changes based on feedback

diff --git a/Test/baseResults/spv.320.meshShaderUserDefined.mesh.out b/Test/baseResults/spv.320.meshShaderUserDefined.mesh.out
new file mode 100644 (file)
index 0000000..ef39907
--- /dev/null
+++ b/Test/baseResults/spv.320.meshShaderUserDefined.mesh.out
@@ -0,0 +1,205 @@
+spv.meshShaderUserDefined.mesh
+// Module Version 10000
+// Generated by (magic number): 80007
+// Id's are bound by 140
+
+                              Capability MeshShadingNV
+                              Extension  "SPV_NV_mesh_shader"
+               1:             ExtInstImport  "GLSL.std.450"
+                              MemoryModel Logical GLSL450
+                              EntryPoint MeshNV 4  "main" 12 19 37 103
+                              ExecutionMode 4 LocalSize 32 1 1
+                              ExecutionMode 4 OutputVertices 81
+                              ExecutionMode 4 OutputPrimitivesNV 32
+                              ExecutionMode 4 OutputTrianglesNV
+                              Source ESSL 320
+                              SourceExtension  "GL_NV_mesh_shader"
+                              Name 4  "main"
+                              Name 8  "iid"
+                              Name 12  "gl_LocalInvocationID"
+                              Name 18  "gid"
+                              Name 19  "gl_WorkGroupID"
+                              Name 33  "myblock"
+                              MemberName 33(myblock) 0  "f"
+                              MemberName 33(myblock) 1  "fArr"
+                              MemberName 33(myblock) 2  "pos"
+                              MemberName 33(myblock) 3  "posArr"
+                              MemberName 33(myblock) 4  "m"
+                              MemberName 33(myblock) 5  "mArr"
+                              Name 37  "blk"
+                              Name 99  "myblock2"
+                              MemberName 99(myblock2) 0  "f"
+                              MemberName 99(myblock2) 1  "pos"
+                              MemberName 99(myblock2) 2  "m"
+                              Name 103  "blk2"
+                              Decorate 12(gl_LocalInvocationID) BuiltIn LocalInvocationId
+                              Decorate 19(gl_WorkGroupID) BuiltIn WorkgroupId
+                              MemberDecorate 33(myblock) 0 PerPrimitiveNV
+                              MemberDecorate 33(myblock) 1 PerPrimitiveNV
+                              MemberDecorate 33(myblock) 2 PerPrimitiveNV
+                              MemberDecorate 33(myblock) 3 PerPrimitiveNV
+                              MemberDecorate 33(myblock) 4 PerPrimitiveNV
+                              MemberDecorate 33(myblock) 5 PerPrimitiveNV
+                              Decorate 33(myblock) Block
+                              Decorate 37(blk) Location 0
+                              Decorate 99(myblock2) Block
+                              Decorate 103(blk2) Location 20
+                              Decorate 139 BuiltIn WorkgroupSize
+               2:             TypeVoid
+               3:             TypeFunction 2
+               6:             TypeInt 32 1
+               7:             TypePointer Function 6(int)
+               9:             TypeInt 32 0
+              10:             TypeVector 9(int) 3
+              11:             TypePointer Input 10(ivec3)
+12(gl_LocalInvocationID):     11(ptr) Variable Input
+              13:      9(int) Constant 0
+              14:             TypePointer Input 9(int)
+19(gl_WorkGroupID):     11(ptr) Variable Input
+              23:             TypeFloat 32
+              24:      9(int) Constant 4
+              25:             TypeArray 23(float) 24
+              26:             TypeVector 23(float) 3
+              27:             TypeVector 23(float) 4
+              28:             TypeArray 27(fvec4) 24
+              29:             TypeMatrix 27(fvec4) 4
+              30:             TypeMatrix 26(fvec3) 3
+              31:      9(int) Constant 2
+              32:             TypeArray 30 31
+     33(myblock):             TypeStruct 23(float) 25 26(fvec3) 28 29 32
+              34:      9(int) Constant 32
+              35:             TypeArray 33(myblock) 34
+              36:             TypePointer Output 35
+         37(blk):     36(ptr) Variable Output
+              39:      6(int) Constant 0
+              40:   23(float) Constant 1093664768
+              41:             TypePointer Output 23(float)
+              44:      6(int) Constant 1
+              52:      6(int) Constant 2
+              54:   23(float) Constant 1096810496
+              55:   23(float) Constant 1097859072
+              56:   23(float) Constant 1095761920
+              57:   26(fvec3) ConstantComposite 54 55 56
+              58:             TypePointer Output 26(fvec3)
+              64:      6(int) Constant 3
+              69:             TypePointer Output 27(fvec4)
+              74:      6(int) Constant 4
+              76:   23(float) Constant 1098907648
+              77:   27(fvec4) ConstantComposite 56 54 55 76
+              82:      6(int) Constant 5
+              85:      9(int) Constant 3
+              88:      9(int) Constant 1
+              93:   23(float) Constant 1099431936
+              94:   23(float) Constant 1099956224
+              95:   23(float) Constant 1100480512
