New nonuniform analysis (#2457)

SPIRV/SpvBuilder.cpp

@@ -2798,8 +2798,9 @@ void Builder::accessChainStore(Id rvalue, Decoration nonUniform, spv::MemoryAcce
 }
 
 // Comments in header
-Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resultType,
-    spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment)
+Id Builder::accessChainLoad(Decoration precision, Decoration l_nonUniform,
+    Decoration r_nonUniform, Id resultType, spv::MemoryAccessMask memoryAccess,
+    spv::Scope scope, unsigned int alignment)
 {
     Id id;
 
@@ -2863,9 +2864,9 @@ Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resu
         // Buffer accesses need the access chain decorated, and this is where
         // loaded image types get decorated. TODO: This should maybe move to
         // createImageTextureFunctionCall.
-        addDecoration(id, nonUniform);
+        addDecoration(id, l_nonUniform);
         id = createLoad(id, precision, memoryAccess, scope, alignment);
-        addDecoration(id, nonUniform);
+        addDecoration(id, r_nonUniform);
     }
 
     // Done, unless there are swizzles to do
@@ -2886,7 +2887,7 @@ Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resu
     if (accessChain.component != NoResult)
         id = setPrecision(createVectorExtractDynamic(id, resultType, accessChain.component), precision);
 
-    addDecoration(id, nonUniform);
+    addDecoration(id, r_nonUniform);
     return id;
 }
 

SPIRV/SpvBuilder.h

@@ -625,6 +625,7 @@ public:
             CoherentFlags operator |=(const CoherentFlags &other) { return *this; }
 #else
             bool isVolatile() const { return volatil; }
+            bool isNonUniform() const { return nonUniform; }
             bool anyCoherent() const {
                 return coherent || devicecoherent || queuefamilycoherent || workgroupcoherent ||
                     subgroupcoherent || shadercallcoherent;
@@ -639,6 +640,7 @@ public:
             unsigned nonprivate : 1;
             unsigned volatil : 1;
             unsigned isImage : 1;
+            unsigned nonUniform : 1;
 
             void clear() {
                 coherent = 0;
@@ -650,6 +652,7 @@ public:
                 nonprivate = 0;
                 volatil = 0;
                 isImage = 0;
+                nonUniform = 0;
             }
 
             CoherentFlags operator |=(const CoherentFlags &other) {
@@ -662,6 +665,7 @@ public:
                 nonprivate |= other.nonprivate;
                 volatil |= other.volatil;
                 isImage |= other.isImage;
+                nonUniform |= other.nonUniform;
                 return *this;
             }
 #endif
@@ -727,7 +731,7 @@ public:
         spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0);
 
     // use accessChain and swizzle to load an r-value
-    Id accessChainLoad(Decoration precision, Decoration nonUniform, Id ResultType,
+    Id accessChainLoad(Decoration precision, Decoration l_nonUniform, Decoration r_nonUniform, Id ResultType,
         spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, spv::Scope scope = spv::ScopeMax,
             unsigned int alignment = 0);
 

Test/baseResults/spv.nonuniform4.frag.out

@@ -23,6 +23,7 @@ spv.nonuniform4.frag
                               Decorate 13(rIndex) Flat
                               Decorate 13(rIndex) Location 3
                               Decorate 15 DecorationNonUniformEXT
+                              Decorate 17 DecorationNonUniformEXT
                               Decorate 21 DecorationNonUniformEXT
                2:             TypeVoid
                3:             TypeFunction 2

Test/spv.nonuniform.frag

@@ -28,10 +28,12 @@ nonuniformEXT int foo(nonuniformEXT int nupi, nonuniformEXT out int f)
 void main()
 {
     nonuniformEXT int nu_li;
+    nonuniformEXT int nu_li2;
     int dyn_i;
 
     int a = foo(nu_li, nu_li);
     nu_li = nonuniformEXT(a) + nonuniformEXT(a * 2);
+    nu_li2 = a + nonuniformEXT(a * 2);
 
     float b;
     b = nu_inv4.x * nu_gf;
@@ -46,16 +48,25 @@ void main()
     b += texelFetch(uniformTexelBuffer[nu_ii], 1).x;
     b += imageLoad(storageTexelBuffer[nu_ii], 1).x;
     b += texture(sampler2D(uniformTexArr[nu_ii], uniformSampler), inTexcoord.xy).x;
+    b += texture(nonuniformEXT(sampler2D(uniformTexArr[nu_ii], uniformSampler)), inTexcoord.xy).x;
 
     nonuniformEXT ivec4 v;
     nonuniformEXT mat4 m;
     nonuniformEXT struct S { int a; } s;
+    nonuniformEXT int arr[10];
     ivec4 uv;
+    mat4 um;
+    struct US { int a[10]; } us;
+    int uarr[10];
     b += uniformBuffer[v.y].a;
     b += uniformBuffer[v[2]].a;
     b += uniformBuffer[uv[nu_ii]].a;
     b += uniformBuffer[int(m[2].z)].a;
     b += uniformBuffer[s.a].a;
+    b += uniformBuffer[arr[2]].a;
+    b += uniformBuffer[int(um[nu_ii].z)].a;
+    b += uniformBuffer[us.a[nu_ii]].a;
+    b += uniformBuffer[uarr[nu_ii]].a;
 
     storageBuffer[nu_ii].b = b;
 }