,7 +176,7 @@ index c26b48a..7fb01e5 100644
 +++ b/simd/i386/jcphuff-sse2.asm
 @@ -281,6 +281,7 @@
      GLOBAL_FUNCTION(jsimd_encode_mcu_AC_first_prepare_sse2)
- 
+
  EXTN(jsimd_encode_mcu_AC_first_prepare_sse2):
 +    _endbr32
      push        ebp
@@ -184,7 +184,7 @@ index c26b48a..7fb01e5 100644
      sub         esp, byte 4
 @@ -460,6 +461,7 @@ EXTN(jsimd_encode_mcu_AC_first_prepare_sse2):
      GLOBAL_FUNCTION(jsimd_encode_mcu_AC_refine_prepare_sse2)
- 
+
  EXTN(jsimd_encode_mcu_AC_refine_prepare_sse2):
 +    _endbr32
      push        ebp
@@ -196,7 +196,7 @@ index 0a20802..46eba8c 100644
 +++ b/simd/i386/jcsample-avx2.asm
 @@ -43,6 +43,7 @@
      GLOBAL_FUNCTION(jsimd_h2v1_downsample_avx2)
- 
+
  EXTN(jsimd_h2v1_downsample_avx2):
 +    _endbr32
      push        ebp
@@ -204,7 +204,7 @@ index 0a20802..46eba8c 100644
  ;   push        ebx                     ; unused
 @@ -216,6 +217,7 @@ EXTN(jsimd_h2v1_downsample_avx2):
      GLOBAL_FUNCTION(jsimd_h2v2_downsample_avx2)
- 
+
  EXTN(jsimd_h2v2_downsample_avx2):
 +    _endbr32
      push        ebp
@@ -216,7 +216,7 @@ index 2c223ee..b2b8ded 100644
 +++ b/simd/i386/jcsample-mmx.asm
 @@ -42,6 +42,7 @@
      GLOBAL_FUNCTION(jsimd_h2v1_downsample_mmx)
- 
+
  EXTN(jsimd_h2v1_downsample_mmx):
 +    _endbr32
      push        ebp
@@ -224,7 +224,7 @@ index 2c223ee..b2b8ded 100644
  ;   push        ebx                     ; unused
 @@ -185,6 +186,7 @@ EXTN(jsimd_h2v1_downsample_mmx):
      GLOBAL_FUNCTION(jsimd_h2v2_downsample_mmx)
- 
+
  EXTN(jsimd_h2v2_downsample_mmx):
 +    _endbr32
      push        ebp
@@ -236,7 +236,7 @@ index 4fea60d..4c22b40 100644
 +++ b/simd/i386/jcsample-sse2.asm
 @@ -42,6 +42,7 @@
      GLOBAL_FUNCTION(jsimd_h2v1_downsample_sse2)
- 
+
  EXTN(jsimd_h2v1_downsample_sse2):
 +    _endbr32
      push        ebp
@@ -244,7 +244,7 @@ index 4fea60d..4c22b40 100644
  ;   push        ebx                     ; unused
 @@ -198,6 +199,7 @@ EXTN(jsimd_h2v1_downsample_sse2):
      GLOBAL_FUNCTION(jsimd_h2v2_downsample_sse2)
- 
+
  EXTN(jsimd_h2v2_downsample_sse2):
 +    _endbr32
      push        ebp
@@ -256,7 +256,7 @@ index 015be04..b076765 100644
 +++ b/simd/i386/jdcolext-avx2.asm
 @@ -43,6 +43,7 @@
      GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_avx2)
- 
+
  EXTN(jsimd_ycc_rgb_convert_avx2):
 +    _endbr32
      push        ebp
@@ -268,7 +268,7 @@ index 5813cfc..150f5b6 100644
 +++ b/simd/i386/jdcolext-mmx.asm
 @@ -42,6 +42,7 @@
      GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_mmx)
- 
+
  EXTN(jsimd_ycc_rgb_convert_mmx):
 +    _endbr32
      push        ebp
@@ -280,7 +280,7 @@ index d5572b3..cd3ac70 100644
 +++ b/simd/i386/jdcolext-sse2.asm
 @@ -42,6 +42,7 @@
      GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_sse2)
- 
+
  EXTN(jsimd_ycc_rgb_convert_sse2):
 +    _endbr32
      push        ebp
@@ -292,7 +292,7 @@ index e35f728..0db0aa4 100644
 +++ b/simd/i386/jdmrgext-avx2.asm
 @@ -43,6 +43,7 @@
      GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_avx2)
- 
+
  EXTN(jsimd_h2v1_merged_upsample_avx2):
 +    _endbr32
      push        ebp
@@ -300,7 +300,7 @@ index e35f728..0db0aa4 100644
      sub         esp, byte 4
 @@ -523,6 +524,7 @@ EXTN(jsimd_h2v1_merged_upsample_avx2):
      GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_avx2)
- 
+
  EXTN(jsimd_h2v2_merged_upsample_avx2):
 +    _endbr32
      push        ebp
@@ -312,7 +312,7 @@ index eb3e36b..6427a1a 100644
 +++ b/simd/i386/jdmrgext-mmx.asm
 @@ -40,6 +40,7 @@
      GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_mmx)
- 
+
  EXTN(jsimd_h2v1_merged_upsample_mmx):
 +    _endbr32
      push        ebp
@@ -320,7 +320,7 @@ index eb3e36b..6427a1a 100644
      sub         esp, byte 4
 @@ -408,6 +409,7 @@ EXTN(jsimd_h2v1_merged_upsample_mmx):
      GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_mmx)
- 
+
  EXTN(jsimd_h2v2_merged_upsample_mmx):
 +    _endbr32
      push        ebp
@@ -332,7 +332,7 @@ index c113dc4..6897fa1 100644
 +++ b/simd/i386/jdmrgext-sse2.asm
 @@ -42,6 +42,7 @@
      GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_sse2)
- 
+
  EXTN(jsimd_h2v1_merged_upsample_sse2):
 +    _endbr32
      push        ebp
@@ -340,7 +340,7 @@ index c113dc4..6897fa1 100644
      sub         esp, byte 4
 @@ -465,6 +466,7 @@ EXTN(jsimd_h2v1_merged_upsample_sse2):
      GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_sse2)
- 
+
  EXTN(jsimd_h2v2_merged_upsample_sse2):
 +    _endbr32
      push        ebp