+              96:   26(fvec3) ConstantComposite 93 94 95
+              98:      9(int) Constant 264
+    99(myblock2):             TypeStruct 23(float) 27(fvec4) 29
+             100:      9(int) Constant 81
+             101:             TypeArray 99(myblock2) 100
+             102:             TypePointer Output 101
+       103(blk2):    102(ptr) Variable Output
+             109:   23(float) Constant 1101004800
+             113:   23(float) Constant 1101529088
+             114:   23(float) Constant 1102053376
+             115:   23(float) Constant 1102577664
+             116:   23(float) Constant 1103101952
+             117:   27(fvec4) ConstantComposite 113 114 115 116
+             129:   23(float) Constant 1105723392
+             139:   10(ivec3) ConstantComposite 34 88 88
+         4(main):           2 Function None 3
+               5:             Label
+          8(iid):      7(ptr) Variable Function
+         18(gid):      7(ptr) Variable Function
+              15:     14(ptr) AccessChain 12(gl_LocalInvocationID) 13
+              16:      9(int) Load 15
+              17:      6(int) Bitcast 16
+                              Store 8(iid) 17
+              20:     14(ptr) AccessChain 19(gl_WorkGroupID) 13
+              21:      9(int) Load 20
+              22:      6(int) Bitcast 21
+                              Store 18(gid) 22
+              38:      6(int) Load 8(iid)
+              42:     41(ptr) AccessChain 37(blk) 38 39
+                              Store 42 40
+              43:      6(int) Load 8(iid)
+              45:      6(int) IAdd 43 44
+              46:      6(int) Load 18(gid)
+              47:      6(int) Load 8(iid)
+              48:     41(ptr) AccessChain 37(blk) 47 39
+              49:   23(float) Load 48
+              50:     41(ptr) AccessChain 37(blk) 45 44 46
+                              Store 50 49
+              51:      6(int) Load 8(iid)
+              53:      6(int) SDiv 51 52
+              59:     58(ptr) AccessChain 37(blk) 53 52
+              60:   26(fvec3) Load 59
+              61:   26(fvec3) VectorShuffle 60 57 5 3 4
+                              Store 59 61
+              62:      6(int) Load 8(iid)
+              63:      6(int) IMul 62 52
+              65:      6(int) Load 8(iid)
+              66:      6(int) SDiv 65 52
+              67:     58(ptr) AccessChain 37(blk) 66 52
+              68:   26(fvec3) Load 67
+              70:     69(ptr) AccessChain 37(blk) 63 64 44
+              71:   27(fvec4) Load 70
+              72:   27(fvec4) VectorShuffle 71 68 0 4 5 6
+                              Store 70 72
+              73:      6(int) Load 8(iid)
+              75:      6(int) SDiv 73 74
+              78:     69(ptr) AccessChain 37(blk) 75 74 52
+              79:   27(fvec4) Load 78
+              80:   27(fvec4) VectorShuffle 79 77 7 6 5 4
+                              Store 78 80
+              81:      6(int) Load 8(iid)
+              83:      6(int) Load 8(iid)
+              84:      6(int) SDiv 83 74
+              86:     41(ptr) AccessChain 37(blk) 84 74 52 85
+              87:   23(float) Load 86
+              89:     41(ptr) AccessChain 37(blk) 81 82 39 44 88
+                              Store 89 87
+              90:      6(int) Load 8(iid)
+              91:      6(int) IMul 90 74
+              92:      6(int) Load 18(gid)
+              97:     58(ptr) AccessChain 37(blk) 91 82 44 92
+                              Store 97 96
+                              MemoryBarrier 88 98
+                              ControlBarrier 31 31 98
+             104:      6(int) Load 8(iid)
+             105:      6(int) Load 8(iid)
+             106:      6(int) ISub 105 44
+             107:     41(ptr) AccessChain 103(blk2) 106 39
+             108:   23(float) Load 107
+             110:   23(float) FAdd 108 109
+             111:     41(ptr) AccessChain 103(blk2) 104 39
+                              Store 111 110
+             112:      6(int) Load 8(iid)
+             118:     69(ptr) AccessChain 103(blk2) 112 44
+                              Store 118 117
+             119:      6(int) Load 8(iid)
+             120:      6(int) IAdd 119 44
+             121:      6(int) Load 18(gid)
+             122:      6(int) Load 8(iid)
+             123:     69(ptr) AccessChain 103(blk2) 122 44
+             124:   27(fvec4) Load 123
+             125:     69(ptr) AccessChain 103(blk2) 120 52 121
+                              Store 125 124
+             126:      6(int) Load 8(iid)
+             127:      6(int) IAdd 126 44
+             128:      6(int) Load 18(gid)
+             130:     41(ptr) AccessChain 103(blk2) 127 52 128 31
+                              Store 130 129
+             131:      6(int) Load 8(iid)
+             132:      6(int) IAdd 131 52
+             133:      6(int) Load 8(iid)
+             134:      6(int) IAdd 133 44
+             135:      6(int) Load 18(gid)
+             136:     69(ptr) AccessChain 103(blk2) 134 52 135
+             137:   27(fvec4) Load 136
+             138:     69(ptr) AccessChain 103(blk2) 132 52 64
+                              Store 138 137
+                              MemoryBarrier 88 98
+                              ControlBarrier 31 31 98
+                              Return
+                              FunctionEnd

