[Capability System] Fix bug where capabilities do not correctly propegate if AST-parent has target+set the AST-child does not (#8175)

Fixes: #8174

Changes:
* To determine if we propagate capabilities, we need to ensure that a
`join` will do nothing (optimization since `join` is expensive + caching
data for the `join` adds up to be expensive). This logic was changed in
`slang-check-decl.cpp` since the current logic was incorrect.
* A parent could have the set `metal+glsl` and the use-site could have
`glsl`. In this case, we will not remove `metal` from the parent since
`{metal+glsl}.implies({glsl})` is true.

---------

Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>

5 files changed, 119 insertions, 48 deletions

diff --git a/source/slang/glsl.meta.slang b/source/slang/glsl.meta.slang
index eeaf2a58c..5a78a9960 100644
--- a/source/slang/glsl.meta.slang
+++ b/source/slang/glsl.meta.slang
@@ -301,7 +301,7 @@ public extension vector<T, 3>
 
 [ForceInline]
 [OverloadRank(15)]
-[require(cpp_cuda_glsl_hlsl_spirv)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
 public bool operator==<T:__BuiltinArithmeticType, let N:int>(vector<T, N> left, vector<T, N> right)
 {
     return all(equal(left, right));
@@ -309,7 +309,7 @@ public bool operator==<T:__BuiltinArithmeticType, let N:int>(vector<T, N> left,
 
 [ForceInline]
 [OverloadRank(15)]
-[require(cpp_cuda_glsl_hlsl_spirv)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
 public bool operator!=<T:__BuiltinArithmeticType, let N:int>(vector<T, N> left, vector<T, N> right)
 {
     return any(notEqual(left, right));
@@ -317,7 +317,7 @@ public bool operator!=<T:__BuiltinArithmeticType, let N:int>(vector<T, N> left,
 
 [ForceInline]
 [OverloadRank(14)]
-[require(cpp_cuda_glsl_hlsl_spirv)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
 public bool operator==<T:__BuiltinFloatingPointType, let N:int>(vector<T, N> left, vector<T, N> right)
 {
     return all(equal(left, right));
@@ -325,7 +325,7 @@ public bool operator==<T:__BuiltinFloatingPointType, let N:int>(vector<T, N> lef
 
 [ForceInline]
 [OverloadRank(14)]
-[require(cpp_cuda_glsl_hlsl_spirv)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
 public bool operator!=<T:__BuiltinFloatingPointType, let N:int>(vector<T, N> left, vector<T, N> right)
 {
     return any(notEqual(left, right));
@@ -333,7 +333,7 @@ public bool operator!=<T:__BuiltinFloatingPointType, let N:int>(vector<T, N> lef
 
 [ForceInline]
 [OverloadRank(14)]
-[require(cpp_cuda_glsl_hlsl_spirv)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
 public bool operator==<T:__BuiltinLogicalType, let N:int>(vector<T, N> left, vector<T, N> right)
 {
     return all(equal(left, right));
@@ -341,7 +341,7 @@ public bool operator==<T:__BuiltinLogicalType, let N:int>(vector<T, N> left, vec
 
 [ForceInline]
 [OverloadRank(14)]
-[require(cpp_cuda_glsl_hlsl_spirv)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
 public bool operator!=<T:__BuiltinLogicalType, let N:int>(vector<T, N> left, vector<T, N> right)
 {
     return any(notEqual(left, right));
@@ -354,7 +354,7 @@ for (auto type : kBaseTypes) {
 }}}}
 [ForceInline]
 [OverloadRank(15)]
-[require(cpp_cuda_glsl_hlsl_spirv)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
 public bool operator==<let N:int>(vector<$(typeName), N> left, vector<$(typeName), N> right)
 {
     return all(equal(left, right));
@@ -362,7 +362,7 @@ public bool operator==<let N:int>(vector<$(typeName), N> left, vector<$(typeName
 
 [ForceInline]
 [OverloadRank(15)]
-[require(cpp_cuda_glsl_hlsl_spirv)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
 public bool operator!=<let N:int>(vector<$(typeName), N> left, vector<$(typeName), N> right)
 {
     return any(notEqual(left, right));
@@ -6264,7 +6264,7 @@ public void traceRayMotionNV(
 
 // GL_KHR_shader_subgroup_basic Built-in Variables
 
-[require(cpp_cuda_glsl_hlsl_spirv_wgsl, subgroup_basic)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl, subgroup_basic)]
 void requireGLSLExtForSubgroupBasicBuiltin() {
     __target_switch
     {
@@ -6288,7 +6288,7 @@ void setupExtForSubgroupBasicBuiltIn() {
 }
 