@@ -352,7 +352,7 @@ index a800c35..7d52708 100644
 +++ b/simd/i386/jdsample-avx2.asm
 @@ -60,6 +60,7 @@ PW_EIGHT times 16 dw 8
      GLOBAL_FUNCTION(jsimd_h2v1_fancy_upsample_avx2)
- 
+
  EXTN(jsimd_h2v1_fancy_upsample_avx2):
 +    _endbr32
      push        ebp
@@ -360,7 +360,7 @@ index a800c35..7d52708 100644
      pushpic     ebx
 @@ -227,6 +228,7 @@ EXTN(jsimd_h2v1_fancy_upsample_avx2):
      GLOBAL_FUNCTION(jsimd_h2v2_fancy_upsample_avx2)
- 
+
  EXTN(jsimd_h2v2_fancy_upsample_avx2):
 +    _endbr32
      push        ebp
@@ -368,7 +368,7 @@ index a800c35..7d52708 100644
      sub         esp, byte 4
 @@ -570,6 +572,7 @@ EXTN(jsimd_h2v2_fancy_upsample_avx2):
      GLOBAL_FUNCTION(jsimd_h2v1_upsample_avx2)
- 
+
  EXTN(jsimd_h2v1_upsample_avx2):
 +    _endbr32
      push        ebp
@@ -376,7 +376,7 @@ index a800c35..7d52708 100644
  ;   push        ebx                     ; unused
 @@ -669,6 +672,7 @@ EXTN(jsimd_h2v1_upsample_avx2):
      GLOBAL_FUNCTION(jsimd_h2v2_upsample_avx2)
- 
+
  EXTN(jsimd_h2v2_upsample_avx2):
 +    _endbr32
      push        ebp
@@ -388,7 +388,7 @@ index 12c49f0..7f2ab40 100644
 +++ b/simd/i386/jdsample-mmx.asm
 @@ -59,6 +59,7 @@ PW_EIGHT times 4 dw 8
      GLOBAL_FUNCTION(jsimd_h2v1_fancy_upsample_mmx)
- 
+
  EXTN(jsimd_h2v1_fancy_upsample_mmx):
 +    _endbr32
      push        ebp
@@ -396,7 +396,7 @@ index 12c49f0..7f2ab40 100644
      pushpic     ebx
 @@ -217,6 +218,7 @@ EXTN(jsimd_h2v1_fancy_upsample_mmx):
      GLOBAL_FUNCTION(jsimd_h2v2_fancy_upsample_mmx)
- 
+
  EXTN(jsimd_h2v2_fancy_upsample_mmx):
 +    _endbr32
      push        ebp
@@ -404,7 +404,7 @@ index 12c49f0..7f2ab40 100644
      sub         esp, byte 4
 @@ -541,6 +543,7 @@ EXTN(jsimd_h2v2_fancy_upsample_mmx):
      GLOBAL_FUNCTION(jsimd_h2v1_upsample_mmx)
- 
+
  EXTN(jsimd_h2v1_upsample_mmx):
 +    _endbr32
      push        ebp
@@ -412,7 +412,7 @@ index 12c49f0..7f2ab40 100644
  ;   push        ebx                     ; unused
 @@ -640,6 +643,7 @@ EXTN(jsimd_h2v1_upsample_mmx):
      GLOBAL_FUNCTION(jsimd_h2v2_upsample_mmx)
- 
+
  EXTN(jsimd_h2v2_upsample_mmx):
 +    _endbr32
      push        ebp
@@ -424,7 +424,7 @@ index 4e28d2f..3311b25 100644
 +++ b/simd/i386/jdsample-sse2.asm
 @@ -59,6 +59,7 @@ PW_EIGHT times 8 dw 8
      GLOBAL_FUNCTION(jsimd_h2v1_fancy_upsample_sse2)
- 
+
  EXTN(jsimd_h2v1_fancy_upsample_sse2):
 +    _endbr32
      push        ebp
@@ -432,7 +432,7 @@ index 4e28d2f..3311b25 100644
      pushpic     ebx
 @@ -216,6 +217,7 @@ EXTN(jsimd_h2v1_fancy_upsample_sse2):
      GLOBAL_FUNCTION(jsimd_h2v2_fancy_upsample_sse2)
- 
+
  EXTN(jsimd_h2v2_fancy_upsample_sse2):
 +    _endbr32
      push        ebp
@@ -440,7 +440,7 @@ index 4e28d2f..3311b25 100644
      sub         esp, byte 4
 @@ -538,6 +540,7 @@ EXTN(jsimd_h2v2_fancy_upsample_sse2):
      GLOBAL_FUNCTION(jsimd_h2v1_upsample_sse2)
- 
+
  EXTN(jsimd_h2v1_upsample_sse2):
 +    _endbr32
      push        ebp
@@ -448,7 +448,7 @@ index 4e28d2f..3311b25 100644
  ;   push        ebx                     ; unused
 @@ -635,6 +638,7 @@ EXTN(jsimd_h2v1_upsample_sse2):
      GLOBAL_FUNCTION(jsimd_h2v2_upsample_sse2)
- 
+
  EXTN(jsimd_h2v2_upsample_sse2):
 +    _endbr32
      push        ebp
@@ -460,7 +460,7 @@ index 322ab16..109e36e 100644
 +++ b/simd/i386/jfdctflt-3dn.asm
 @@ -56,6 +56,7 @@ PD_1_306 times 2 dd 1.306562964876376527856643
      GLOBAL_FUNCTION(jsimd_fdct_float_3dnow)
- 
+
  EXTN(jsimd_fdct_float_3dnow):
 +    _endbr32
      push        ebp
@@ -472,7 +472,7 @@ index 86952c6..b1e0576 100644
 +++ b/simd/i386/jfdctflt-sse.asm
 @@ -67,6 +67,7 @@ PD_1_306 times 4 dd 1.306562964876376527856643
      GLOBAL_FUNCTION(jsimd_fdct_float_sse)
- 
+
  EXTN(jsimd_fdct_float_sse):
 +    _endbr32
      push        ebp
@@ -484,7 +484,7 @@ index 80645a5..be84fdb 100644
 +++ b/simd/i386/jfdctfst-mmx.asm
 @@ -81,6 +81,7 @@ PW_F1306 times 4 dw F_1_306 << CONST_SHIFT
      GLOBAL_FUNCTION(jsimd_fdct_ifast_mmx)
- 
+
  EXTN(jsimd_fdct_ifast_mmx):
 +    _endbr32
      push        ebp
@@ -496,7 +496,7 @@ index 446fa7a..945f9cf 100644
 +++ b/simd/i386/jfdctfst-sse2.asm
 @@ -82,6 +82,7 @@ PW_F1306 times 8 dw F_1_306 << CONST_SHIFT
      GLOBAL_FUNCTION(jsimd_fdct_ifast_sse2)
- 
+
  EXTN(jsimd_fdct_ifast_sse2):
 +    _endbr32
      push        ebp
@@ -508,7 +508,7 @@ index 23cf733..56acb63 100644
 +++ b/simd/i386/jfdctint-avx2.asm
 @@ -260,6 +260,7 @@ PW_1_NEG1                  times 8  dw  1
      GLOBAL_FUNCTION(jsimd_fdct_islow_avx2)
