condition jump instruction seems lost when optimization opened with -O1 or -O2 or -Oz

[haifengdeng]

when i use below command to compile FFmpeg 4.3.1 libavformat/utils.c with Clang 14 on Macmini M1 macOS 14.3.1:

/opt/homebrew/opt/llvm@14/bin/clang -I. -I./ --sysroot=/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS17.2.sdk -D_ISOC99_SOURCE -D_LARGEFILE_SOURCE -I./compat/dispatch_semaphore -DPIC -DZLIB_CONST -DHAVE_AV_CONFIG_H -DBUILDING_avformat -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS17.2.sdk -arch arm64 -mios-version-min=11.0 -no-canonical-prefixes -Wno-invalid-command-line-argument -Wno-unused-command-line-argument -funwind-tables -fstack-protector-strong -fno-addrsig -Werror=return-type -Werror=int-to-pointer-cast -Werror=implicit-function-declaration -fPIC -DTARGET_OS_IPHONE=1 -Wall -pipe -std=c99 -ffast-math -g -O0 -DDEBUG_ENABLE -I/Users/yue/gitlab/libaveditor_6/buildbin_ios/tmp/arm64/include -std=c11 -Werror=partial-availability -fomit-frame-pointer -fPIC -pthread  -DX264_API_IMPORTS -I/opt/homebrew/Cellar/x264/r3108/include -ggdb -Wdeclaration-after-statement -Wall -Wdisabled-optimization -Wpointer-arith -Wredundant-decls -Wwrite-strings -Wtype-limits -Wundef -Wmissing-prototypes -Wno-pointer-to-int-cast -Wstrict-prototypes -Wempty-body -Wno-parentheses -Wno-switch -Wno-format-zero-length -Wno-pointer-sign -Wno-unused-const-variable -Wno-bool-operation -Wno-char-subscripts -Oz -fno-math-errno -fno-signed-zeros -mstack-alignment=16 -Qunused-arguments -Werror=implicit-function-declaration -Werror=missing-prototypes -Werror=return-type  -MMD -MF libavformat/utils.d -MT libavformat/utils.o -c -o libavformat/utils.o libavformat/utils.c

then I disassembe like below:

/Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/objdump -S libavformat/utils.o > utils14.txt

I found code if (st->index_entries) { in function ff_seek_frame_binary compiled like below, when I run to here, it always crashes:

;     av_log(s, AV_LOG_TRACE, "read_seek: %d %s\n", stream_index, av_ts2str(target_ts));
    2b00: 4f00e400     	movi.16b	v0, #0
    2b04: ad0403e0     	stp	q0, q0, [sp, #128]
    2b08: 910203f4     	add	x20, sp, #128
    2b0c: 910203e0     	add	x0, sp, #128
    2b10: aa0203e1     	mov	x1, x2
    2b14: 94000000     	bl	0x2b14 <_ff_seek_frame_binary+0x5c>
    2b18: a90053f7     	stp	x23, x20, [sp]
    2b1c: 90000002     	adrp	x2, 0x2000 <_ff_seek_frame_binary+0x64>
    2b20: 91000042     	add	x2, x2, #0
    2b24: 94000000     	bl	0x2b24 <_ff_seek_frame_binary+0x6c>
;     st = s->streams[stream_index];
    2b28: f9401a68     	ldr	x8, [x19, #48]
    2b2c: f8775914     	ldr	x20, [x8, w23, uxtw  #3]
**;     if (st->index_entries) {
    2b30: f940e69a     	ldr	x26, [x20, #456]**
;                                           flags | AVSEEK_FLAG_BACKWARD);
    2b34: 320002c2     	orr	w2, w22, #0x1
    2b38: 94000000     	bl	0x2b38 <_ff_seek_frame_binary+0x80>
;         index = FFMAX(index, 0);
    2b3c: 0aa07c08     	bic	w8, w0, w0, asr #31
;         if (e->timestamp <= target_ts || e->pos == e->min_distance) {
    2b40: 52800309     	mov	w9, #24
    2b44: 9ba96909     	umaddl	x9, w8, w9, x26
    2b48: a9406139     	ldp	x25, x24, [x9]
    2b4c: eb15031f     	cmp	x24, x21
    2b50: 540000cd     	b.le	0x2b68 <_ff_seek_frame_binary+0xb0>
    2b54: 52800309     	mov	w9, #24
;         if (e->timestamp <= target_ts || e->pos == e->min_distance) {

but when I use Clang 13 with same the compile command, it generates the code below:

;     av_log(s, AV_LOG_TRACE, "read_seek: %d %s\n", stream_index, av_ts2str(target_ts));
    2b14: 4f00e400     	movi.16b	v0, #0
    2b18: ad0383e0     	stp	q0, q0, [sp, #112]
    2b1c: 9101c3f4     	add	x20, sp, #112
    2b20: 9101c3e0     	add	x0, sp, #112
    2b24: aa0203e1     	mov	x1, x2
    2b28: 94000000     	bl	0x2b28 <_ff_seek_frame_binary+0x58>
    2b2c: a90053f7     	stp	x23, x20, [sp]
    2b30: 90000002     	adrp	x2, 0x2000 <_ff_seek_frame_binary+0x60>
    2b34: 91000042     	add	x2, x2, #0
    2b38: 94000000     	bl	0x2b38 <_ff_seek_frame_binary+0x68>
;     st = s->streams[stream_index];
    2b3c: f9401a68     	ldr	x8, [x19, #48]
    2b40: f8775914     	ldr	x20, [x8, w23, uxtw  #3]
**;     if (st->index_entries) {
    2b44: f940e69a     	ldr	x26, [x20, #456]
    2b48: b400033a     	cbz	x26, 0x2bac <_ff_seek_frame_binary+0xdc>**
;                                           flags | AVSEEK_FLAG_BACKWARD);
    2b4c: 320002c2     	orr	w2, w22, #0x1
    2b50: 94000000     	bl	0x2b50 <_ff_seek_frame_binary+0x80>
;         index = FFMAX(index, 0);
    2b54: 0aa07c08     	bic	w8, w0, w0, asr #31
;         if (e->timestamp <= target_ts || e->pos == e->min_distance) {

someone can help me find if my compile command has an error somewhere?