 __spirv_version(1.3)
-[require(cpp_cuda_glsl_hlsl_spirv_wgsl, subgroup_ballot)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl, subgroup_ballot)]
 void requireGLSLExtForSubgroupBallotBuiltin() {
     __target_switch
     {
@@ -6301,7 +6301,7 @@ void requireGLSLExtForSubgroupBallotBuiltin() {
 }
 
 __spirv_version(1.3)
-[require(cpp_cuda_glsl_hlsl_spirv_wgsl, subgroup_ballot)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl, subgroup_ballot)]
 void setupExtForSubgroupBallotBuiltIn() {
     __target_switch
     {
@@ -8204,7 +8204,7 @@ __generic<T : __BuiltinType>
 __spirv_version(1.3)
 __glsl_extension(GL_KHR_shader_subgroup_quad)
 [ForceInline]
-[require(glsl_hlsl_spirv, subgroup_quad)]
+[require(glsl_hlsl_metal_spirv, subgroup_quad)]
 public T subgroupQuadBroadcast(T value, uint id)
 {
     shader_subgroup_preamble<T>();
@@ -8247,9 +8247,9 @@ public T subgroupQuadSwapDiagonal(T value)
 
 __generic<T : __BuiltinType, let N : int>
 __spirv_version(1.3)
-__glsl_extension(GL_KHR_shader_subgroup_quad)
+    __glsl_extension(GL_KHR_shader_subgroup_quad)
 [ForceInline]
-[require(glsl_hlsl_spirv, subgroup_quad)]
+    [require(glsl_hlsl_metal_spirv, subgroup_quad)]
 public vector<T,N> subgroupQuadBroadcast(vector<T,N> value, uint id)
 {
     shader_subgroup_preamble<T>();
diff --git a/source/slang/hlsl.meta.slang b/source/slang/hlsl.meta.slang
index b8243d6e4..740f48320 100644
--- a/source/slang/hlsl.meta.slang
+++ b/source/slang/hlsl.meta.slang
@@ -4804,11 +4804,11 @@ __intrinsic_op($(kIROp_ByteAddressBufferLoad))
 T __byteAddressBufferLoad<T>(RasterizerOrderedByteAddressBuffer buffer, uint offset, uint alignment);
 
 __intrinsic_op($(kIROp_ByteAddressBufferStore))
-[require(cpp_cuda_glsl_hlsl_metal_spirv, byteaddressbuffer_rw)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl, byteaddressbuffer_rw)]
 void __byteAddressBufferStore<T>(RWByteAddressBuffer buffer, uint offset, uint alignment, T value);
 
 __intrinsic_op($(kIROp_ByteAddressBufferStore))
-[require(cpp_cuda_glsl_hlsl_metal_spirv, byteaddressbuffer_rw)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl, byteaddressbuffer_rw)]
 void __byteAddressBufferStore<T>(RasterizerOrderedByteAddressBuffer buffer, uint offset, uint alignment, T value);
 
 __intrinsic_op($(kIROp_GetUntypedBufferPtr))
@@ -5095,7 +5095,7 @@ struct $(item.name)
     /// When targeting non-HLSL, the status is always 0.
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     uint2 Load2(uint location)
     {
         __target_switch
@@ -5108,7 +5108,7 @@ struct $(item.name)
 
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     uint2 Load2Aligned(uint location, uint alignment)
     {
         __target_switch
@@ -5125,7 +5125,7 @@ struct $(item.name)
     ///@return `uint2` Two 32-bit unsigned integers loaded from the buffer.
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     uint2 Load2Aligned(uint location)
     {
         __target_switch
@@ -5163,7 +5163,7 @@ struct $(item.name)
     /// When targeting non-HLSL, the status is always 0.
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     uint3 Load3(uint location)
     {
         __target_switch
@@ -5176,7 +5176,7 @@ struct $(item.name)
 
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     uint3 Load3Aligned(uint location, uint alignment)
     {
         __target_switch
@@ -5193,7 +5193,7 @@ struct $(item.name)
     ///@return `uint3` Three 32-bit unsigned integer value loaded from the buffer.
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     uint3 Load3Aligned(uint location)
     {
         __target_switch
@@ -5230,7 +5230,7 @@ struct $(item.name)
     /// If any values were taken from an unmapped tile, `CheckAccessFullyMapped` returns FALSE.
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     uint4 Load4(uint location)
     {
         __target_switch
@@ -5243,7 +5243,7 @@ struct $(item.name)
 
