[DNI] Add -fgl-remap-z to remap SV_Position.z for GLSL vertex output

[nv-slang-bot]

Motivation

Shaders authored against the OpenGL clip-space convention emit SV_Position.z
in the [-w, w] range (NDC depth [-1, 1] after the perspective divide). The
standard convention used by Vulkan/SPIR-V/D3D/Metal — and the default the GLSL
backend targets — is [0, w] (NDC depth [0, 1]). When such a shader is
cross-compiled to textual GLSL for a desktop-GL pipeline that was set up for
the [0, 1] convention (e.g. via glClipControl(... GL_ZERO_TO_ONE) or a host
that does its own depth handling), the depth comes out wrong with no per-shader
knob to correct it.

This is not a DXC-parity gap (D3D, Vulkan, and Metal all already share the
[0, 1] depth range) — it is the one place desktop GL differs, so it needs a
GLSL-scoped opt-in rather than a target-wide behavior change.

Concretely, given:

struct VOutput { float4 v : SV_Position; }
VOutput main() { VOutput o; o.v = float4(1, 2, 3, 4); return o; }

-target glsl ... -fgl-remap-z now emits, just before the write to gl_Position:

vec4 _S1 = output_0.v_0;
_S1.z = (output_0.v_0.z + output_0.v_0.w) / 2.0;
gl_Position = _S1;

i.e. z' = (z + w) / 2, mapping [-w, w] to [0, w].

Proposed solution

Add a new compiler option -fgl-remap-z that runs a small IR fix-up pass on the
position output, directly mirroring the existing -fvk-invert-y /
-fvk-use-dx-position-w machinery. The remap is affine and reads both the z
(index 2) and w (index 3) components of the position vector, so unlike the
single-component y inversion it must operate on the full float4.

Scope is deliberately narrow, per maintainer direction on #11599:

• Target: textual GLSL only. SPIR-V also passes isKhronosTarget, so the
  pass is gated on target == CodeGenTarget::GLSL inside that block to keep
  SPIR-V/Vulkan untouched.
• Stage: vertex only. Because the pass rewrites the whole module's position
  outputs, it runs only when codegen is for exactly one entry point and that
  entry point is a vertex shader; any other case (including a mixed-stage
  whole-program request) is skipped. The stage is checked at the scheduling site,
  keeping the IR pass itself stage-agnostic and simple.
• Direction: [-1, 1] -> [0, 1] only (z' = (z + w) / 2). The reverse and
  bidirectional cases are out of scope.

The pass reuses the existing IRGLPositionOutputDecoration (no new decoration)
and is off by default, so it is a pure opt-in with no effect on existing output.

Change summary

File	What
include/slang.h	Append CompilerOptionName::GLSLRemapZ = 153 immediately before CountOf (ABI-safe append; explicit next value).
source/slang/slang-options.cpp	Register the -fgl-remap-z command-line option and route it through the existing boolean-option dispatch group.
source/slang/slang-compiler-options.cpp	Round-trip the option in writeCommandLineArgs (boolean group).
source/slang/slang-ir-vk-invert-y.{h,cpp}	New remapZOfPositionOutput pass + _remapZOfVector helper computing z' = (z + w) / 2, modeled on invertYOfPositionOutput / _invertYOfVector.
source/slang/slang-emit.cpp	Schedule the pass for GLSL + vertex only, inside the Khronos/HLSL position-fixup block.
tests/cross-compile/gl-remap-z.slang	FileCheck test for the GLSL emit, including composition with -fvk-invert-y.

Concepts and vocabulary

• IRGLPositionOutputDecoration — marks the global that lowers to
  gl_Position; both the invert-y and this remap pass find their store target by
  scanning for it (and following getElementPtr chains).
• Affine z remap — D3D/VK [0, w] <-> GL [-w, w]: z_gl = 2*z - w, inverse
  z_std = (z + w) / 2. This PR implements only the GL->standard inverse.
• isKhronosTarget — true for both textual GLSL and SPIR-V; the inner
  target == CodeGenTarget::GLSL check is what excludes SPIR-V here.