utils.zip

[dmitrys99]

I found similar issue https://gitlab.com/embeddable-common-lisp/ecl/-/issues/737 and can confirm problem exists.

cl_object si_make_seq_iterator_va(cl_narg narg, ...)
{
 cl_object x[2];
 va_list args; va_start(args,narg);
 x[0] = va_arg(args,cl_object);
 for (int i = 1; i < narg; i++){
  x[i] = va_arg(args,cl_object);
 }
 va_end(args);
 return si_make_seq_iterator(narg, x[0], x[1]);
}

When -O2 or -O3 flag given, loop condition silently skipped and loop body is executed yet it shouldn't.

Without optimization things work correctly.

Broken function

si_make_seq_iterator_va:
	sub    sp, sp, #0x40            
	stp    x29, x30, [sp, #0x30]    
	add    x29, sp, #0x30           
	adrp   x8, 642                  
	ldr    x8, [x8, #0x58]          
	ldr    x8, [x8]                 
	stur   x8, [x29, #-0x8]         
	add    x8, x29, #0x10           
	orr    x8, x8, #0x8             
	str    x8, [sp, #0x10]          
	ldr    x1, [x29, #0x10]         
	mov    w8, #0x1                 
	add    x9, sp, #0x18            
	ldr    x10, [sp, #0x10]         
	add    x11, x10, #0x8           
	str    x11, [sp, #0x10]         
	ldr    x10, [x10]               
	str    x10, [x9, x8, lsl #3]    
	add    x8, x8, #0x1             
	cmp    x0, x8                   
	b.ne   0x1ef24                    ; <+52> at seq.c:480:10
	ldr    x8, [sp, #0x20]          
	str    x8, [sp]                 
	bl     0x1ed80                    ; si_make_seq_iterator at seq.c:404
	ldur   x8, [x29, #-0x8]         
	adrp   x9, 642                  
	ldr    x9, [x9, #0x58]          
	ldr    x9, [x9]                 
	cmp    x9, x8                   
	b.ne   0x1ef74                    ; <+132> at seq.c
	ldp    x29, x30, [sp, #0x30]    
	add    sp, sp, #0x40            
	ret                             
	bl     0x1947e4                   ; symbol stub for: __stack_chk_fail

Correct function

si_make_seq_iterator_va:
	sub    sp, sp, #0x60            
	stp    x29, x30, [sp, #0x50]    
	add    x29, sp, #0x50           
	adrp   x8, 985                  
	ldr    x8, [x8, #0x60]          
	ldr    x8, [x8]                 
	stur   x8, [x29, #-0x8]         
	stur   x0, [x29, #-0x20]        
	add    x9, sp, #0x28            
	add    x8, x29, #0x10           
	str    x8, [x9]                 
	ldr    x9, [sp, #0x28]          
	ldr    x8, [x9]                 
	add    x9, x9, #0x8             
	str    x9, [sp, #0x28]          
	str    x8, [sp, #0x20]          
	ldr    x8, [sp, #0x20]          
	stur   x8, [x29, #-0x18]        
	mov    w8, #0x1                 
	str    w8, [sp, #0x1c]          
	b      0x3f61c                    ; <+84> at seq.c:479:18
	ldrsw  x8, [sp, #0x1c]          
	ldur   x9, [x29, #-0x20]        
	subs   x8, x8, x9               
	cset   w8, ge                   
	tbnz   w8, #0x0, 0x3f66c          ; <+164> at seq.c:483:30
	b      0x3f634                    ; <+108> at seq.c:480:10
	ldr    x9, [sp, #0x28]          
	ldr    x8, [x9]                 
	add    x9, x9, #0x8             
	str    x9, [sp, #0x28]          
	str    x8, [sp, #0x10]          
	ldr    x8, [sp, #0x10]          
	ldrsw  x10, [sp, #0x1c]         
	sub    x9, x29, #0x18           
	str    x8, [x9, x10, lsl #3]    
	b      0x3f65c                    ; <+148> at seq.c:479:29
	ldr    w8, [sp, #0x1c]          
	add    w8, w8, #0x1             
	str    w8, [sp, #0x1c]          
	b      0x3f61c                    ; <+84> at seq.c:479:18
	ldur   x0, [x29, #-0x20]        
	ldur   x1, [x29, #-0x18]        
	ldur   x8, [x29, #-0x10]        
	mov    x9, sp                   
	str    x8, [x9]                 
	bl     0x3f228                    ; si_make_seq_iterator at seq.c:404
	str    x0, [sp, #0x8]           
	ldur   x9, [x29, #-0x8]         
	adrp   x8, 985                  
	ldr    x8, [x8, #0x60]          
	ldr    x8, [x8]                 
	subs   x8, x8, x9               
	cset   w8, eq                   
	tbnz   w8, #0x0, 0x3f6ac          ; <+228> at seq.c
	b      0x3f6a8                    ; <+224> at seq.c:483:2
	bl     0x30e424                   ; symbol stub for: __stack_chk_fail
	ldr    x0, [sp, #0x8]           
	ldp    x29, x30, [sp, #0x50]    
	add    sp, sp, #0x60            
	ret       

Checked with latest git version of LLVM/Clang.