diff --git a/Test/baseResults/spv.meshShaderUserDefined.mesh.out b/Test/baseResults/spv.meshShaderUserDefined.mesh.out
index ef39907..c3ec915 100644 (file)
--- a/Test/baseResults/spv.meshShaderUserDefined.mesh.out
+++ b/Test/baseResults/spv.meshShaderUserDefined.mesh.out
@@ -1,205 +1,203 @@
 spv.meshShaderUserDefined.mesh
 // Module Version 10000
 // Generated by (magic number): 80007
-// Id's are bound by 140
+// Id's are bound by 138
 
                               Capability MeshShadingNV
                               Extension  "SPV_NV_mesh_shader"
                1:             ExtInstImport  "GLSL.std.450"
                               MemoryModel Logical GLSL450
-                              EntryPoint MeshNV 4  "main" 12 19 37 103
+                              EntryPoint MeshNV 4  "main" 11 17 34 101
                               ExecutionMode 4 LocalSize 32 1 1
                               ExecutionMode 4 OutputVertices 81
                               ExecutionMode 4 OutputPrimitivesNV 32
                               ExecutionMode 4 OutputTrianglesNV
-                              Source ESSL 320
+                              Source GLSL 450
                               SourceExtension  "GL_NV_mesh_shader"
                               Name 4  "main"
                               Name 8  "iid"
-                              Name 12  "gl_LocalInvocationID"
-                              Name 18  "gid"
-                              Name 19  "gl_WorkGroupID"
-                              Name 33  "myblock"
-                              MemberName 33(myblock) 0  "f"
-                              MemberName 33(myblock) 1  "fArr"
-                              MemberName 33(myblock) 2  "pos"
-                              MemberName 33(myblock) 3  "posArr"
-                              MemberName 33(myblock) 4  "m"
-                              MemberName 33(myblock) 5  "mArr"
-                              Name 37  "blk"
-                              Name 99  "myblock2"
-                              MemberName 99(myblock2) 0  "f"
-                              MemberName 99(myblock2) 1  "pos"
-                              MemberName 99(myblock2) 2  "m"
-                              Name 103  "blk2"
-                              Decorate 12(gl_LocalInvocationID) BuiltIn LocalInvocationId
-                              Decorate 19(gl_WorkGroupID) BuiltIn WorkgroupId
-                              MemberDecorate 33(myblock) 0 PerPrimitiveNV
-                              MemberDecorate 33(myblock) 1 PerPrimitiveNV
-                              MemberDecorate 33(myblock) 2 PerPrimitiveNV
-                              MemberDecorate 33(myblock) 3 PerPrimitiveNV
-                              MemberDecorate 33(myblock) 4 PerPrimitiveNV
-                              MemberDecorate 33(myblock) 5 PerPrimitiveNV
-                              Decorate 33(myblock) Block
-                              Decorate 37(blk) Location 0
-                              Decorate 99(myblock2) Block
-                              Decorate 103(blk2) Location 20
-                              Decorate 139 BuiltIn WorkgroupSize
+                              Name 11  "gl_LocalInvocationID"
+                              Name 16  "gid"
+                              Name 17  "gl_WorkGroupID"
+                              Name 30  "myblock"
+                              MemberName 30(myblock) 0  "f"
+                              MemberName 30(myblock) 1  "fArr"
+                              MemberName 30(myblock) 2  "pos"
+                              MemberName 30(myblock) 3  "posArr"
+                              MemberName 30(myblock) 4  "m"
+                              MemberName 30(myblock) 5  "mArr"
+                              Name 34  "blk"
+                              Name 97  "myblock2"
+                              MemberName 97(myblock2) 0  "f"
+                              MemberName 97(myblock2) 1  "pos"
+                              MemberName 97(myblock2) 2  "m"
+                              Name 101  "blk2"
+                              Decorate 11(gl_LocalInvocationID) BuiltIn LocalInvocationId
+                              Decorate 17(gl_WorkGroupID) BuiltIn WorkgroupId
+                              MemberDecorate 30(myblock) 0 PerPrimitiveNV
+                              MemberDecorate 30(myblock) 1 PerPrimitiveNV
+                              MemberDecorate 30(myblock) 2 PerPrimitiveNV
+                              MemberDecorate 30(myblock) 3 PerPrimitiveNV
+                              MemberDecorate 30(myblock) 4 PerPrimitiveNV
+                              MemberDecorate 30(myblock) 5 PerPrimitiveNV
+                              Decorate 30(myblock) Block
+                              Decorate 34(blk) Location 0
+                              Decorate 97(myblock2) Block
+                              Decorate 101(blk2) Location 20
