Fix issue 20012 - clear diagnostic for CTFE array cast in UDA context by usefahmed07 · Pull Request #23820 · dlang/dmd

usefahmed07 · 2026-09-09T10:13:14Z

Fixes #20012

When an array cast that changes element size (e.g. cast(dstring) from a string literal) appeared in a context requiring compile-time evaluation but going through the codegen-lowering path (such as a UDA), the compiler rewrote it to a call to object.__ArrayCast!(...). CTFE then had to interpret that call's body, including onArrayCastError()'s use of pureMalloc for building the error message - which cannot be evaluated at compile time. This surfaced a confusing, unrelated error (e.g. `fakePureErrno` cannot be interpreted at compile time) instead of a clear message, unlike the equivalent enum case which hits dinterpret.d's CTFE-aware array cast check directly and gives a clear message. Added a __ctfe branch in onArrayCastError() (druntime) that builds the same error message using plain string concatenation instead of pureMalloc, since that's evaluable during CTFE. The non-CTFE path is untouched. Added test/fail_compilation/fail20012.d covering the original UDA repro from the issue. Verified casting.d's existing unittests (covering the ordinary runtime path) still pass, and the full fail_compilation/compilable suites pass with no regressions.

thewilsonator · 2026-09-09T12:36:10Z

no, runnable/betterc.d fails every time on every platform, it looks very much related.

dkorpel

This is still printing multiple redundant messages

The CTFE branch of onArrayCastError used ~ and .idup, which are heap-allocating operations. Since __ArrayCast is @nogc, the D compiler's attribute inference marked onArrayCastError as non-@nogc based on this branch, breaking the @nogc call from __ArrayCast and producing a confusing secondary 'CTFE failed because of previous errors' message alongside the real one. Build the CTFE error message on a stack-allocated char[2048] buffer instead, with no ~ or .idup, so attribute inference keeps the function @nogc. This also fixes the same issue under -betterC. Fixes https://issues.dlang.org/show_bug.cgi?id=22546

The expected error message hardcodes '4LU' for the toElemSize argument printed by the compiler's call-site diagnostics. This literal suffix reflects size_t being ulong on 64-bit platforms; on 32-bit platforms size_t is uint, so the compiler prints '4u' instead, causing a text mismatch. Disable the test on 32-bit linux/windows targets rather than attempt to match both formats, consistent with existing tests like chkformat_clong_smalllong.d that hit the same size_t size-dependent formatting issue.

Adding the DISABLED: linux32 win32 line shifted the source code down by one line, so the expected error line numbers (10, 11) no longer matched the actual output (11, 12). Update the expected TEST_OUTPUT line numbers accordingly.

Previous CI failure was an unrelated apt-get/Google Chrome mirror hash mismatch during runner setup, unrelated to this change.

Previous CI failure was an unrelated Google Chrome apt mirror hash mismatch during runner setup, unrelated to this change.

…rror issue)

Previous CI failure was an unrelated Google Chrome apt mirror hash mismatch during runner setup, unrelated to this change.

usefahmed07 · 2026-09-09T21:03:51Z

@thewilsonator Looking good now — btw the buildkite/dmd CI (which is the required check) was failing on all previous PRs, but it's passing now.

dkorpel · 2026-09-10T11:06:04Z

My comment has not been addressed yet

thewilsonator · 2026-09-10T11:31:17Z

This adds one branch if (__ctfe) that ends in assert(false, ...);, that replaces the error for the second line to

/dlang/dmd-beta/linux/bin64/../../src/druntime/import/core/memory.d(1060): Error: `fakePureErrno` cannot be interpreted at compile time, because it has no available source code
    const errnosave = fakePureErrno;
                      ^

There are two errors given, because each line of the test case is an error.