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     uint4 Load4Aligned(uint location, uint alignment)
     {
         __target_switch
@@ -5260,7 +5260,7 @@ struct $(item.name)
     ///@return `uint4` Four 32-bit unsigned integer value loaded from the buffer.
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     uint4 Load4Aligned(uint location)
     {
         __target_switch
@@ -5284,7 +5284,7 @@ struct $(item.name)
 
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     T Load<T>(uint location)
     {
         return __byteAddressBufferLoad<T>(this, location, 0);
@@ -5292,7 +5292,7 @@ struct $(item.name)
 
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     T LoadAligned<T>(uint location, uint alignment)
     {
         return __byteAddressBufferLoad<T>(this, location, alignment);
@@ -5305,7 +5305,7 @@ struct $(item.name)
     ///Currently, this function only supports when `T` is scalar, vector, or matrix type.
     [__NoSideEffect]
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     T LoadAligned<T>(uint location)
     {
         return __byteAddressBufferLoad<T>(this, location, __naturalStrideOf<T>());
@@ -5828,7 +5828,7 @@ ${{{{
     ///@param address The input address in bytes, which must be a multiple of 4.
     ///@param alignment Specifies the alignment of the location, which must be a multiple of 4.
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store(uint address, uint value)
     {
         __target_switch
@@ -5845,7 +5845,7 @@ ${{{{
     ///@param value Two input values.
     ///@param alignment Specifies the alignment of the location, which must be a multiple of 4.
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store2(uint address, uint2 value)
     {
         __target_switch
@@ -5858,7 +5858,7 @@ ${{{{
 
 
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store2(uint address, uint2 value, uint alignment)
     {
         __target_switch
@@ -5874,7 +5874,7 @@ ${{{{
     ///@param address The input address in bytes, which must be a multiple of 8.
     ///@param value Two input values.
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store2Aligned(uint address, uint2 value)
     {
         __target_switch
@@ -5890,7 +5890,7 @@ ${{{{
     ///@param value Three input values.
     ///@param alignment Specifies the alignment of the location, which must be a multiple of 4.
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store3(uint address, uint3 value)
     {
         __target_switch
@@ -5902,7 +5902,7 @@ ${{{{
     }
 
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store3(uint address, uint3 value, uint alignment)
     {
         __target_switch
@@ -5918,7 +5918,7 @@ ${{{{
     ///@param address The input address in bytes, which must be a multiple of 12.
     ///@param value Three input values.
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store3Aligned(uint address, uint3 value)
     {
         __target_switch
@@ -5934,7 +5934,7 @@ ${{{{
     ///@param value Four input values.
     ///@param alignment Specifies the alignment of the location, which must be a multiple of 4.
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store4(uint address, uint4 value)
     {
         __target_switch
@@ -5947,7 +5947,7 @@ ${{{{
 
 
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store4(uint address, uint4 value, uint alignment)
     {
         __target_switch
@@ -5963,7 +5963,7 @@ ${{{{
     ///@param address The input address in bytes, which must be a multiple of 16.
     ///@param value Four input values.
     [ForceInline]
-    [require(cpp_cuda_glsl_hlsl_metal_spirv)]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store4Aligned(uint address, uint4 value)
     {
         __target_switch
@@ -5975,12 +5975,14 @@ ${{{{
     }
 
     [ForceInline]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store<T>(uint address, T value)
     {
         __byteAddressBufferStore(this, address, 0, value);
     }
 
     [ForceInline]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void Store<T>(uint address, T value, uint alignment)
     {
         __byteAddressBufferStore(this, address, alignment, value);
@@ -5992,6 +5994,7 @@ ${{{{
     ///@param address The input address in bytes, which must be a multiple of size of `T`.
     ///@param value The input value.
     [ForceInline]
+    [require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
     void StoreAligned<T>(uint address, T value)
     {
         __byteAddressBufferStore(this, address, __naturalStrideOf<T>(), value);
@@ -8806,7 +8809,7 @@ vector<T,N> fdim(vector<T,N> x, vector<T,N> y)
 /// @category math
 __generic<T : __BuiltinFloatingPointType>
 [__readNone]
-[require(cpp_cuda_glsl_hlsl_metal_spirv)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
 T divide(T x, T y)
 {
     __target_switch
@@ -8819,7 +8822,7 @@ T divide(T x, T y)
 