+                              Decorate 137 BuiltIn WorkgroupSize
                2:             TypeVoid
                3:             TypeFunction 2
-               6:             TypeInt 32 1
+               6:             TypeInt 32 0
                7:             TypePointer Function 6(int)
-               9:             TypeInt 32 0
-              10:             TypeVector 9(int) 3
-              11:             TypePointer Input 10(ivec3)
-12(gl_LocalInvocationID):     11(ptr) Variable Input
-              13:      9(int) Constant 0
-              14:             TypePointer Input 9(int)
-19(gl_WorkGroupID):     11(ptr) Variable Input
-              23:             TypeFloat 32
-              24:      9(int) Constant 4
-              25:             TypeArray 23(float) 24
-              26:             TypeVector 23(float) 3
-              27:             TypeVector 23(float) 4
-              28:             TypeArray 27(fvec4) 24
-              29:             TypeMatrix 27(fvec4) 4
-              30:             TypeMatrix 26(fvec3) 3
-              31:      9(int) Constant 2
-              32:             TypeArray 30 31
-     33(myblock):             TypeStruct 23(float) 25 26(fvec3) 28 29 32
-              34:      9(int) Constant 32
-              35:             TypeArray 33(myblock) 34
-              36:             TypePointer Output 35
-         37(blk):     36(ptr) Variable Output
-              39:      6(int) Constant 0
-              40:   23(float) Constant 1093664768
-              41:             TypePointer Output 23(float)
-              44:      6(int) Constant 1
-              52:      6(int) Constant 2
-              54:   23(float) Constant 1096810496
-              55:   23(float) Constant 1097859072
-              56:   23(float) Constant 1095761920
-              57:   26(fvec3) ConstantComposite 54 55 56
-              58:             TypePointer Output 26(fvec3)
-              64:      6(int) Constant 3
-              69:             TypePointer Output 27(fvec4)
-              74:      6(int) Constant 4
-              76:   23(float) Constant 1098907648
-              77:   27(fvec4) ConstantComposite 56 54 55 76
-              82:      6(int) Constant 5
-              85:      9(int) Constant 3
-              88:      9(int) Constant 1
-              93:   23(float) Constant 1099431936
-              94:   23(float) Constant 1099956224
-              95:   23(float) Constant 1100480512
-              96:   26(fvec3) ConstantComposite 93 94 95
-              98:      9(int) Constant 264
-    99(myblock2):             TypeStruct 23(float) 27(fvec4) 29
-             100:      9(int) Constant 81
-             101:             TypeArray 99(myblock2) 100
-             102:             TypePointer Output 101
-       103(blk2):    102(ptr) Variable Output
-             109:   23(float) Constant 1101004800
-             113:   23(float) Constant 1101529088
-             114:   23(float) Constant 1102053376
-             115:   23(float) Constant 1102577664
-             116:   23(float) Constant 1103101952
-             117:   27(fvec4) ConstantComposite 113 114 115 116
-             129:   23(float) Constant 1105723392
-             139:   10(ivec3) ConstantComposite 34 88 88
+               9:             TypeVector 6(int) 3
+              10:             TypePointer Input 9(ivec3)
+11(gl_LocalInvocationID):     10(ptr) Variable Input
+              12:      6(int) Constant 0
+              13:             TypePointer Input 6(int)
+17(gl_WorkGroupID):     10(ptr) Variable Input
+              20:             TypeFloat 32
+              21:      6(int) Constant 4
+              22:             TypeArray 20(float) 21
+              23:             TypeVector 20(float) 3
+              24:             TypeVector 20(float) 4
+              25:             TypeArray 24(fvec4) 21
+              26:             TypeMatrix 24(fvec4) 4
+              27:             TypeMatrix 23(fvec3) 3
+              28:      6(int) Constant 2
+              29:             TypeArray 27 28
+     30(myblock):             TypeStruct 20(float) 22 23(fvec3) 25 26 29
+              31:      6(int) Constant 32
+              32:             TypeArray 30(myblock) 31
+              33:             TypePointer Output 32
+         34(blk):     33(ptr) Variable Output
+              36:             TypeInt 32 1
+              37:     36(int) Constant 0
+              38:   20(float) Constant 1093664768
+              39:             TypePointer Output 20(float)
+              42:      6(int) Constant 1
+              44:     36(int) Constant 1
+              52:     36(int) Constant 2
+              53:   20(float) Constant 1096810496
+              54:   20(float) Constant 1097859072
+              55:   20(float) Constant 1095761920
+              56:   23(fvec3) ConstantComposite 53 54 55
+              57:             TypePointer Output 23(fvec3)
+              63:     36(int) Constant 3
+              68:             TypePointer Output 24(fvec4)