What do you mean that "This is still printing multiple redundant messages"?

dkorpel · 2026-09-10T12:02:37Z

The expected error message "Error: array cast from string to dstring is not supported at compile time" is not printed in the UDA case, instead this PR duplicates a bunch of druntime code to be able to print the run-time error message at ctfe. It's inconsistent, exposes an ugly druntime template trace in the message, creates code duplication, and is symptom fighting while keeping the root cause of the issue very much alive.

UserAttributeDeclaration.semantic evaluated attribute expressions with a normal (non-CTFE) scope, unlike enum initializers which use sc.startCTFE(). This caused array casts inside UDA context to be lowered to a runtime __ArrayCast call instead of being evaluated directly by the CTFE interpreter, producing a messy druntime error trace instead of the clean "array cast from ... is not supported at compile time" message. Wrap the UDA attribute semantic in startCTFE()/endCTFE(), matching the enum case, so CTFE handles the diagnostic directly.

The previous approach duplicated druntime code in onArrayCastError to print a CTFE-safe message, and tweaked the __ArrayCast lowering condition in expressionsem.d. Both are reverted here in favor of a minimal fix to UDA attribute semantic (see next commit).

The previous revert commit mistakenly used origin/master as the target, which also discarded three unrelated legitimate commits on this branch (issue 5010, issue 18586, and the AA array comparison fix) that touched the same files. This restores expressionsem.d and casting.d to the state right before the CTFE-array-cast duplication commits, keeping those legitimate fixes intact.

github-actions · 2026-09-10T18:49:54Z

DMD perf check

No differences above the noise thresholds.

All measurements

Metric	Base	PR	Δ
compile hello.d (instr)	213.4 M	213.4 M	+0.009%
compile hello.d -O -release (instr)	231.6 M	231.7 M	+0.008%
compile Phobos (instr)	5,121.1 M	5,121.1 M	+0.001%
compile Phobos codegen (instr)	1,472.5 M	1,472.5 M	-0.001%
compile vibe.d (instr)	15,113.5 M	15,113.7 M	+0.001%
dmd binary size (stripped)	6.87 MB	6.87 MB	0.00%
hello binary size (stripped)	0.72 MB	0.72 MB	0.00%
peak RSS (compile hello.d)	43.13 MB	43.18 MB	+0.11%
peak RSS (compile Phobos)	617.7 MB	619.9 MB	+0.37%
peak RSS (compile vibe.d)	1919 MB	1919 MB	+0.03%
compile dmd itself (wall)	12.0 s	12.1 s	+0.89%
compile hello.d (wall)	63.6 ms	63.1 ms	-0.66%
compile Phobos (wall)	1,546 ms	1,519 ms	-1.75%

_{2379886 vs merge-base 70145d6 · about these metrics}

usefahmed07 · 2026-09-10T22:28:20Z

@dkorpel @thewilsonator Pushed a new fix.

Root cause: UDA attributes weren't evaluated in a CTFE scope like enum initializers are, so the cast got lowered to a runtime __ArrayCast() call instead of being handled directly by CTFE.

Fix: wrap UDA semantic in startCTFE()/endCTFE() (5-line change in dsymbolsem.d). Reverted all the druntime duplication and the expressionsem.d change from before.

Now both enum and UDA cases print the same clean message, no druntime trace, no __ctfe branch. Full test suite passes.

GitHub issue dlang#20012 (linked by dkorpel) confirms this was transferred from bugzilla bug 22546 by dlangBugzillaToGithub. The bugzilla page at issues.dlang.org/show_bug.cgi?id=20012 is an unrelated bug (extern(C) mangling in template mixins) using bugzilla's own numbering, which differs from GitHub's issue numbering after the migration. The original id=22546 link was correct all along.

