Associate II

Question

Bug with GCC compiler (for STM32F030C6)

Forum|Forum|6 years ago
October 11, 2019
3 replies
1846 views

We encountered what appears to be a compiler bug as rewriting the code slightly by adding intermediate variables solved the issue.

Relevant code :

#define HT16K33_DISPLAY_NB_COLUMNS (8)
#define HT16K33_DISPLAY_NB_ROWS (16)
#define HT16K33_DISPLAY_NB_COLUMN_BYTES (HT16K33_DISPLAY_NB_ROWS/8)
 
class HT16K33_SevenSegment {
 void writeASCII(const uint8_t ledPos, const char character);
 
 private:
 uint8_t screen[HT16K33_DISPLAY_NB_COLUMNS*HT16K33_DISPLAY_NB_COLUMN_BYTES]; 
}
 
const char AlphaNum[] = { /* ... list of bytes ... */ }; 
 
void
HT16K33_SevenSegment::writeASCII(const uint8_t ledPos, const char character) {
	assert(ledPos<3);
	assert(character<100);
#if 1
	screen[ledPos*HT16K33_DISPLAY_NB_COLUMN_BYTES] = AlphaNum[character-'0'-7];
#else
	{
		int charIdx=character-'0'-7;
		char charVal=AlphaNum[charIdx];
		screen[ledPos*HT16K33_DISPLAY_NB_COLUMN_BYTES]=charVal;
	}
#endif
}

In the above extract, the code in the first leg of the preprocessor "if" does not work - "screen" does not get the correct data value. The evaluation of the right hand expression by the debugger shows "0x39" or 57, but the value that is actually written is 0x40 which is not in the AlphaNum table.

With the code in the "else" part, we do get the correct value.

I think that the cause can be detemined by examining the generated code:

Dysfunctionnal code with assembly:

/**
 * Write a 1 ASCII Character from position 0.
 */
void
HT16K33_SevenSegment::writeASCII(const uint8_t ledPos, const char character) {
 8003470:	b580 	push	{r7, lr}
 8003472:	b082 	sub	sp, #8
 8003474:	af00 	add	r7, sp, #0
 8003476:	6078 	str	r0, [r7, #4]
 8003478:	0008 	movs	r0, r1
 800347a:	0011 	movs	r1, r2
 800347c:	1cfb 	adds	r3, r7, #3
 800347e:	1c02 	adds	r2, r0, #0
 8003480:	701a 	strb	r2, [r3, #0]
 8003482:	1cbb 	adds	r3, r7, #2
 8003484:	1c0a 	adds	r2, r1, #0
 8003486:	701a 	strb	r2, [r3, #0]
	assert(ledPos<3);
 8003488:	1cfb 	adds	r3, r7, #3
 800348a:	781b 	ldrb	r3, [r3, #0]
 800348c:	2b02 	cmp	r3, #2
 800348e:	d905 	bls.n	800349c <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x2c>
 8003490:	4b10 	ldr	r3, [pc, #64]	; (80034d4 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x64>)
 8003492:	4a11 	ldr	r2, [pc, #68]	; (80034d8 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x68>)
 8003494:	4811 	ldr	r0, [pc, #68]	; (80034dc <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x6c>)
 8003496:	21e0 	movs	r1, #224	; 0xe0
 8003498:	f001 f9fa 	bl	8004890 <__assert_func>
	assert(character<100);
 800349c:	1cbb 	adds	r3, r7, #2
 800349e:	781b 	ldrb	r3, [r3, #0]
 80034a0:	2b63 	cmp	r3, #99	; 0x63
 80034a2:	d905 	bls.n	80034b0 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x40>
 80034a4:	4b0e 	ldr	r3, [pc, #56]	; (80034e0 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x70>)
 80034a6:	4a0c 	ldr	r2, [pc, #48]	; (80034d8 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x68>)
 80034a8:	480c 	ldr	r0, [pc, #48]	; (80034dc <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x6c>)
 80034aa:	21e1 	movs	r1, #225	; 0xe1
 80034ac:	f001 f9f0 	bl	8004890 <__assert_func>
#if 1
	screen[ledPos*HT16K33_DISPLAY_NB_COLUMN_BYTES] = AlphaNum[character-'0'-7];
 80034b0:	1cbb 	adds	r3, r7, #2
 80034b2:	781b 	ldrb	r3, [r3, #0]
 80034b4:	3b37 	subs	r3, #55	; 0x37
 80034b6:	001a 	movs	r2, r3
 80034b8:	1cfb 	adds	r3, r7, #3
 80034ba:	781b 	ldrb	r3, [r3, #0]
 80034bc:	005b 	lsls	r3, r3, #1
 80034be:	4909 	ldr	r1, [pc, #36]	; (80034e4 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x74>)
 80034c0:	5c89 	ldrb	r1, [r1, r2]
 80034c2:	687a 	ldr	r2, [r7, #4]
 80034c4:	18d3 	adds	r3, r2, r3
 80034c6:	1c0a 	adds	r2, r1, #0
 80034c8:	72da 	strb	r2, [r3, #11]
		int charIdx=character-'0'-7;
		char charVal=AlphaNum[charIdx];
		screen[ledPos*HT16K33_DISPLAY_NB_COLUMN_BYTES]=charVal;
	}
#endif
}