- 
+
  EXTN(jsimd_fdct_islow_avx2):
 +    _endbr32
      push        ebp
@@ -520,7 +520,7 @@ index 34a43b9..4d1e773 100644
 +++ b/simd/i386/jfdctint-mmx.asm
 @@ -102,6 +102,7 @@ PW_DESCALE_P2X times 4 dw  1 << (PASS1_BITS - 1)
      GLOBAL_FUNCTION(jsimd_fdct_islow_mmx)
- 
+
  EXTN(jsimd_fdct_islow_mmx):
 +    _endbr32
      push        ebp
@@ -532,7 +532,7 @@ index 6f8e18c..3954c8f 100644
 +++ b/simd/i386/jfdctint-sse2.asm
 @@ -103,6 +103,7 @@ PW_DESCALE_P2X times 8 dw  1 << (PASS1_BITS - 1)
      GLOBAL_FUNCTION(jsimd_fdct_islow_sse2)
- 
+
  EXTN(jsimd_fdct_islow_sse2):
 +    _endbr32
      push        ebp
@@ -544,7 +544,7 @@ index 8795191..eb49902 100644
 +++ b/simd/i386/jidctflt-3dn.asm
 @@ -65,6 +65,7 @@ PB_CENTERJSAMP  times 8 db CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_float_3dnow)
- 
+
  EXTN(jsimd_idct_float_3dnow):
 +    _endbr32
      push        ebp
@@ -556,7 +556,7 @@ index b27ecfd..ffe54f8 100644
 +++ b/simd/i386/jidctflt-sse.asm
 @@ -75,6 +75,7 @@ PB_CENTERJSAMP times 8 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_float_sse)
- 
+
  EXTN(jsimd_idct_float_sse):
 +    _endbr32
      push        ebp
@@ -568,7 +568,7 @@ index c646eae..fd1fe35 100644
 +++ b/simd/i386/jidctflt-sse2.asm
 @@ -75,6 +75,7 @@ PB_CENTERJSAMP  times 16 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_float_sse2)
- 
+
  EXTN(jsimd_idct_float_sse2):
 +    _endbr32
      push        ebp
@@ -580,7 +580,7 @@ index 24622d4..00940b8 100644
 +++ b/simd/i386/jidctfst-mmx.asm
 @@ -96,6 +96,7 @@ PB_CENTERJSAMP times 8 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_ifast_mmx)
- 
+
  EXTN(jsimd_idct_ifast_mmx):
 +    _endbr32
      push        ebp
@@ -592,7 +592,7 @@ index 19704ff..1f4af33 100644
 +++ b/simd/i386/jidctfst-sse2.asm
 @@ -94,6 +94,7 @@ PB_CENTERJSAMP times 16 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_ifast_sse2)
- 
+
  EXTN(jsimd_idct_ifast_sse2):
 +    _endbr32
      push        ebp
@@ -604,7 +604,7 @@ index 199c7df..2eb606a 100644
 +++ b/simd/i386/jidctint-avx2.asm
 @@ -296,6 +296,7 @@ PW_1_NEG1                  times 8  dw  1
      GLOBAL_FUNCTION(jsimd_idct_islow_avx2)
- 
+
  EXTN(jsimd_idct_islow_avx2):
 +    _endbr32
      push        ebp
@@ -616,7 +616,7 @@ index f15c8d3..2d91b7e 100644
 +++ b/simd/i386/jidctint-mmx.asm
 @@ -109,6 +109,7 @@ PB_CENTERJSAMP times 8 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_islow_mmx)
- 
+
  EXTN(jsimd_idct_islow_mmx):
 +    _endbr32
      push        ebp
@@ -628,7 +628,7 @@ index 43e3201..804be19 100644
 +++ b/simd/i386/jidctint-sse2.asm
 @@ -107,6 +107,7 @@ PB_CENTERJSAMP times 16 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_islow_sse2)
- 
+
  EXTN(jsimd_idct_islow_sse2):
 +    _endbr32
      push        ebp
@@ -640,7 +640,7 @@ index e2307e1..cb43106 100644
 +++ b/simd/i386/jidctred-mmx.asm
 @@ -117,6 +117,7 @@ PB_CENTERJSAMP  times 8 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_4x4_mmx)
- 
+
  EXTN(jsimd_idct_4x4_mmx):
 +    _endbr32
      push        ebp
@@ -648,7 +648,7 @@ index e2307e1..cb43106 100644
      sub         esp, byte 4
 @@ -504,6 +505,7 @@ EXTN(jsimd_idct_4x4_mmx):
      GLOBAL_FUNCTION(jsimd_idct_2x2_mmx)
- 
+
  EXTN(jsimd_idct_2x2_mmx):
 +    _endbr32
      push        ebp
@@ -660,7 +660,7 @@ index 6e56494..2a61b9e 100644
 +++ b/simd/i386/jidctred-sse2.asm
 @@ -115,6 +115,7 @@ PB_CENTERJSAMP  times 16 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_4x4_sse2)
- 
+
  EXTN(jsimd_idct_4x4_sse2):
 +    _endbr32
      push        ebp
@@ -668,7 +668,7 @@ index 6e56494..2a61b9e 100644
      sub         esp, byte 4
 @@ -425,6 +426,7 @@ EXTN(jsimd_idct_4x4_sse2):
      GLOBAL_FUNCTION(jsimd_idct_2x2_sse2)
- 
+
  EXTN(jsimd_idct_2x2_sse2):
 +    _endbr32
      push        ebp
@@ -680,7 +680,7 @@ index 5cb60ca..a0599eb 100644
 +++ b/simd/i386/jquant-3dn.asm
 @@ -36,6 +36,7 @@
      GLOBAL_FUNCTION(jsimd_convsamp_float_3dnow)
- 
+
  EXTN(jsimd_convsamp_float_3dnow):
 +    _endbr32
      push        ebp
@@ -688,7 +688,7 @@ index 5cb60ca..a0599eb 100644
      push        ebx
 @@ -138,6 +139,7 @@ EXTN(jsimd_convsamp_float_3dnow):
      GLOBAL_FUNCTION(jsimd_quantize_float_3dnow)
- 
+
  EXTN(jsimd_quantize_float_3dnow):
 +    _endbr32
      push        ebp
@@ -700,7 +700,7 @@ index 61305c6..080021b 100644
 +++ b/simd/i386/jquant-mmx.asm
 @@ -36,6 +36,7 @@
      GLOBAL_FUNCTION(jsimd_convsamp_mmx)
- 
+
  EXTN(jsimd_convsamp_mmx):
 +    _endbr32
      push        ebp