+              74:     36(int) Constant 4
+              75:   20(float) Constant 1098907648
+              76:   24(fvec4) ConstantComposite 55 53 54 75
+              81:     36(int) Constant 5
+              84:      6(int) Constant 3
+              91:   20(float) Constant 1099431936
+              92:   20(float) Constant 1099956224
+              93:   20(float) Constant 1100480512
+              94:   23(fvec3) ConstantComposite 91 92 93
+              96:      6(int) Constant 264
+    97(myblock2):             TypeStruct 20(float) 24(fvec4) 26
+              98:      6(int) Constant 81
+              99:             TypeArray 97(myblock2) 98
+             100:             TypePointer Output 99
+       101(blk2):    100(ptr) Variable Output
+             107:   20(float) Constant 1101004800
+             111:   20(float) Constant 1101529088
+             112:   20(float) Constant 1102053376
+             113:   20(float) Constant 1102577664
+             114:   20(float) Constant 1103101952
+             115:   24(fvec4) ConstantComposite 111 112 113 114
+             127:   20(float) Constant 1105723392
+             137:    9(ivec3) ConstantComposite 31 42 42
          4(main):           2 Function None 3
                5:             Label
           8(iid):      7(ptr) Variable Function
-         18(gid):      7(ptr) Variable Function
-              15:     14(ptr) AccessChain 12(gl_LocalInvocationID) 13
-              16:      9(int) Load 15
-              17:      6(int) Bitcast 16
-                              Store 8(iid) 17
-              20:     14(ptr) AccessChain 19(gl_WorkGroupID) 13
-              21:      9(int) Load 20
-              22:      6(int) Bitcast 21
-                              Store 18(gid) 22
-              38:      6(int) Load 8(iid)
-              42:     41(ptr) AccessChain 37(blk) 38 39
-                              Store 42 40
-              43:      6(int) Load 8(iid)
-              45:      6(int) IAdd 43 44
-              46:      6(int) Load 18(gid)
-              47:      6(int) Load 8(iid)
-              48:     41(ptr) AccessChain 37(blk) 47 39
-              49:   23(float) Load 48
-              50:     41(ptr) AccessChain 37(blk) 45 44 46
-                              Store 50 49
-              51:      6(int) Load 8(iid)
-              53:      6(int) SDiv 51 52
-              59:     58(ptr) AccessChain 37(blk) 53 52
-              60:   26(fvec3) Load 59
-              61:   26(fvec3) VectorShuffle 60 57 5 3 4
-                              Store 59 61
-              62:      6(int) Load 8(iid)
-              63:      6(int) IMul 62 52
-              65:      6(int) Load 8(iid)
-              66:      6(int) SDiv 65 52
-              67:     58(ptr) AccessChain 37(blk) 66 52
-              68:   26(fvec3) Load 67
-              70:     69(ptr) AccessChain 37(blk) 63 64 44
-              71:   27(fvec4) Load 70
-              72:   27(fvec4) VectorShuffle 71 68 0 4 5 6
-                              Store 70 72
-              73:      6(int) Load 8(iid)
-              75:      6(int) SDiv 73 74
-              78:     69(ptr) AccessChain 37(blk) 75 74 52
-              79:   27(fvec4) Load 78
-              80:   27(fvec4) VectorShuffle 79 77 7 6 5 4
-                              Store 78 80
-              81:      6(int) Load 8(iid)
-              83:      6(int) Load 8(iid)
-              84:      6(int) SDiv 83 74
-              86:     41(ptr) AccessChain 37(blk) 84 74 52 85
-              87:   23(float) Load 86
-              89:     41(ptr) AccessChain 37(blk) 81 82 39 44 88
-                              Store 89 87
-              90:      6(int) Load 8(iid)
-              91:      6(int) IMul 90 74
-              92:      6(int) Load 18(gid)
-              97:     58(ptr) AccessChain 37(blk) 91 82 44 92
-                              Store 97 96
-                              MemoryBarrier 88 98
-                              ControlBarrier 31 31 98
-             104:      6(int) Load 8(iid)
-             105:      6(int) Load 8(iid)
-             106:      6(int) ISub 105 44
-             107:     41(ptr) AccessChain 103(blk2) 106 39
-             108:   23(float) Load 107
-             110:   23(float) FAdd 108 109
-             111:     41(ptr) AccessChain 103(blk2) 104 39
-                              Store 111 110
-             112:      6(int) Load 8(iid)
-             118:     69(ptr) AccessChain 103(blk2) 112 44
-                              Store 118 117
-             119:      6(int) Load 8(iid)
-             120:      6(int) IAdd 119 44
-             121:      6(int) Load 18(gid)
-             122:      6(int) Load 8(iid)
-             123:     69(ptr) AccessChain 103(blk2) 122 44
-             124:   27(fvec4) Load 123
-             125:     69(ptr) AccessChain 103(blk2) 120 52 121
-                              Store 125 124
-             126:      6(int) Load 8(iid)
-             127:      6(int) IAdd 126 44
