https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92637
Bug ID: 92637
Summary: runtime issue with -ftree-coalesce-vars
Product: gcc
Version: 7.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: dongjianqiang2 at huawei dot com
Target Milestone: ---
Created attachment 47338
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=47338&action=edit
testsuite
hi, I compiled the attached with aarch64-linux-gnu-gcc -c -O2 -march=armv8.1-a
testsuite.c -o testsuite.o, it had runtime error, and I found x10 was
overwriten ar row 214.
00000000000001d0 <find_next_zero_area>:
1d0: a9bf7bfd stp x29, x30, [sp, #-16]!
1d4: aa0403eb mov x11, x4
1d8: aa0003ea mov x10, x0
1dc: aa0103e9 mov x9, x1
1e0: 910003fd mov x29, sp
1e4: aa2403ed mvn x13, x4
1e8: 2a0303ec mov w12, w3
1ec: 14000005 b 200 <find_next_zero_area+0x30>
1f0: 94000000 bl f0 <find_next_bit>
1f4: eb00011f cmp x8, x0
1f8: 54000209 b.ls 238 <find_next_zero_area+0x68> // b.plast
1fc: 91000402 add x2, x0, #0x1
200: aa0903e1 mov x1, x9
204: aa0a03e0 mov x0, x10
208: 94000000 bl 0 <find_next_zero_bit>
20c: 8b0b0007 add x7, x0, x11
210: 8a0d00e7 and x7, x7, x13
214: aa0a03e0 mov x0, x10 // HERE x10 is overwriten
by glibc
218: 8b0c00e8 add x8, x7, x12
21c: aa0703e2 mov x2, x7
220: aa0803e1 mov x1, x8
224: eb08013f cmp x9, x8
228: 54fffe42 b.cs 1f0 <find_next_zero_area+0x20> // b.hs,
b.nlast
22c: aa0803e0 mov x0, x8
230: a8c17bfd ldp x29, x30, [sp], #16
234: d65f03c0 ret
238: aa0703e8 mov x8, x7
23c: aa0803e0 mov x0, x8
240: a8c17bfd ldp x29, x30, [sp], #16
244: d65f03c0 ret
But when I add -fno-tree-coalesce-vars, x10 changed to x9, following assembler
is running correctly.
00000000000001e0 <find_next_zero_area>:
1e0: a9bf7bfd stp x29, x30, [sp, #-16]!
1e4: aa0403ea mov x10, x4
1e8: aa0003e9 mov x9, x0
1ec: aa0103e8 mov x8, x1
1f0: 910003fd mov x29, sp
1f4: aa2403ec mvn x12, x4
1f8: 2a0303eb mov w11, w3
1fc: 14000005 b 210 <find_next_zero_area+0x30>
200: 94000000 bl f8 <find_next_bit>
204: eb0000ff cmp x7, x0
208: 54000209 b.ls 248 <find_next_zero_area+0x68> // b.plast
20c: 91000402 add x2, x0, #0x1
210: aa0803e1 mov x1, x8
214: aa0903e0 mov x0, x9 //HERE is correct
218: 94000000 bl 0 <find_next_zero_bit>
21c: 8b0a0006 add x6, x0, x10
220: 8a0c00c6 and x6, x6, x12
224: aa0903e0 mov x0, x9
228: 8b0b00c7 add x7, x6, x11
22c: aa0603e2 mov x2, x6
230: aa0703e1 mov x1, x7
234: eb07011f cmp x8, x7
238: 54fffe42 b.cs 200 <find_next_zero_area+0x20> // b.hs,
b.nlast
23c: aa0703e0 mov x0, x7
240: a8c17bfd ldp x29, x30, [sp], #16
244: d65f03c0 ret
248: aa0603e0 mov x0, x6
24c: a8c17bfd ldp x29, x30, [sp], #16
250: d65f03c0 ret
I checked x9-x15 are caller-saved register, why compiler doesn't push these
registers onto the stack across calls?
