Ticket #446 (Fixed)Tue Jan 16 18:05:38 UTC 2018

cc 5.75: compiler fails to mark R12 as containing pertinant data and reuses R12 for different subexpression results

Details by André Timmermans (100):

Here is a minimal case:
<pre>
<code>
typedef struct CRect
{
int x0;
int y0;
int x1;
int y1;
} CRect;

typedef struct CSize
{
int cx;
int cy;
} CSize;

typedef struct WListCore WListCore;

struct WListCore
{
CSize m_Size;
CSize m_Offset;

unsigned int m_flags;
};

#define EWList_RightToLeft 0×00000040
#define EWList_BottomToTop 0×00000080

CRect WListCore_RowColsToPixels(const WListCore* This, const CRect* pbox)
{
CRect box;

if (This->m_flags & EWList_RightToLeft)
{
box.×0 = This>m_Offset.cx – pbox->x1 * This->m_Size.cx;
box.×1 = This>m_Offset.cx – pbox->x0 * This->m_Size.cx;
}
else
{
box.×0 = This->m_Offset.cx + pbox->x0 * This->m_Size.cx;
box.×1 = This->m_Offset.cx + pbox->x1 * This->m_Size.cx;
}

if (This->m_flags & EWList_BottomToTop)
{
box.y0 = This->m_Offset.cy + pbox->y0 * This->m_Size.cy;
box.y1 = This->m_Offset.cy + pbox->y1 * This->m_Size.cy;
}
else
{
box.y0 = This>m_Offset.cy – pbox->y1 * This->m_Size.cy;
box.y1 = This>m_Offset.cy pbox>y0 * This->m_Size.cy;
}

return box;
}
</code>
</pre>

Just drop it on !cc and compile it.

The generated output which I somewhat commented is as follow:
<pre>
<code>
STMDB R13!,{R4-R6,R14}
LDR R5,[R1,#16] ; R5 = This->m_flags
LDR R12,[R1,#8] ; R12 = This->m_Offset.cx
LDR R6,[R2,#8] ; R6 = pbox->x1
LDR R3,[R1,#0] ; R3 = This->m_Size.cx
TST R5,#&40 ; if (This->m_flags & EWList_RightToLeft)
RSBNE R12,R12,#0 ; R12 = This>m_Offset.cx
MULNE R12,R3,R6 ; R12 = pbox->x1 * This->m_Size.cx !!! R12 !!!
LDR R14,[R2,#0] ; R14 = pbox->x0
SUBNE R4,R12,R12 ; R4 = 0 !!!
MULNE R12,R3,R14 ; R12 = pbox->x0 * This->m_Size.cx !!! R12 !!!
MLAEQ R4,R3,R14,R12
SUBNE R12,R12,R12
LDR R14,[R2,#12]
MLAEQ R12,R3,R6,R12
TST R5,#&80
LDR R5,[R1,#4]
LDR R1,[R1,#12]
MOV R3,R12
MULEQ R12,R5,R14
LDR R2,[R2,#4]
RSBEQ R1,R1,#0
SUBEQ R12,R1,R12
MLANE R12,R5,R2,R1
MLANE R1,R5,R14,R1
MULEQ R12,R5,R2
STMIA R0!,{R4,R12}
SUBEQ R1,R1,R12
STR R3,[R0],#4
STR R1,[R0,#0]
LDMIA R13!,{R4-R6,PC}
</code>
</pre>

As you can see at several occasions it stores results into R12 when it should use an extra register.