-             128:      6(int) Load 18(gid)
-             130:     41(ptr) AccessChain 103(blk2) 127 52 128 31
-                              Store 130 129
+         16(gid):      7(ptr) Variable Function
+              14:     13(ptr) AccessChain 11(gl_LocalInvocationID) 12
+              15:      6(int) Load 14
+                              Store 8(iid) 15
+              18:     13(ptr) AccessChain 17(gl_WorkGroupID) 12
+              19:      6(int) Load 18
+                              Store 16(gid) 19
+              35:      6(int) Load 8(iid)
+              40:     39(ptr) AccessChain 34(blk) 35 37
+                              Store 40 38
+              41:      6(int) Load 8(iid)
+              43:      6(int) IAdd 41 42
+              45:      6(int) Load 16(gid)
+              46:      6(int) Load 8(iid)
+              47:     39(ptr) AccessChain 34(blk) 46 37
+              48:   20(float) Load 47
+              49:     39(ptr) AccessChain 34(blk) 43 44 45
+                              Store 49 48
+              50:      6(int) Load 8(iid)
+              51:      6(int) UDiv 50 28
+              58:     57(ptr) AccessChain 34(blk) 51 52
+              59:   23(fvec3) Load 58
+              60:   23(fvec3) VectorShuffle 59 56 5 3 4
+                              Store 58 60
+              61:      6(int) Load 8(iid)
+              62:      6(int) IMul 61 28
+              64:      6(int) Load 8(iid)
+              65:      6(int) UDiv 64 28
+              66:     57(ptr) AccessChain 34(blk) 65 52
+              67:   23(fvec3) Load 66
+              69:     68(ptr) AccessChain 34(blk) 62 63 44
+              70:   24(fvec4) Load 69
+              71:   24(fvec4) VectorShuffle 70 67 0 4 5 6
+                              Store 69 71
+              72:      6(int) Load 8(iid)
+              73:      6(int) UDiv 72 21
+              77:     68(ptr) AccessChain 34(blk) 73 74 52
+              78:   24(fvec4) Load 77
+              79:   24(fvec4) VectorShuffle 78 76 7 6 5 4
+                              Store 77 79
+              80:      6(int) Load 8(iid)
+              82:      6(int) Load 8(iid)
+              83:      6(int) UDiv 82 21
+              85:     39(ptr) AccessChain 34(blk) 83 74 52 84
+              86:   20(float) Load 85
+              87:     39(ptr) AccessChain 34(blk) 80 81 37 44 42
+                              Store 87 86
+              88:      6(int) Load 8(iid)
+              89:      6(int) IMul 88 21
+              90:      6(int) Load 16(gid)
+              95:     57(ptr) AccessChain 34(blk) 89 81 44 90
+                              Store 95 94
+                              MemoryBarrier 42 96
+                              ControlBarrier 28 28 96
+             102:      6(int) Load 8(iid)
+             103:      6(int) Load 8(iid)
+             104:      6(int) ISub 103 42
+             105:     39(ptr) AccessChain 101(blk2) 104 37
+             106:   20(float) Load 105
+             108:   20(float) FAdd 106 107
+             109:     39(ptr) AccessChain 101(blk2) 102 37
+                              Store 109 108
+             110:      6(int) Load 8(iid)
+             116:     68(ptr) AccessChain 101(blk2) 110 44
+                              Store 116 115
+             117:      6(int) Load 8(iid)
+             118:      6(int) IAdd 117 42
+             119:      6(int) Load 16(gid)
+             120:      6(int) Load 8(iid)
+             121:     68(ptr) AccessChain 101(blk2) 120 44
+             122:   24(fvec4) Load 121
+             123:     68(ptr) AccessChain 101(blk2) 118 52 119
+                              Store 123 122
+             124:      6(int) Load 8(iid)
+             125:      6(int) IAdd 124 42
+             126:      6(int) Load 16(gid)
+             128:     39(ptr) AccessChain 101(blk2) 125 52 126 28
+                              Store 128 127
+             129:      6(int) Load 8(iid)
+             130:      6(int) IAdd 129 28
              131:      6(int) Load 8(iid)
-             132:      6(int) IAdd 131 52
-             133:      6(int) Load 8(iid)
-             134:      6(int) IAdd 133 44
-             135:      6(int) Load 18(gid)
-             136:     69(ptr) AccessChain 103(blk2) 134 52 135
-             137:   27(fvec4) Load 136
-             138:     69(ptr) AccessChain 103(blk2) 132 52 64
-                              Store 138 137
-                              MemoryBarrier 88 98
-                              ControlBarrier 31 31 98
+             132:      6(int) IAdd 131 42
+             133:      6(int) Load 16(gid)
+             134:     68(ptr) AccessChain 101(blk2) 132 52 133
+             135:   24(fvec4) Load 134
+             136:     68(ptr) AccessChain 101(blk2) 130 52 63
+                              Store 136 135
+                              MemoryBarrier 42 96
+                              ControlBarrier 28 28 96
                               Return
                               FunctionEnd