@@ -708,7 +708,7 @@ index 61305c6..080021b 100644
      push        ebx
 @@ -145,6 +146,7 @@ EXTN(jsimd_convsamp_mmx):
      GLOBAL_FUNCTION(jsimd_quantize_mmx)
- 
+
  EXTN(jsimd_quantize_mmx):
 +    _endbr32
      push        ebp
@@ -720,7 +720,7 @@ index 218adc9..cacd2a9 100644
 +++ b/simd/i386/jquant-sse.asm
 @@ -36,6 +36,7 @@
      GLOBAL_FUNCTION(jsimd_convsamp_float_sse)
- 
+
  EXTN(jsimd_convsamp_float_sse):
 +    _endbr32
      push        ebp
@@ -728,7 +728,7 @@ index 218adc9..cacd2a9 100644
      push        ebx
 @@ -138,6 +139,7 @@ EXTN(jsimd_convsamp_float_sse):
      GLOBAL_FUNCTION(jsimd_quantize_float_sse)
- 
+
  EXTN(jsimd_quantize_float_sse):
 +    _endbr32
      push        ebp
@@ -740,7 +740,7 @@ index a881ab5..6f4789c 100644
 +++ b/simd/i386/jquantf-sse2.asm
 @@ -36,6 +36,7 @@
      GLOBAL_FUNCTION(jsimd_convsamp_float_sse2)
- 
+
  EXTN(jsimd_convsamp_float_sse2):
 +    _endbr32
      push        ebp
@@ -748,7 +748,7 @@ index a881ab5..6f4789c 100644
      push        ebx
 @@ -115,6 +116,7 @@ EXTN(jsimd_convsamp_float_sse2):
      GLOBAL_FUNCTION(jsimd_quantize_float_sse2)
- 
+
  EXTN(jsimd_quantize_float_sse2):
 +    _endbr32
      push        ebp
@@ -760,7 +760,7 @@ index 5ed6bec..efcddd2 100644
 +++ b/simd/i386/jquanti-avx2.asm
 @@ -37,6 +37,7 @@
      GLOBAL_FUNCTION(jsimd_convsamp_avx2)
- 
+
  EXTN(jsimd_convsamp_avx2):
 +    _endbr32
      push        ebp
@@ -768,7 +768,7 @@ index 5ed6bec..efcddd2 100644
      push        ebx
 @@ -130,6 +131,7 @@ EXTN(jsimd_convsamp_avx2):
      GLOBAL_FUNCTION(jsimd_quantize_avx2)
- 
+
  EXTN(jsimd_quantize_avx2):
 +    _endbr32
      push        ebp
@@ -780,7 +780,7 @@ index 0a50940..98d39e0 100644
 +++ b/simd/i386/jquanti-sse2.asm
 @@ -36,6 +36,7 @@
      GLOBAL_FUNCTION(jsimd_convsamp_sse2)
- 
+
  EXTN(jsimd_convsamp_sse2):
 +    _endbr32
      push        ebp
@@ -788,20 +788,20 @@ index 0a50940..98d39e0 100644
      push        ebx
 @@ -121,6 +122,7 @@ EXTN(jsimd_convsamp_sse2):
      GLOBAL_FUNCTION(jsimd_quantize_sse2)
- 
+
  EXTN(jsimd_quantize_sse2):
 +    _endbr32
      push        ebp
      mov         ebp, esp
  ;   push        ebx                     ; unused
 diff --git a/simd/nasm/jsimdext.inc b/simd/nasm/jsimdext.inc
-index e8d50b0..f140d4d 100644
+index bebcb20..4b0cd69 100644
 --- a/simd/nasm/jsimdext.inc
 +++ b/simd/nasm/jsimdext.inc
-@@ -512,6 +512,14 @@ const_base:
- 
+@@ -513,6 +513,14 @@ const_base:
+
  %endif
- 
+
 +%imacro _endbr32 0
 +    dd 0xfb1e0ff3
 +%endmacro
@@ -814,437 +814,435 @@ index e8d50b0..f140d4d 100644
  ;  Defines picked up from the C headers
  ;
 diff --git a/simd/x86_64/jccolext-avx2.asm b/simd/x86_64/jccolext-avx2.asm
-index ffb527d..5c01670 100644
+index dd7ea39..e6c8283 100644
 --- a/simd/x86_64/jccolext-avx2.asm
 +++ b/simd/x86_64/jccolext-avx2.asm
-@@ -40,6 +40,7 @@
+@@ -41,6 +41,7 @@
      GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_avx2)
- 
+
  EXTN(jsimd_rgb_ycc_convert_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jccolext-sse2.asm b/simd/x86_64/jccolext-sse2.asm
-index af70ed6..99c1889 100644
+index bc1e817..4bb1af6 100644
 --- a/simd/x86_64/jccolext-sse2.asm
 +++ b/simd/x86_64/jccolext-sse2.asm
-@@ -39,6 +39,7 @@
+@@ -40,6 +40,7 @@
      GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_sse2)
- 
+
  EXTN(jsimd_rgb_ycc_convert_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jcgryext-avx2.asm b/simd/x86_64/jcgryext-avx2.asm
-index ddcc2c0..a5b9619 100644
+index c8c8d12..12ecb7e 100644
 --- a/simd/x86_64/jcgryext-avx2.asm
 +++ b/simd/x86_64/jcgryext-avx2.asm
-@@ -40,6 +40,7 @@
+@@ -41,6 +41,7 @@
      GLOBAL_FUNCTION(jsimd_rgb_gray_convert_avx2)
- 
+
  EXTN(jsimd_rgb_gray_convert_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jcgryext-sse2.asm b/simd/x86_64/jcgryext-sse2.asm
-index f1d399a..550fc2f 100644
+index 7e5a0f2..e3a2413 100644
 --- a/simd/x86_64/jcgryext-sse2.asm
 +++ b/simd/x86_64/jcgryext-sse2.asm
-@@ -39,6 +39,7 @@
+@@ -40,6 +40,7 @@
      GLOBAL_FUNCTION(jsimd_rgb_gray_convert_sse2)
- 
+
  EXTN(jsimd_rgb_gray_convert_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jchuff-sse2.asm b/simd/x86_64/jchuff-sse2.asm
-index 9ea6df9..7e5ca30 100644
+index 0c2cdd6..44ea81d 100644
 --- a/simd/x86_64/jchuff-sse2.asm
 +++ b/simd/x86_64/jchuff-sse2.asm
-@@ -259,6 +259,7 @@ times 1 << 15 db 16
+@@ -261,6 +261,7 @@ times 1 << 15 db 16
      GLOBAL_FUNCTION(jsimd_huff_encode_one_block_sse2)
- 
+
  EXTN(jsimd_huff_encode_one_block_sse2):