 __generic<T : __BuiltinFloatingPointType, let N: int>
 [__readNone]
-[require(cpp_cuda_glsl_hlsl_metal_spirv)]
+[require(cpp_cuda_glsl_hlsl_metal_spirv_wgsl)]
 vector<T,N> divide(vector<T,N> x, vector<T,N> y)
 {
     __target_switch
diff --git a/source/slang/slang-capability.cpp b/source/slang/slang-capability.cpp
index a965ecd93..0c883b3b8 100644
--- a/source/slang/slang-capability.cpp
+++ b/source/slang/slang-capability.cpp
@@ -529,36 +529,68 @@ bool CapabilitySet::implies(CapabilityAtom atom) const
     return this->implies(tmpSet);
 }
 
+// Implication depends heavily on context as per the `ImpliesFlags`.
 CapabilitySet::ImpliesReturnFlags CapabilitySet::_implies(
     CapabilitySet const& otherSet,
     ImpliesFlags flags) const
 {
-    // x implies (c | d) only if (x implies c) and (x implies d).
+    // By default (`ImpliesFlags::None`): x implies (c | d) only if (x implies c) and (x implies d).
 
     bool onlyRequireSingleImply = ((int)flags & (int)ImpliesFlags::OnlyRequireASingleValidImply);
+    bool cannotHaveMoreTargetAndStageSets =
+        ((int)flags & (int)ImpliesFlags::CannotHaveMoreTargetAndStageSets);
+    bool canHaveSubsetOfTargetAndStageSets =
+        ((int)flags & (int)ImpliesFlags::CanHaveSubsetOfTargetAndStageSets);
+
     int flagsCollected = (int)CapabilitySet::ImpliesReturnFlags::NotImplied;
 
     if (otherSet.isEmpty())
         return CapabilitySet::ImpliesReturnFlags::Implied;
 
-    for (const auto& otherTarget : otherSet.m_targetSets)
+    // If empty, and the other is not empty, it does not matter what flags are used,
+    // `this` is considered to "not imply" another set. This is important since
+    // `T.join(U)` causes `T == U`.
+    if (this->isEmpty())
+        return CapabilitySet::ImpliesReturnFlags::NotImplied;
+
+    if (cannotHaveMoreTargetAndStageSets &&
+        this->getCapabilityTargetSets().getCount() > otherSet.getCapabilityTargetSets().getCount())
     {
-        auto thisTarget = this->m_targetSets.tryGetValue(otherTarget.first);
+        return CapabilitySet::ImpliesReturnFlags::NotImplied;
+    }
+
+    for (const auto& otherTargetPair : otherSet.m_targetSets)
+    {
+        auto thisTarget = this->m_targetSets.tryGetValue(otherTargetPair.first);
+        const auto& otherTarget = otherTargetPair.second;
         if (!thisTarget)
         {
             if (onlyRequireSingleImply)
                 continue;
+
+            if (canHaveSubsetOfTargetAndStageSets)
+                continue;
             // 'this' lacks a target 'other' has.
             return CapabilitySet::ImpliesReturnFlags::NotImplied;
         }
 
-        for (const auto& otherStage : otherTarget.second.shaderStageSets)
+        if (cannotHaveMoreTargetAndStageSets && thisTarget->getShaderStageSets().getCount() >
+                                                    otherTarget.getShaderStageSets().getCount())
+        {
+            return CapabilitySet::ImpliesReturnFlags::NotImplied;
+        }
+
+        for (const auto& otherStagePair : otherTarget.getShaderStageSets())
         {
-            auto thisStage = thisTarget->shaderStageSets.tryGetValue(otherStage.first);
+            auto thisStage = thisTarget->shaderStageSets.tryGetValue(otherStagePair.first);
+            const auto& otherStage = otherStagePair.second;
             if (!thisStage)
             {
                 if (onlyRequireSingleImply)
                     continue;
+
+                if (canHaveSubsetOfTargetAndStageSets)
+                    continue;
                 // 'this' lacks a stage 'other' has.
                 return CapabilitySet::ImpliesReturnFlags::NotImplied;
             }
@@ -567,9 +599,9 @@ CapabilitySet::ImpliesReturnFlags CapabilitySet::_implies(
             if (thisStage->atomSet)
             {
                 auto& thisStageSet = thisStage->atomSet.value();
-                if (otherStage.second.atomSet)
+                if (otherStage.atomSet)
                 {
-                    auto contained = thisStageSet.contains(otherStage.second.atomSet.value());
+                    auto contained = thisStageSet.contains(otherStage.atomSet.value());
                     if (!onlyRequireSingleImply && !contained)
                     {
                         return CapabilitySet::ImpliesReturnFlags::NotImplied;
@@ -593,12 +625,20 @@ bool CapabilitySet::implies(CapabilitySet const& other) const
     return (int)_implies(other, ImpliesFlags::None) &
            (int)CapabilitySet::ImpliesReturnFlags::Implied;
 }
+
 CapabilitySet::ImpliesReturnFlags CapabilitySet::atLeastOneSetImpliedInOther(
     CapabilitySet const& other) const
 {
     return _implies(other, ImpliesFlags::OnlyRequireASingleValidImply);
 }
 