diff --git a/Test/spv.320.meshShaderUserDefined.mesh b/Test/spv.320.meshShaderUserDefined.mesh
new file mode 100644 (file)
index 0000000..60a1452
--- /dev/null
+++ b/Test/spv.320.meshShaderUserDefined.mesh
@@ -0,0 +1,59 @@
+#version 320 es
+
+#define MAX_VER  81
+#define MAX_PRIM 32
+
+#define BARRIER() \
+    memoryBarrierShared(); \
+    barrier();
+
+#extension GL_NV_mesh_shader : enable
+
+layout(local_size_x = 32) in;
+
+layout(max_vertices=MAX_VER) out;
+layout(max_primitives=MAX_PRIM) out;
+layout(triangles) out;
+
+// test use of user defined interface out blocks:
+
+// per-primitive block
+perprimitiveNV layout(location=0) out myblock {
+    float f;
+    float fArr[4];
+    vec3 pos;
+    vec4 posArr[4];
+    mat4 m;
+    mat3 mArr[2];
+} blk[];
+
+// per-vertex block
+layout(location=20) out myblock2 {
+    float f;
+    vec4 pos;
+    mat4 m;
+} blk2[];
+
+void main()
+{
+    int iid = int(gl_LocalInvocationID.x);
+    int gid = int(gl_WorkGroupID.x);
+
+    blk[iid].f               = 11.0;
+    blk[iid+1].fArr[gid]     = blk[iid].f;
+    blk[iid/2].pos.yzx       = vec3(14.0, 15.0, 13.0);
+    blk[iid*2].posArr[1].yzw = blk[iid/2].pos;
+    blk[iid/4].m[2].wzyx     = vec4(13.0, 14.0, 15.0, 16.0);
+    blk[iid].mArr[0][1][1]   = blk[iid/4].m[2].w;
+    blk[iid*4].mArr[1][gid]  = vec3(17.0, 18.0, 19.0);
+
+    BARRIER();
+
+    blk2[iid].f           = blk2[iid-1].f + 20.0;
+    blk2[iid].pos         = vec4(21.0, 22.0, 23.0, 24.0);
+    blk2[iid+1].m[gid]    = blk2[iid].pos;
+    blk2[iid+1].m[gid][2] = 29.0;
+    blk2[iid+2].m[3]      = blk2[iid+1].m[gid];
+
+    BARRIER();
+}