 +    _endbr64
- 
- %ifdef WIN64
- 
+     push        rbp
+     mov         rbp, rsp
+
 diff --git a/simd/x86_64/jcphuff-sse2.asm b/simd/x86_64/jcphuff-sse2.asm
-index 01b5c02..4625c11 100644
+index 11db4b2..2157e97 100644
 --- a/simd/x86_64/jcphuff-sse2.asm
 +++ b/simd/x86_64/jcphuff-sse2.asm
-@@ -281,6 +281,7 @@
+@@ -282,6 +282,7 @@
      GLOBAL_FUNCTION(jsimd_encode_mcu_AC_first_prepare_sse2)
- 
+
  EXTN(jsimd_encode_mcu_AC_first_prepare_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
-@@ -449,6 +450,7 @@ EXTN(jsimd_encode_mcu_AC_first_prepare_sse2):
+     mov         rbp, rsp
+     and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
+@@ -445,6 +446,7 @@ EXTN(jsimd_encode_mcu_AC_first_prepare_sse2):
      GLOBAL_FUNCTION(jsimd_encode_mcu_AC_refine_prepare_sse2)
- 
+
  EXTN(jsimd_encode_mcu_AC_refine_prepare_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
 diff --git a/simd/x86_64/jcsample-avx2.asm b/simd/x86_64/jcsample-avx2.asm
-index b32527a..28bf70a 100644
+index 589c52b..7d8d4e0 100644
 --- a/simd/x86_64/jcsample-avx2.asm
 +++ b/simd/x86_64/jcsample-avx2.asm
 @@ -44,6 +44,7 @@
      GLOBAL_FUNCTION(jsimd_h2v1_downsample_avx2)
- 
+
  EXTN(jsimd_h2v1_downsample_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
-@@ -206,6 +207,7 @@ EXTN(jsimd_h2v1_downsample_avx2):
+     collect_args 6
+@@ -205,6 +206,7 @@ EXTN(jsimd_h2v1_downsample_avx2):
      GLOBAL_FUNCTION(jsimd_h2v2_downsample_avx2)
- 
+
  EXTN(jsimd_h2v2_downsample_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     collect_args 6
 diff --git a/simd/x86_64/jcsample-sse2.asm b/simd/x86_64/jcsample-sse2.asm
-index 2fcfe45..2093090 100644
+index 7a4f1bc..8932b94 100644
 --- a/simd/x86_64/jcsample-sse2.asm
 +++ b/simd/x86_64/jcsample-sse2.asm
 @@ -43,6 +43,7 @@
      GLOBAL_FUNCTION(jsimd_h2v1_downsample_sse2)
- 
+
  EXTN(jsimd_h2v1_downsample_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
-@@ -188,6 +189,7 @@ EXTN(jsimd_h2v1_downsample_sse2):
+     collect_args 6
+@@ -187,6 +188,7 @@ EXTN(jsimd_h2v1_downsample_sse2):
      GLOBAL_FUNCTION(jsimd_h2v2_downsample_sse2)
- 
+
  EXTN(jsimd_h2v2_downsample_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     collect_args 6
 diff --git a/simd/x86_64/jdcolext-avx2.asm b/simd/x86_64/jdcolext-avx2.asm
-index 2370fda..406fc01 100644
+index 070436c..6c0a212 100644
 --- a/simd/x86_64/jdcolext-avx2.asm
 +++ b/simd/x86_64/jdcolext-avx2.asm
-@@ -41,6 +41,7 @@
+@@ -42,6 +42,7 @@
      GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_avx2)
- 
+
  EXTN(jsimd_ycc_rgb_convert_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jdcolext-sse2.asm b/simd/x86_64/jdcolext-sse2.asm
-index e07c8d7..84acb50 100644
+index bba3a30..28d05e4 100644
 --- a/simd/x86_64/jdcolext-sse2.asm
 +++ b/simd/x86_64/jdcolext-sse2.asm
-@@ -40,6 +40,7 @@
+@@ -41,6 +41,7 @@
      GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_sse2)
- 
+
  EXTN(jsimd_ycc_rgb_convert_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jdmrgext-avx2.asm b/simd/x86_64/jdmrgext-avx2.asm
-index 8b264b4..c42f395 100644
+index 1191645..f58384c 100644
 --- a/simd/x86_64/jdmrgext-avx2.asm
 +++ b/simd/x86_64/jdmrgext-avx2.asm
-@@ -41,6 +41,7 @@
+@@ -42,6 +42,7 @@
      GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_avx2)
- 
+
  EXTN(jsimd_h2v1_merged_upsample_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
-@@ -505,6 +506,7 @@ EXTN(jsimd_h2v1_merged_upsample_avx2):
+     mov         rbp, rsp
+     push        r15
+@@ -506,6 +507,7 @@ EXTN(jsimd_h2v1_merged_upsample_avx2):
      GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_avx2)
- 
+
  EXTN(jsimd_h2v2_merged_upsample_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     collect_args 4
 diff --git a/simd/x86_64/jdmrgext-sse2.asm b/simd/x86_64/jdmrgext-sse2.asm
-index eb3ab9d..8347620 100644
+index 8988dd0..8641aa3 100644
 --- a/simd/x86_64/jdmrgext-sse2.asm
 +++ b/simd/x86_64/jdmrgext-sse2.asm
-@@ -40,6 +40,7 @@
+@@ -41,6 +41,7 @@
      GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_sse2)
- 
+
  EXTN(jsimd_h2v1_merged_upsample_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
-@@ -447,6 +448,7 @@ EXTN(jsimd_h2v1_merged_upsample_sse2):
+     mov         rbp, rsp
+     push        r15
+@@ -448,6 +449,7 @@ EXTN(jsimd_h2v1_merged_upsample_sse2):
      GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_sse2)
- 
+
  EXTN(jsimd_h2v2_merged_upsample_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     collect_args 4
 diff --git a/simd/x86_64/jdsample-avx2.asm b/simd/x86_64/jdsample-avx2.asm
-index 1e4979f..b80d607 100644
+index c6ddbb5..c5594ed 100644
 --- a/simd/x86_64/jdsample-avx2.asm
 +++ b/simd/x86_64/jdsample-avx2.asm
-@@ -61,6 +61,7 @@ PW_EIGHT times 16 dw 8
+@@ -62,6 +62,7 @@ PW_EIGHT times 16 dw 8
      GLOBAL_FUNCTION(jsimd_h2v1_fancy_upsample_avx2)
- 
+
  EXTN(jsimd_h2v1_fancy_upsample_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     push_xmm    3
 @@ -215,6 +216,7 @@ EXTN(jsimd_h2v1_fancy_upsample_avx2):
      GLOBAL_FUNCTION(jsimd_h2v2_fancy_upsample_avx2)
- 
+
  EXTN(jsimd_h2v2_fancy_upsample_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 @@ -524,6 +526,7 @@ EXTN(jsimd_h2v2_fancy_upsample_avx2):
      GLOBAL_FUNCTION(jsimd_h2v1_upsample_avx2)
- 
+
  EXTN(jsimd_h2v1_upsample_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
-@@ -613,6 +616,7 @@ EXTN(jsimd_h2v1_upsample_avx2):
+     collect_args 4
+@@ -612,6 +615,7 @@ EXTN(jsimd_h2v1_upsample_avx2):
      GLOBAL_FUNCTION(jsimd_h2v2_upsample_avx2)
- 
+
  EXTN(jsimd_h2v2_upsample_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     collect_args 4
 diff --git a/simd/x86_64/jdsample-sse2.asm b/simd/x86_64/jdsample-sse2.asm
-index 38dbcee..6a1e012 100644
+index 24cd389..27bf771 100644
 --- a/simd/x86_64/jdsample-sse2.asm
 +++ b/simd/x86_64/jdsample-sse2.asm
-@@ -60,6 +60,7 @@ PW_EIGHT times 8 dw 8
+@@ -61,6 +61,7 @@ PW_EIGHT times 8 dw 8
      GLOBAL_FUNCTION(jsimd_h2v1_fancy_upsample_sse2)
- 
+
  EXTN(jsimd_h2v1_fancy_upsample_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     collect_args 4
 @@ -202,6 +203,7 @@ EXTN(jsimd_h2v1_fancy_upsample_sse2):
      GLOBAL_FUNCTION(jsimd_h2v2_fancy_upsample_sse2)
- 
+
  EXTN(jsimd_h2v2_fancy_upsample_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 @@ -497,6 +499,7 @@ EXTN(jsimd_h2v2_fancy_upsample_sse2):
      GLOBAL_FUNCTION(jsimd_h2v1_upsample_sse2)
- 
+
  EXTN(jsimd_h2v1_upsample_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
-@@ -584,6 +587,7 @@ EXTN(jsimd_h2v1_upsample_sse2):
+     collect_args 4
+@@ -583,6 +586,7 @@ EXTN(jsimd_h2v1_upsample_sse2):
      GLOBAL_FUNCTION(jsimd_h2v2_upsample_sse2)
- 
+
  EXTN(jsimd_h2v2_upsample_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     collect_args 4
 diff --git a/simd/x86_64/jfdctflt-sse.asm b/simd/x86_64/jfdctflt-sse.asm
-index ef27966..90d06eb 100644
+index 3595496..b5abeee 100644
 --- a/simd/x86_64/jfdctflt-sse.asm
 +++ b/simd/x86_64/jfdctflt-sse.asm
-@@ -65,6 +65,7 @@ PD_1_306 times 4 dd 1.306562964876376527856643
+@@ -66,6 +66,7 @@ PD_1_306 times 4 dd 1.306562964876376527856643
      GLOBAL_FUNCTION(jsimd_fdct_float_sse)
- 
+
  EXTN(jsimd_fdct_float_sse):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jfdctfst-sse2.asm b/simd/x86_64/jfdctfst-sse2.asm
-index 2e1bfe6..5a74aaf 100644
+index d33c58a..78af5e4 100644
 --- a/simd/x86_64/jfdctfst-sse2.asm
 +++ b/simd/x86_64/jfdctfst-sse2.asm
-@@ -80,6 +80,7 @@ PW_F1306 times 8 dw F_1_306 << CONST_SHIFT
+@@ -81,6 +81,7 @@ PW_F1306 times 8 dw F_1_306 << CONST_SHIFT
      GLOBAL_FUNCTION(jsimd_fdct_ifast_sse2)
- 
+
  EXTN(jsimd_fdct_ifast_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jfdctint-avx2.asm b/simd/x86_64/jfdctint-avx2.asm
-index e56258b..2ac6d4f 100644
+index d0afe5e..b083fbb 100644
 --- a/simd/x86_64/jfdctint-avx2.asm
 +++ b/simd/x86_64/jfdctint-avx2.asm
 @@ -260,6 +260,7 @@ PW_1_NEG1                  times 8  dw  1
      GLOBAL_FUNCTION(jsimd_fdct_islow_avx2)
- 
+
  EXTN(jsimd_fdct_islow_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     collect_args 1
 diff --git a/simd/x86_64/jfdctint-sse2.asm b/simd/x86_64/jfdctint-sse2.asm
-index ec1f383..0950825 100644
+index 024ce90..88ea491 100644
 --- a/simd/x86_64/jfdctint-sse2.asm
 +++ b/simd/x86_64/jfdctint-sse2.asm
-@@ -101,6 +101,7 @@ PW_DESCALE_P2X times 8 dw  1 << (PASS1_BITS - 1)
+@@ -102,6 +102,7 @@ PW_DESCALE_P2X times 8 dw  1 << (PASS1_BITS - 1)
      GLOBAL_FUNCTION(jsimd_fdct_islow_sse2)
- 
+
  EXTN(jsimd_fdct_islow_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jidctflt-sse2.asm b/simd/x86_64/jidctflt-sse2.asm
-index 60bf961..f04902a 100644
+index 952fbe3..8610710 100644
 --- a/simd/x86_64/jidctflt-sse2.asm
 +++ b/simd/x86_64/jidctflt-sse2.asm
 @@ -76,6 +76,7 @@ PB_CENTERJSAMP  times 16 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_float_sse2)
- 
+
  EXTN(jsimd_idct_float_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jidctfst-sse2.asm b/simd/x86_64/jidctfst-sse2.asm
