Yet more definitions moved into the stdlib (#1263)

The only big catch that I ran into with this batch was that I found the `float.getPi()` function was being emitted to the output GLSL even when that function wasn't being used. This seems to have been a latent problem in the earlier PR, but was only surfaced in the tests once a Slang->GLSL test started using another intrinsic that led to the `float : __BuiltinFloatingPointType` witness table being live in the IR.

The fix for the gotcha here was to add a late IR pass that basically empties out all witness tables in the IR, so that functions that are only referenced by witness tables can then be removed as dead code. This pass is something we should *not* apply if/when we start supporting real dynamic dispatch through witness tables, but that is a problem to be solved on another day.

The remaining tricky pieces of this change were:

* Needed to remember to mark functions as target intrinsics on HLSL and/or GLSL as appropriate (hopefully I caught all the cases) so they don't get emitted as source there.

* The `msad4` function in HLSL is very poorly documented, so filling in its definition was tricky. I made my best effort based on how it is described on MSDN, but it is likely that if anybody wants to rely on this function they will need us to vet our results with some tests.

6 files changed, 211 insertions, 55 deletions

diff --git a/source/slang/hlsl.meta.slang b/source/slang/hlsl.meta.slang
index 572b64b21..d9e40dd4f 100644
--- a/source/slang/hlsl.meta.slang
+++ b/source/slang/hlsl.meta.slang
@@ -1199,9 +1199,24 @@ matrix<T, N, M> floor(matrix<T, N, M> x)
 }
 
 // