[AArch64] Incorrect load after #142941 #150004
Description
After the merge of #142941 (e8a891b), the following Fortran program prints incorrect result on -O2 or higher.
program main
integer :: i, k
k = 0
write(10, *) 0
write(10, *) 1
write(10, *) 2
write(10, *) 3
write(10, *) 4
write(10, *) 5
write(10, *) 6
write(10, *) 7
write(10, *) 8
write(10, *) 9
rewind 10
read(10, *) i; if (i /= 0) k = k + 1
read(10, *) i; if (i /= 1) k = k + 1
read(10, *) i; if (i /= 2) k = k + 1
read(10, *) i; if (i /= 3) k = k + 1
read(10, *) i; if (i /= 4) k = k + 1
read(10, *) i; if (i /= 5) k = k + 1
read(10, *) i; if (i /= 6) k = k + 1
read(10, *) i; if (i /= 7) k = k + 1
read(10, *) i; if (i /= 8) k = k + 1
read(10, *) i; if (i /= 9) k = k + 1
print *, k
end program main$ flang -O0 test2.f90 && ./a.out
0
$ flang -O1 test2.f90 && ./a.out
0
$ flang -O2 test2.f90 && ./a.out
4
$ flang -O3 test2.f90 && ./a.out
4This program writes ten values to a file and read the written values. If a read value is not the expected one, it increments k.
Before e8a891b, it prints 0 for all optimization levels. It's the expected behavior.
Comparing assembly before and after the commit shows that there are differences in how i values are loaded after read.
@@ -188,12 +188,9 @@
bl _FortranAioInputInteger
mov x0, x20
bl _FortranAioEndIoStatement
- ldr q0, [sp, #16]
mov w0, #10
mov x1, x19
mov w2, #22
- ld1 { v0.s }[1], [x21]
- str q0, [sp, #16]
bl _FortranAioBeginExternalListInput
add x1, x29, #28
mov w2, #4
@@ -201,12 +198,9 @@
bl _FortranAioInputInteger
mov x0, x20
bl _FortranAioEndIoStatement
- ldr q0, [sp, #16]
mov w0, #10
mov x1, x19
mov w2, #23
- ld1 { v0.s }[2], [x21]
- str q0, [sp, #16]
bl _FortranAioBeginExternalListInput
add x1, x29, #28
mov w2, #4
@@ -214,11 +208,14 @@
bl _FortranAioInputInteger
mov x0, x20
bl _FortranAioEndIoStatement
- ldr q0, [sp, #16]
+ ldr q1, [sp, #16]
mov w0, #10
mov x1, x19
mov w2, #24
- ld1 { v0.s }[3], [x21]
+ ld1 { v1.s }[1], [x22]
+ ldr s0, [x22]
+ ld1 { v0.s }[1], [x22]
+ zip1 v0.2d, v1.2d, v0.2d
str q0, [sp, #16]
bl _FortranAioBeginExternalListInput
add x1, x29, #28#149703 does not fix this problem.
@jcohen-apple Could you take a look?
The LLVM IR by flang -O2 -S -emit-llvm test.f90, assembly before e8a891b, assembly after e8a891b are attached.
I ported this program to C using fprintf/fscanf but I could not reproduce the problem.