-index cb97fdf..d004da9 100644
+index a3da8d8..351ee06 100644
 --- a/simd/x86_64/jidctfst-sse2.asm
 +++ b/simd/x86_64/jidctfst-sse2.asm
 @@ -95,6 +95,7 @@ PB_CENTERJSAMP times 16 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_ifast_sse2)
- 
+
  EXTN(jsimd_idct_ifast_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jidctint-avx2.asm b/simd/x86_64/jidctint-avx2.asm
-index ca7e317..5c65fa7 100644
+index 528da01..00087c7 100644
 --- a/simd/x86_64/jidctint-avx2.asm
 +++ b/simd/x86_64/jidctint-avx2.asm
 @@ -282,6 +282,7 @@ PW_1_NEG1                  times 8  dw  1
      GLOBAL_FUNCTION(jsimd_idct_islow_avx2)
- 
+
  EXTN(jsimd_idct_islow_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
      mov         rbp, rsp                     ; rbp = aligned rbp
+     push_xmm    4
 diff --git a/simd/x86_64/jidctint-sse2.asm b/simd/x86_64/jidctint-sse2.asm
-index 7aa869b..a428c9e 100644
+index 92f633e..9301e81 100644
 --- a/simd/x86_64/jidctint-sse2.asm
 +++ b/simd/x86_64/jidctint-sse2.asm
 @@ -108,6 +108,7 @@ PB_CENTERJSAMP times 16 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_islow_sse2)
- 
+
  EXTN(jsimd_idct_islow_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
+     mov         rbp, rsp
+     push        r15
 diff --git a/simd/x86_64/jidctred-sse2.asm b/simd/x86_64/jidctred-sse2.asm
-index 4ece9d8..5fac624 100644
+index 1ec500c..e74162d 100644
 --- a/simd/x86_64/jidctred-sse2.asm
 +++ b/simd/x86_64/jidctred-sse2.asm
 @@ -116,6 +116,7 @@ PB_CENTERJSAMP  times 16 db  CENTERJSAMPLE
      GLOBAL_FUNCTION(jsimd_idct_4x4_sse2)
- 
+
  EXTN(jsimd_idct_4x4_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp                     ; rax = original rbp
-     sub         rsp, byte 4
-@@ -414,6 +415,7 @@ EXTN(jsimd_idct_4x4_sse2):
+     mov         rbp, rsp
+     push        r15
+@@ -413,6 +414,7 @@ EXTN(jsimd_idct_4x4_sse2):
      GLOBAL_FUNCTION(jsimd_idct_2x2_sse2)
- 
+
  EXTN(jsimd_idct_2x2_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     collect_args 4
 diff --git a/simd/x86_64/jquantf-sse2.asm b/simd/x86_64/jquantf-sse2.asm
-index ab2e395..a260e32 100644
+index 232bbb2..2d65986 100644
 --- a/simd/x86_64/jquantf-sse2.asm
 +++ b/simd/x86_64/jquantf-sse2.asm
 @@ -37,6 +37,7 @@
      GLOBAL_FUNCTION(jsimd_convsamp_float_sse2)
- 
+
  EXTN(jsimd_convsamp_float_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
-@@ -110,6 +111,7 @@ EXTN(jsimd_convsamp_float_sse2):
+     collect_args 3
+@@ -109,6 +110,7 @@ EXTN(jsimd_convsamp_float_sse2):
      GLOBAL_FUNCTION(jsimd_quantize_float_sse2)
- 
+
  EXTN(jsimd_quantize_float_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     collect_args 3
 diff --git a/simd/x86_64/jquanti-avx2.asm b/simd/x86_64/jquanti-avx2.asm
-index 70fe811..c48c73f 100644
+index 66104d7..a7ea496 100644
 --- a/simd/x86_64/jquanti-avx2.asm
 +++ b/simd/x86_64/jquanti-avx2.asm
 @@ -38,6 +38,7 @@
      GLOBAL_FUNCTION(jsimd_convsamp_avx2)
- 
+
  EXTN(jsimd_convsamp_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
-@@ -116,6 +117,7 @@ EXTN(jsimd_convsamp_avx2):
+     collect_args 3
+@@ -115,6 +116,7 @@ EXTN(jsimd_convsamp_avx2):
      GLOBAL_FUNCTION(jsimd_quantize_avx2)
- 
+
  EXTN(jsimd_quantize_avx2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
+     collect_args 3
 diff --git a/simd/x86_64/jquanti-sse2.asm b/simd/x86_64/jquanti-sse2.asm
-index 3ee4420..fb1b4ae 100644
+index 11e9f4c..24cb1cf 100644
 --- a/simd/x86_64/jquanti-sse2.asm
 +++ b/simd/x86_64/jquanti-sse2.asm
 @@ -37,6 +37,7 @@
      GLOBAL_FUNCTION(jsimd_convsamp_sse2)
- 
+
  EXTN(jsimd_convsamp_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
-@@ -116,6 +117,7 @@ EXTN(jsimd_convsamp_sse2):
+     collect_args 3
+@@ -115,6 +116,7 @@ EXTN(jsimd_convsamp_sse2):
      GLOBAL_FUNCTION(jsimd_quantize_sse2)
- 
+
  EXTN(jsimd_quantize_sse2):
 +    _endbr64
      push        rbp
-     mov         rax, rsp
      mov         rbp, rsp
--- 
-2.31.1
-
+     collect_args 3
+--
diff --git a/libjpeg-turbo-cmake.patch b/libjpeg-turbo-cmake.patch
index 5cc9a3e..ef4c14d 100644
--- a/libjpeg-turbo-cmake.patch
+++ b/libjpeg-turbo-cmake.patch
@@ -1,38 +1,13 @@
 diff --git a/CMakeLists.txt b/CMakeLists.txt
-index cca2966..4875bf3 100644
+index adb0ca4..902c271 100644
 --- a/CMakeLists.txt
 +++ b/CMakeLists.txt
-@@ -1434,8 +1434,6 @@ if(WITH_TURBOJPEG)
-       ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