Point directly to dlang#20012 (where this issue now lives) instead of the old bugzilla URL it was transferred from.

thewilsonator approved these changes Sep 9, 2026

View reviewed changes

usefahmed07 force-pushed the fix-ctfe-arraycast-uda-msg-20012 branch from 5441c48 to 7402c95 Compare September 9, 2026 10:23

usefahmed07 requested a review from thewilsonator September 9, 2026 11:20

dkorpel requested changes Sep 9, 2026

View reviewed changes

usefahmed07 added 11 commits September 9, 2026 19:21

chore: trigger CI re-run

2408300

Previous CI failure was an unrelated apt-get/Google Chrome mirror hash mismatch during runner setup, unrelated to this change.

chore: trigger CI re-run

e45383a

Previous CI failure was an unrelated apt-get/Google Chrome mirror hash mismatch during runner setup, unrelated to this change.

chore: trigger CI re-run

d7eadd9

Previous CI failure was an unrelated apt-get/Google Chrome mirror hash mismatch during runner setup, unrelated to this change.

chore: trigger CI re-run

8879645

Previous CI failure was an unrelated apt-get/Google Chrome mirror hash mismatch during runner setup, unrelated to this change.

chore: trigger CI re-run

f65b8c4

Previous CI failure was an unrelated Google Chrome apt mirror hash mismatch during runner setup, unrelated to this change.

chore: trigger CI re-run (unrelated GitHub runner image Chrome apt mi…

48c0164

…rror issue)

chore: trigger CI re-run (unrelated GitHub runner image Chrome apt mi…

71612da

…rror issue)

chore: trigger CI re-run

f23adf6

Previous CI failure was an unrelated Google Chrome apt mirror hash mismatch during runner setup, unrelated to this change.

thewilsonator requested a review from dkorpel September 10, 2026 08:00

usefahmed07 added 3 commits September 10, 2026 21:03

dkorpel reviewed Sep 10, 2026

View reviewed changes

Comment thread compiler/test/fail_compilation/fail20012.d Outdated

Fix wrong issue number in fail20012.d comment (22546 -> 20012)

05f4230

usefahmed07 requested a review from dkorpel September 10, 2026 22:42

usefahmed07 added 3 commits September 11, 2026 01:53

Remove incorrect issue link from fail20012.d (no matching issue found)

a0b8e41

Use GitHub issue link instead of bugzilla in fail20012.d

2379886

Point directly to dlang#20012 (where this issue now lives) instead of the old bugzilla URL it was transferred from.

thewilsonator approved these changes Sep 11, 2026

View reviewed changes

usefahmed07 requested a review from thewilsonator September 11, 2026 12:25

Navigation Menu

Uh oh!

Uh oh!

Fix issue 20012 - clear diagnostic for CTFE array cast in UDA context - #23820

Fix issue 20012 - clear diagnostic for CTFE array cast in UDA context#23820
usefahmed07 wants to merge 19 commits into
dlang:masterfrom
usefahmed07:fix-ctfe-arraycast-uda-msg-20012

usefahmed07 commented Sep 9, 2026

Uh oh!

thewilsonator commented Sep 9, 2026

Uh oh!

dkorpel left a comment

Uh oh!

usefahmed07 commented Sep 9, 2026

Uh oh!

dkorpel commented Sep 10, 2026

Uh oh!

thewilsonator commented Sep 10, 2026

Uh oh!

dkorpel commented Sep 10, 2026

Uh oh!

github-actions Bot commented Sep 10, 2026 •

edited

Loading

Uh oh!

usefahmed07 commented Sep 10, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Uh oh!

Conversation

usefahmed07 commented Sep 9, 2026

Uh oh!

thewilsonator commented Sep 9, 2026

Uh oh!

dkorpel left a comment

Choose a reason for hiding this comment

Uh oh!

usefahmed07 commented Sep 9, 2026

Uh oh!

dkorpel commented Sep 10, 2026

Uh oh!

thewilsonator commented Sep 10, 2026

Uh oh!

dkorpel commented Sep 10, 2026

Uh oh!

github-actions Bot commented Sep 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

DMD perf check

Uh oh!

usefahmed07 commented Sep 10, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

github-actions Bot commented Sep 10, 2026 •

edited

Loading