Bug 16087 – Alignment (.alignof) and stack space incorrect for SIMD types.

The following code compiled and run with dmd-2.071 prints: SIMD vector types inherit their base type's alignment: float8.alignof yields 4 but should yield 32. void16.alignof yields 1 but should yield 16. When embedded, a float4 exposes some hidden alignment, as it changes to 16. The float8 needs a 32 byte alignment though or we SEGFAULT as soon as we load c[0] or c[1]. Even worse, the stack space occupied by these 32 byte vectors is only 16! ------------------------- void main() { import core.simd; import std.stdio; import std.math; union Matrix4x4 { float[16] a; float4[4] b; float8[2] c; alias a this; } float8 onStackA; float8 onStackB; writeln("SIMD vector types inherit their base type's alignment:"); writefln("float8.alignof yields %s but should yield 32.", float8.alignof); writefln("void16.alignof yields %s but should yield 16.", void16.alignof); writefln("When embedded, a float4 exposes some hidden alignment, as it changes to %s.", Matrix4x4.alignof); writeln("The float8 needs a 32 byte alignment though or we SEGFAULT as soon as we load c[0] or c[1]."); writefln("Even worse, the stack space occupied by these %s byte vectors is only %s!", float8.sizeof, abs(cast(ptrdiff_t)&onStackB - cast(ptrdiff_t)&onStackA)); } ------------------------- Compiled programs with too little stack space or incorrect alignment of SIMD types crash.

Just a clarification, that doesn't change the subject of this report: It occurred to me today that the stack is aligned to 16 bytes and therefore we cannot guarantee correct alignment for AVX vectors placed on the stack.

Ok, different sources report that major compilers (llvm, gcc, icc) respect alignments >= 16 on the stack. I'll open a separate issue for that.

Partial fix: https://github.com/dlang/dmd/pull/6265 dmd simply doesn't support 32 byte vectors at the moment, so the rest will have to wait.

(In reply to Marco Leise from comment #0) > writefln("float8.alignof yields %s but should yield 32.", > float8.alignof); > writefln("void16.alignof yields %s but should yield 16.", > void16.alignof); Fixed in https://github.com/dlang/dmd/pull/6265 > writefln("When embedded, a float4 exposes some hidden alignment, as it > changes to %s.", Matrix4x4.alignof); Fixed in https://github.com/dlang/dmd/pull/6582 (issue 17237)