-       LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
-       RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
--    install(TARGETS tjbench
--      RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
-     if(NOT CMAKE_VERSION VERSION_LESS "3.1" AND MSVC AND
-       CMAKE_C_LINKER_SUPPORTS_PDB)
-       install(FILES "$<TARGET_PDB_FILE:turbojpeg>"
-@@ -1446,15 +1444,6 @@ if(WITH_TURBOJPEG)
-     install(TARGETS turbojpeg-static EXPORT ${CMAKE_PROJECT_NAME}Targets
-       INCLUDES DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
-       ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR})
--    if(NOT ENABLE_SHARED)
--      if(MSVC_IDE OR XCODE)
--        set(DIR "${CMAKE_CURRENT_BINARY_DIR}/\${CMAKE_INSTALL_CONFIG_NAME}")
--      else()
--        set(DIR ${CMAKE_CURRENT_BINARY_DIR})
--      endif()
--      install(PROGRAMS ${DIR}/tjbench-static${EXE}
--        DESTINATION ${CMAKE_INSTALL_BINDIR} RENAME tjbench${EXE})
--    endif()
-   endif()
-   install(FILES ${CMAKE_CURRENT_SOURCE_DIR}/turbojpeg.h
-     DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
-@@ -1481,18 +1470,6 @@ endif()
- 
+@@ -1768,18 +1768,6 @@ endif()
+
  install(TARGETS rdjpgcom wrjpgcom RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
- 
+
 -install(FILES ${CMAKE_CURRENT_SOURCE_DIR}/README.ijg
--  ${CMAKE_CURRENT_SOURCE_DIR}/README.md ${CMAKE_CURRENT_SOURCE_DIR}/example.txt
+-  ${CMAKE_CURRENT_SOURCE_DIR}/README.md ${CMAKE_CURRENT_SOURCE_DIR}/example.c
 -  ${CMAKE_CURRENT_SOURCE_DIR}/tjexample.c
 -  ${CMAKE_CURRENT_SOURCE_DIR}/libjpeg.txt
 -  ${CMAKE_CURRENT_SOURCE_DIR}/structure.txt
@@ -46,12 +21,13 @@ index cca2966..4875bf3 100644
  if(UNIX OR MINGW)
    install(FILES ${CMAKE_CURRENT_SOURCE_DIR}/cjpeg.1
      ${CMAKE_CURRENT_SOURCE_DIR}/djpeg.1 ${CMAKE_CURRENT_SOURCE_DIR}/jpegtran.1
-@@ -1516,7 +1493,7 @@ install(EXPORT ${CMAKE_PROJECT_NAME}Targets
- 
+@@ -1803,7 +1791,7 @@ install(EXPORT ${CMAKE_PROJECT_NAME}Targets
+
  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/jconfig.h
    ${CMAKE_CURRENT_SOURCE_DIR}/jerror.h ${CMAKE_CURRENT_SOURCE_DIR}/jmorecfg.h
 -  ${CMAKE_CURRENT_SOURCE_DIR}/jpeglib.h
 +  ${CMAKE_CURRENT_SOURCE_DIR}/jpeglib.h ${CMAKE_CURRENT_SOURCE_DIR}/jpegint.h
    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
- 
+
  include(cmakescripts/BuildPackages.cmake)
+--
diff --git a/libjpeg-turbo.spec b/libjpeg-turbo.spec
index 83e0198..0b92d7c 100644
--- a/libjpeg-turbo.spec
+++ b/libjpeg-turbo.spec
@@ -1,11 +1,11 @@
 Name:           libjpeg-turbo
-Version:        2.1.4
-Release:        6%{?dist}
+Version:        3.0.2
+Release:        1%{?dist}
 Summary:        A MMX/SSE2/SIMD accelerated library for manipulating JPEG image files
 License:        Zlib AND BSD-3-Clause AND MIT AND IJG
-URL:            http://sourceforge.net/projects/libjpeg-turbo
+URL:            https://github.com/%{name}/%{name}
 
-Source0:        http://downloads.sourceforge.net/%{name}/%{name}-%{version}.tar.gz
+Source0:        %{url}/releases/download/%{version}/%{name}-%{version}.tar.gz
 Patch0:         libjpeg-turbo-cmake.patch
 Patch1:         libjpeg-turbo-CET.patch
 
@@ -92,6 +92,9 @@ export LDFLAGS="$RPM_LD_FLAGS -Wl,-z,ibt -Wl,-z,shstk"
 %cmake_install
 find %{buildroot} -name "*.la" -delete
 
+# Remove tjbench
+rm -f %{buildroot}/%{_bindir}/tjbench
+
 # Fix perms
 chmod -x README.md
 
@@ -146,7 +149,7 @@ export LD_LIBRARY_PATH=%{buildroot}%{_libdir}
 %{_libdir}/libjpeg.so.62*
 
 %files devel
-%doc coderules.txt jconfig.txt libjpeg.txt structure.txt example.txt
+%doc coderules.txt jconfig.txt libjpeg.txt structure.txt
 %{_includedir}/jconfig*.h
 %{_includedir}/jerror.h
 %{_includedir}/jmorecfg.h
@@ -181,6 +184,9 @@ export LD_LIBRARY_PATH=%{buildroot}%{_libdir}
 %{_libdir}/pkgconfig/libturbojpeg.pc
 
 %changelog
+* Mon Feb 05 2024 Frantisek Zatloukal <fzatlouk@redhat.com> - 3.0.2-1
+- New upstream release 3.0.2 (Fixes RHBZ#2256228 and RHBZ#2166459 and RHBZ#2208448)
+
 * Mon Jan 29 2024 Matej Mužila <mmuzila@redhat.com> - 2.1.4-6
 - migrated to SPDX license
 
diff --git a/sources b/sources
index 55ec2e1..2573f54 100644
--- a/sources
+++ b/sources
@@ -1 +1 @@
-SHA512 (libjpeg-turbo-2.1.4.tar.gz) = 511f065767c022da06b6c36299686fa44f83441646f7e33b766c6cfab03f91b0e6bfa456962184071dadaed4057ba9a29cba685383f3eb86a4370a1a53731a70
+SHA512 (libjpeg-turbo-3.0.2.tar.gz) = f5eadda0712feb810a8c3bb2621fda24a4c30574998ce30f423b3ffa25225c7a87cb14b696232bc0270485f422a2853a5c32eafb65bc5eeab1b41d8aeb32ad29