Process report

• include/slang.h — A new public CompilerOptionName is required to carry
  the boolean. Appended as GLSLRemapZ = 153 (the next value after
  TraceCoverageBoolean = 152) immediately before the CountOf sentinel, per
  the ABI rule for this enum. The historical CountOfParsableOptions = 111
  decoy sentinel is intentionally not touched.
• source/slang/slang-options.cpp — Adds the option's help-table entry and
  the case OptionKind::GLSLRemapZ: in the boolean dispatch, alongside
  VulkanInvertY / VulkanUseDxPositionW. OptionKind is a typedef of
  CompilerOptionName, so this is the same enum.
• source/slang/slang-compiler-options.cpp — Adds GLSLRemapZ to the
  boolean group in writeCommandLineArgs so the option survives serialization,
  matching its siblings.
• source/slang/slang-ir-vk-invert-y.cpp — _remapZOfVector reads z and w
  with emitSwizzle, computes (z + w) / 2 with emitAdd / emitDiv /
  getFloatValue, and writes z back with emitSwizzleSet on the full vector.
  remapZOfPositionOutput reuses invertYOfPositionOutput's store/getElementPtr
  traversal. Input-shape check: the position output is lowered as a single
  full-float4 store, so the store value is always a vector; rather than guarding
  against a non-vector store (which would silently mask an impossible shape),
  _remapZOfVector SLANG_ASSERTs the vector invariant — identical to how
  _invertYOfVector already asserts it. The shape is therefore a genuine, valid
  input produced by lowering, and the producer is correct and left untouched.
  Unlike y-negation, the remap is store-only: z' = (z + w) / 2 is not its
  own inverse, and the scope is the written stage output, so (unlike invert-y)
  there is intentionally no read-back/load branch.
• source/slang/slang-emit.cpp — remapZOfPositionOutput rewrites every
  position output in the linked module, so the gate must guarantee the module is
  a single vertex shader: the pass runs only when target == CodeGenTarget::GLSL,
  the GLSLRemapZ option is set, getEntryPointCount() == 1, and that single
  entry point's stage is Stage::Vertex. This is deliberately strict — a
  mixed-stage whole-program GLSL request (which reaches this path with
  getEntryPointCount() > 1, e.g. vertex + mesh) is left untouched rather than
  risk remapping a non-vertex stage's position. The GLSL/vertex gate lives here,
  not in the IR pass, so the IR pass stays target/stage-agnostic.
• tests/cross-compile/gl-remap-z.slang — Pins the GLSL emit shape (a temp
  whose .z is set from a z/w expression that is divided by 2.0, then
  written to gl_Position); a second run that composes -fvk-invert-y -fgl-remap-z, asserting (CHECK-DAG, order-independent) that y is additively
  inverted and z is (z + w) / 2-remapped — the two transforms touch
  independent components and compose cleanly; and a third run without the flag
  that asserts the position is written unmodified (no / 2.0), locking the
  off-by-default opt-in.

Notes for the reviewer

• Flag name is open. -fgl-remap-z is a proposal scoped to the GLSL target;
  happy to rename to whatever fits the option-naming convention — not blocking on it.
• Cherry-pickable / self-contained. This is an additive, off-by-default,
  GLSL-vertex-only opt-in with no effect on existing output, intended to be easy
  to cherry-pick per the request on the issue.
• Mesh shaders (-fvk-invert-y doesn't work on Mesh shader position output. #5761): out of scope here — the gate is vertex-stage only, so
  mesh/other stages are unaffected. Behavior for those stages is unchanged.
• Build/verification disclosure: the local build environment is disk-full,
  so this change was not compiled locally — final build and test verification
  is left to CI. To compensate, every touched API was checked against HEAD
  (CodeGenContext::getEntryPointCount/getEntryPoint/getSingleEntryPointIndex,
  EntryPoint::getStage, Stage::Vertex, SLANG_PASS, the IRBuilder emit
  signatures, the ABI append), and the GLSL swizzleSet emit shape the test pins
  was sanity-checked against the existing -fvk-invert-y output from a prebuilt
  slangc. C++ changes pass clang-format locally.

Closes #11599

[jkwak-work]

Looks good to me.
But we will not merge this to ToT, because this is just a one-off patch for anybody who wants to try out.

Slang doesn't support the legacy behavior of GLSL.

[nv-slang-bot]

Understood, and thank you for the review and the clear direction. 👍

Acknowledging the disposition: this stays a one-off, cherry-pickable reference for anyone who wants to try the OpenGL [-w, w] → [0, w] depth remap on the textual-GLSL path, and is not intended for ToT — Slang does not support the legacy GLSL depth convention. The change is deliberately additive, off-by-default, and GLSL-vertex-only, so it carries no effect on existing output and remains easy to cherry-pick as-is.

Leaving the PR open as that reference unless you'd prefer it closed — happy to close on your word.

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}

[nv-slang-bot]

/regenerate-cmdline-ref

[nv-slang-bot]

CI status note (for anyone cherry-picking this reference):

The only failing check is check-cmdline-ref (and its check-ci rollup) — not a code defect. It's that the auto-generated docs/command-line-slangc-reference.md wasn't regenerated after -fgl-remap-z was added to the slangc -help table, so the CI diff-check trips.

One-line fix for anyone with a build:

slangc -help-style markdown -h > docs/command-line-slangc-reference.md

(or /regenerate-cmdline-ref from a maintainer whose account has slash-command dispatch rights).

I couldn't run it here — no local build (disk-full, disclosed above), and the bot's /regenerate-cmdline-ref comment did not dispatch the workflow. I'm deliberately not hand-editing the auto-generated file, since without a build I can't diff-verify a byte-exact match and a near-miss would keep the check red. All correctness/build/test checks are unaffected; this is purely doc staleness.

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}

[jkwak-work]

/regenerate-cmdline-ref

[slangbot]

🌈 Regenerated command line reference, please merge the changes from this PR