diff --git a/Test/spv.meshShaderUserDefined.mesh b/Test/spv.meshShaderUserDefined.mesh
index 60a1452..8b0937d 100644 (file)
--- a/Test/spv.meshShaderUserDefined.mesh
+++ b/Test/spv.meshShaderUserDefined.mesh
@@ -1,4 +1,4 @@
-#version 320 es
+#version 450
 
 #define MAX_VER  81
 #define MAX_PRIM 32
@@ -36,8 +36,8 @@ layout(location=20) out myblock2 {
 
 void main()
 {
-    int iid = int(gl_LocalInvocationID.x);
-    int gid = int(gl_WorkGroupID.x);
+    uint iid = gl_LocalInvocationID.x;
+    uint gid = gl_WorkGroupID.x;
 
     blk[iid].f               = 11.0;
     blk[iid+1].fArr[gid]     = blk[iid].f;

diff --git a/glslang/MachineIndependent/ParseHelper.cpp b/glslang/MachineIndependent/ParseHelper.cpp
index 3a0fa43..9f9e5c5 100644 (file)
--- a/glslang/MachineIndependent/ParseHelper.cpp
+++ b/glslang/MachineIndependent/ParseHelper.cpp
@@ -3606,6 +3606,7 @@ void TParseContext::arraySizesCheck(const TSourceLoc& loc, const TQualifier& qua
             if ((profile == EEsProfile && version >= 320) ||
                 extensionTurnedOn(E_GL_NV_mesh_shader))
                 return;
+        break;
 #endif
     default:
         break;

diff --git a/glslang/MachineIndependent/Versions.cpp b/glslang/MachineIndependent/Versions.cpp
index 4d7dd8b..6182f81 100644 (file)
--- a/glslang/MachineIndependent/Versions.cpp
+++ b/glslang/MachineIndependent/Versions.cpp
@@ -835,8 +835,8 @@ void TParseVersions::updateExtensionBehavior(const char* extension, TExtensionBe
 }
 
 #ifdef NV_EXTENSIONS
-// Validate if extension name is used with correct shader stage.
-bool TParseVersions::validateExtensionName(const TSourceLoc& loc, const char * const extension)
+// Check if extension is used with correct shader stage.
+bool TParseVersions::checkShaderStageForNVExtensions(const TSourceLoc& loc, const char * const extension)
 {
     int lNumErrors = getNumErrors();
 
@@ -845,6 +845,8 @@ bool TParseVersions::validateExtensionName(const TSourceLoc& loc, const char * c
         requireStage(loc, (EShLanguageMask)(EShLangTaskNVMask | EShLangMeshNVMask),
                      "#extension GL_NV_mesh_shader");
 
+    // TODO: need to add error checks for other nvidia turing extensions
+
     if (getNumErrors() > lNumErrors)
         return false;
     return true;

diff --git a/glslang/MachineIndependent/parseVersions.h b/glslang/MachineIndependent/parseVersions.h
index 3d36e50..4504803 100644 (file)
--- a/glslang/MachineIndependent/parseVersions.h
+++ b/glslang/MachineIndependent/parseVersions.h
@@ -104,7 +104,7 @@ public:
     virtual bool checkExtensionsRequested(const TSourceLoc&, int numExtensions, const char* const extensions[], const char* featureDesc);
     virtual void updateExtensionBehavior(const char* const extension, TExtensionBehavior);
 #ifdef NV_EXTENSIONS
-    virtual bool validateExtensionName(const TSourceLoc&, const char* const extension);
+    virtual bool checkShaderStageForNVExtensions(const TSourceLoc&, const char* const extension);
 #endif
 
     virtual void C_DECL error(const TSourceLoc&, const char* szReason, const char* szToken,

diff --git a/glslang/MachineIndependent/preprocessor/Pp.cpp b/glslang/MachineIndependent/preprocessor/Pp.cpp
index efc81fd..7f39c34 100644 (file)
--- a/glslang/MachineIndependent/preprocessor/Pp.cpp
+++ b/glslang/MachineIndependent/preprocessor/Pp.cpp
@@ -875,7 +875,7 @@ int TPpContext::CPPextension(TPpToken* ppToken)
     }
 
 #ifdef NV_EXTENSIONS
-    if (!parseContext.validateExtensionName(ppToken->loc, extensionName))
+    if (!parseContext.checkShaderStageForNVExtensions(ppToken->loc, extensionName))
         return token;
 #endif