conv needs to be picky about aliases and introduces a temp for addr conv
#24818
Conversation
|
so, taking strict aliasing into account, the correct code to generate in c is something like: void main(NU64* v) {
unsigned long long tmp = (unsigned long long)(*v);
foo(&tmp);
*v = tmp
}is that the outcome after this PR? ie casting the pointer itself will lead to UB. |
|
Okay, I was going to generate a temp for |
In this PR, It introduces a temp for the case of T1_ = ((unsigned long long) (m_1));
foo(((&T1_)));The rest could be resolved in the future It requires #24817 to make CI green |
ringabout
changed the title
conv needs to be picky about aliasesconv needs to be picky about aliases and introduces a temp for addr conv
compiler/ccgexprs.nim
Outdated
| @@ -962,6 +962,11 @@ proc cowBracket(p: BProc; n: PNode) = | |||
| proc cow(p: BProc; n: PNode) {.inline.} = | |||
| if n.kind == nkHiddenAddr: cowBracket(p, n[0]) | |||
|
|
|||
| template noSomeCast(e: PNode): bool = | |||
There was a problem hiding this comment.
What is noSomeCast? Name it ignoreConv or something that describes what this does.
|
Thanks for your hard work on this PR! Hint: mm: orc; opt: speed; options: -d:release |
…r conv` (#24818) ref #24817 ref #24815 ref status-im/nim-eth#784 ```nim {.emit:""" void foo(unsigned long long* x) { } """.} proc foo(x: var culonglong) {.importc: "foo", nodecl.} proc main(x: var uint64) = # var s: culonglong = u # TODO: var m = uint64(12) # var s = culonglong(m) foo(culonglong m) var u = uint64(12) main(u) ``` Notes that this code gives incompatible errors in 2.0.0, 2.2.0 and the devel branch. With this PR, `conv` is kept, but it seems to go back to #24807 (cherry picked from commit f9c8775)
ref #24817
ref #24815
ref status-im/nim-eth#784
{.emit:""" void foo(unsigned long long* x) { } """.} proc foo(x: var culonglong) {.importc: "foo", nodecl.} proc main(x: var uint64) = # var s: culonglong = u # TODO: var m = uint64(12) # var s = culonglong(m) foo(culonglong m) var u = uint64(12) main(u)Notes that this code gives incompatible errors in 2.0.0, 2.2.0 and the devel branch. With this PR,
convis kept, but it seems to go back to #24807