Working code with assembly:

/**
 * Write a 1 ASCII Character from position 0.
 */
void
HT16K33_SevenSegment::writeASCII(const uint8_t ledPos, const char character) {
 8003470:	b580 	push	{r7, lr}
 8003472:	b084 	sub	sp, #16
 8003474:	af00 	add	r7, sp, #0
 8003476:	6078 	str	r0, [r7, #4]
 8003478:	0008 	movs	r0, r1
 800347a:	0011 	movs	r1, r2
 800347c:	1cfb 	adds	r3, r7, #3
 800347e:	1c02 	adds	r2, r0, #0
 8003480:	701a 	strb	r2, [r3, #0]
 8003482:	1cbb 	adds	r3, r7, #2
 8003484:	1c0a 	adds	r2, r1, #0
 8003486:	701a 	strb	r2, [r3, #0]
	assert(ledPos<3);
 8003488:	1cfb 	adds	r3, r7, #3
 800348a:	781b 	ldrb	r3, [r3, #0]
 800348c:	2b02 	cmp	r3, #2
 800348e:	d905 	bls.n	800349c <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x2c>
 8003490:	4b13 	ldr	r3, [pc, #76]	; (80034e0 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x70>)
 8003492:	4a14 	ldr	r2, [pc, #80]	; (80034e4 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x74>)
 8003494:	4814 	ldr	r0, [pc, #80]	; (80034e8 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x78>)
 8003496:	21e0 	movs	r1, #224	; 0xe0
 8003498:	f001 fa00 	bl	800489c <__assert_func>
	assert(character<100);
 800349c:	1cbb 	adds	r3, r7, #2
 800349e:	781b 	ldrb	r3, [r3, #0]
 80034a0:	2b63 	cmp	r3, #99	; 0x63
 80034a2:	d905 	bls.n	80034b0 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x40>
 80034a4:	4b11 	ldr	r3, [pc, #68]	; (80034ec <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x7c>)
 80034a6:	4a0f 	ldr	r2, [pc, #60]	; (80034e4 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x74>)
 80034a8:	480f 	ldr	r0, [pc, #60]	; (80034e8 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x78>)
 80034aa:	21e1 	movs	r1, #225	; 0xe1
 80034ac:	f001 f9f6 	bl	800489c <__assert_func>
 
 
 
#if 0
	screen[ledPos*HT16K33_DISPLAY_NB_COLUMN_BYTES] = AlphaNum[character-'0'-7];
#else
 {
	int charIdx=character-'0'-7;
 80034b0:	1cbb 	adds	r3, r7, #2
 80034b2:	781b 	ldrb	r3, [r3, #0]
 80034b4:	3b37 	subs	r3, #55	; 0x37
 80034b6:	60fb 	str	r3, [r7, #12]
		char charVal=AlphaNum[charIdx];
 80034b8:	200b 	movs	r0, #11
 80034ba:	183b 	adds	r3, r7, r0
 80034bc:	490c 	ldr	r1, [pc, #48]	; (80034f0 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x80>)
 80034be:	68fa 	ldr	r2, [r7, #12]
 80034c0:	188a 	adds	r2, r1, r2
 80034c2:	7812 	ldrb	r2, [r2, #0]
 80034c4:	701a 	strb	r2, [r3, #0]
		screen[ledPos*HT16K33_DISPLAY_NB_COLUMN_BYTES]=charVal;
 80034c6:	1cfb 	adds	r3, r7, #3
 80034c8:	781b 	ldrb	r3, [r3, #0]
 80034ca:	005b 	lsls	r3, r3, #1
 80034cc:	687a 	ldr	r2, [r7, #4]
 80034ce:	18d3 	adds	r3, r2, r3
 80034d0:	183a 	adds	r2, r7, r0
 80034d2:	7812 	ldrb	r2, [r2, #0]
 80034d4:	72da 	strb	r2, [r3, #11]
 }
#endif
}

Where should we report this?

