https://bugs.llvm.org/show_bug.cgi?id=40243
Bug ID: 40243
Summary: [x86] horizontal op codegen miscompile
Product: libraries
Version: trunk
Hardware: PC
OS: All
Status: NEW
Severity: enhancement
Priority: P
Component: Backend: X86
Assignee: [email protected]
Reporter: [email protected]
CC: [email protected], [email protected],
[email protected], [email protected]
#include <immintrin.h>
#include <stdio.h>
#include <stdalign.h> // C11 defines _Alignas(). This header defines
alignas()
void print(__m256 in) {
alignas(32) float v[8];
_mm256_store_ps(v, in);
printf("element 4: %.1f\n", v[4]);
printf("element 7: %.1f\n", v[7]);
}
__m256 hadd(__m256 a, __m256 b) {
float r4 = a[4] + a[5];
float r7 = b[6] + b[7];
__m256 r = (__m256){ a[0], a[2], b[0], b[2], r4, a[6], b[4], r7 };
return __builtin_shufflevector(r, a, -1, -1, -1, -1, 4, -1, -1, 7);
}
int main() {
__m256 a = _mm256_set_ps(7.0, 6.0, 5.0, 4.0, 3.0, 2.0, 1.0, 0.0);
__m256 b = _mm256_set_ps(15.0, 14.0, 13.0, 12.0, 11.0, 10.0, 9.0, 8.0);
print(hadd(a, b));
return 0;
}
-----------------------------------------------------------------------------
$ clang checkhadd.c -mavx -O0 ; ./a.out
element 4: 9.0
element 7: 29.0
$ clang checkhadd.c -mavx -O1 ; ./a.out
element 4: 9.0
element 7: 13.0 <--- 14.0 + 15.0 ???
------------------------------------------------------------------------------
I think this is a long-standing bug in LowerToHorizontalOp() -- it's visible on
the shipping macOS version of clang.
There's probably a small fix that could correct this, but given that we're also
missing horizontal patterns completely and only partially matching/optimizing
other patterns, I wonder if we should start over on that function.
--
You are receiving this mail because:
You are on the CC list for the bug._______________________________________________
llvm-bugs mailing list
[email protected]
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