+bool CapabilitySet::joinWithOtherWillChangeThis(CapabilitySet const& other) const
+{
+    return !(
+        (int)_implies(other, ImpliesFlags::CannotHaveMoreTargetAndStageSets) &
+        (int)CapabilitySet::ImpliesReturnFlags::Implied);
+}
+
 void CapabilityTargetSet::unionWith(const CapabilityTargetSet& other)
 {
     for (auto otherStageSet : other.shaderStageSets)
diff --git a/source/slang/slang-capability.h b/source/slang/slang-capability.h
index 43f933620..c72d1d7d6 100644
--- a/source/slang/slang-capability.h
+++ b/source/slang/slang-capability.h
@@ -168,9 +168,12 @@ public:
         Implied = 1 << 0,
     };
     /// Does this capability set imply all the capabilities in `other`?
+    /// `this` can have excess target+stage sets.
     bool implies(CapabilitySet const& other) const;
     /// Does this capability set imply at least 1 set in other.
     ImpliesReturnFlags atLeastOneSetImpliedInOther(CapabilitySet const& other) const;
+    /// Will a `join` with `other` change `this`?
+    bool joinWithOtherWillChangeThis(CapabilitySet const& other) const;
 
     /// Does this capability set imply the atomic capability `other`?
     bool implies(CapabilityAtom other) const;
@@ -353,8 +356,29 @@ private:
 
     enum class ImpliesFlags
     {
+        // All permutations of target+stage from `other` must be implied by a target+stage
+        // in `this`.
         None = 0,
+        // Given a single target+stage permutation, if 1 permutation is implied in `other`,
+        // return true.
         OnlyRequireASingleValidImply = 1 << 0,
+        // The target+stage permuations in `this` cannot have extra permutations
+        // relative to `other`.
+        // Ex: `{metal|glsl}.implies({glsl})` is false
+        //     `{glsl}.implies({glsl|metal})` is false
+        //     `{glsl}.implies({glsl|glsl})` is true
+        CannotHaveMoreTargetAndStageSets = 1 << 1,
+        // The target+stage permuations in `this` can have less permutations
+        // than `other`. This means, only for the shared permutations of `this`
+        // and `other` does `thisSet[target][stage].imply(otherSet)` have to be
+        // true.
+        // If `this` is empty, `this` is not able to imply `other` unless `other`
+        // is empty.
+        // Ex: `{glsl}.implies({glsl|metal})` is true since we only compare shared-permutations.
+        CanHaveSubsetOfTargetAndStageSets = 1 << 2,
+
+        WillAJoinWithOtherModifyThis =
+            CannotHaveMoreTargetAndStageSets | CanHaveSubsetOfTargetAndStageSets
     };
     ImpliesReturnFlags _implies(CapabilitySet const& other, ImpliesFlags flags) const;
 };
diff --git a/source/slang/slang-check-decl.cpp b/source/slang/slang-check-decl.cpp
index 0a9853012..507e12fa6 100644
--- a/source/slang/slang-check-decl.cpp
+++ b/source/slang/slang-check-decl.cpp
@@ -14268,7 +14268,11 @@ static void _propagateRequirement(
         ensureDecl(visitor, referencedDecl, DeclCheckState::CapabilityChecked);
     }
 
-    if (resultCaps.implies(nodeCaps))
+    // If we do not have the same target+stage, we need to `join` to remove excess target+stage.
+    //
+    // If we have the same target+stage but current capabilities do not imply incoming capabilities,
+    // we need to `join`.
+    if (!resultCaps.joinWithOtherWillChangeThis(nodeCaps))
         return;
 
     auto oldCaps = resultCaps;