We're using STMCubeIDE V1.0.2 and "com.st.stm32cube.ide.mcu.externaltools.gnu-tools-for-stm32.7-2018-q2-update.win32_1.0.0.201904181610" for the compiler which comes with the IDE.

This topic has been closed for replies.

TDK

Super User

It's clearly trying to do something that looks similar in both cases. I'm not going to try to step through it exactly. Not convinced it's a bug, but could be.

You should add assert(character >= '0' + 7) to verify you're in bounds. That could be one explanation.

Bug reports for the compiler probably (?) should go here:

https://bugs.launchpad.net/gcc-arm-embedded

"If you feel a post has answered your question, please click ""Accept as Solution""."

M_3Author

Associate II

Thank you for your reply.

In my humble experience, it's very likely a bug. I have detected more than 10 confirmed compiler bugs in my carreer ;-).

While debugging this, code execution was deterministic, the debugger showed the expected value for AlphaNum[character-'0'-7], and the index value character-'0'-'7' was 11 for the closely inspected execution, well within the length of the array. This is part of a 3 Digit 7 segment LED driver and all three digits were wrong with the original code (we were trying to show "AbC" and the display showed "rubbish"), but correct with the alternative code (the display was showing "AbC").

I'll probably post the issue on launchpad.

TDK

Super User

Oh well then. With more than 10 confirmed bugs I would imagine you would already know where to post them? Excuse me while I roll my eyes.

"If you feel a post has answered your question, please click ""Accept as Solution""."

M_3Author

Associate II

I've analysed the code in more detail - the assembly code seems to do the same thing, I'll have to check the byte/halfword/word consistency.

void
HT16K33_SevenSegment::writeASCII(const uint8_t ledPos, const char character) {
 8003470:	b580 	push	{r7, lr} // Store current stack pointer
 8003472:	b082 	sub	sp, #8 // Reserve stack
 8003474:	af00 	add	r7, sp, #0 // Get new stack pointer
 8003476:	6078 	str	r0, [r7, #4] // Store 'this' on stack Location #4
 8003478:	0008 	movs	r0, r1 // r0(ledpos)=r1(ledpos)
 800347a:	0011 	movs	r1, r2 // r1(character)=r0(character)
 800347c:	1cfb 	adds	r3, r7, #3 // r3=Address for Location #3
 800347e:	1c02 	adds	r2, r0, #0 // r2(ledpos)=r0(ledpos)
 8003480:	701a 	strb	r2, [r3, #0] // Store r2(ledpos) in Location #3
 8003482:	1cbb 	adds	r3, r7, #2 // r3=address for Location #2
 8003484:	1c0a 	adds	r2, r1, #0 // r2(character)=r1(character)
 8003486:	701a 	strb	r2, [r3, #0] // Store r2(character) in Location #2
 
 // At this point:
 // r7#2: character
 // r7#3: ledpos
 // r7#4: this
 
	screen[ledPos*HT16K33_DISPLAY_NB_COLUMN_BYTES] = AlphaNum[character-'0'-7];
 80034b0:	1cbb 	adds	r3, r7, #2 // Address for r7#2 - character.
 80034b2:	781b 	ldrb	r3, [r3, #0] // r3=character (* r7#2)
 80034b4:	3b37 	subs	r3, #55	; 0x37 // r3-= '0'+7;
 80034b6:	001a 	movs	r2, r3 // r2=r3=character-'0'+7;
 80034b8:	1cfb 	adds	r3, r7, #3 // Address for r7#3 - ledpos
 80034ba:	781b 	ldrb	r3, [r3, #0] // r3=ledpos (* r7#2)
 80034bc:	005b 	lsls	r3, r3, #1 // r3*=2 (r3=2*ledpos=ledPos*HT16K33_DISPLAY_NB_COLUMN_BYTES)
 80034be:	4909 	ldr	r1, [pc, #36]	; (80034e4 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x74>) // r1=Pointer to AlphaNum
 80034c0:	5c89 	ldrb	r1, [r1, r2] // r1=AlphaNum[r2]
 80034c2:	687a 	ldr	r2, [r7, #4] // r2=this
 80034c4:	18d3 	adds	r3, r2, r3 // r3=this +2*ledpos
 80034c6:	1c0a 	adds	r2, r1, #0 // r2=r1=AlphaNum[character-'0'+7]
 80034c8:	72da 	strb	r2, [r3, #11] // store r2 in screen[2*ledpos]

Working code:

/**
 * Write a 1 ASCII Character from position 0.
 */
void
HT16K33_SevenSegment::writeASCII(const uint8_t ledPos, const char character) {
 8003470:	b580 	push	{r7, lr} // Store current stack pointer
 8003472:	b084 	sub	sp, #16 // Reserve stack
 8003474:	af00 	add	r7, sp, #0 // Get new stack pointer
 8003476:	6078 	str	r0, [r7, #4] // Store 'this' on stack Location #4
 8003478:	0008 	movs	r0, r1 // r0(ledpos)=r1(ledpos)
 800347a:	0011 	movs	r1, r2 // r1(character)=r0(character)
 800347c:	1cfb 	adds	r3, r7, #3 // r3=Address for Location #3
 800347e:	1c02 	adds	r2, r0, #0 // r2(ledpos)=r0(ledpos)
 8003480:	701a 	strb	r2, [r3, #0] // Store r2(ledpos) in Location #3
 8003482:	1cbb 	adds	r3, r7, #2 // r3=address for Location #2
 8003484:	1c0a 	adds	r2, r1, #0 // r2(character)=r1(character)
 8003486:	701a 	strb	r2, [r3, #0] // Store r2(character) in Location #2
 
 
 
 {
	int charIdx=character-'0'-7;
 80034b0:	1cbb 	adds	r3, r7, #2 // Address for r7#2 - character.
 80034b2:	781b 	ldrb	r3, [r3, #0] // r3=character (* r7#2)
 80034b4:	3b37 	subs	r3, #55	; 0x37 // r3-= '0'+7;
 80034b6:	60fb 	str	r3, [r7, #12] // Store r3 in "charIdx" #12 on stack
		char charVal=AlphaNum[charIdx];
 80034b8:	200b 	movs	r0, #11 // r0=11
 80034ba:	183b 	adds	r3, r7, r0 // r3=r7#11
 80034bc:	490c 	ldr	r1, [pc, #48]	; (80034f0 <_ZN20HT16K33_SevenSegment10writeASCIIEhc+0x80>) // r1=ptr to AlphaNum
 80034be:	68fa 	ldr	r2, [r7, #12] // r2 = *r7"12 = charIdx
 80034c0:	188a 	adds	r2, r1, r2 // r2=ptr to AlphaNum[charIdx]
 80034c2:	7812 	ldrb	r2, [r2, #0] // r2=AlphaNum[charIdx]
 80034c4:	701a 	strb	r2, [r3, #0] // *r7#11(charVal)=r2=AlphaNum[charIdx]
		screen[ledPos*HT16K33_DISPLAY_NB_COLUMN_BYTES]=charVal;
 80034c6:	1cfb 	adds	r3, r7, #3 // r3=ptr to ledpos (r7#3)
 80034c8:	781b 	ldrb	r3, [r3, #0] // r3=ledpos (*r7#3)
 80034ca:	005b 	lsls	r3, r3, #1 // r3*=2 (2*ledpos)
 80034cc:	687a 	ldr	r2, [r7, #4] // r2=ptr to this.
 80034ce:	18d3 	adds	r3, r2, r3 // r2+=2*ledpos
 80034d0:	183a 	adds	r2, r7, r0 // r2=r7+r0==r7#11=charVal
 80034d2:	7812 	ldrb	r2, [r2, #0] // r2=charVal
 80034d4:	72da 	strb	r2, [r3, #11] // store r2 in screen[2*ledpos]
 }

waclawek.jan

Super User

Why don't you single-step through the disasm? You'd probably see immediately where the strange results come from.

M_3Author

Associate II

We've moved on in the mean time, so still have to see if we can reproduce it next Monday.

Before being able to step throught ha ASM efficiently, the code had to be annotated.

Sign up

Login with SSO

Login to the community

Login with SSO

Scanning file for viruses.

This file cannot be downloaded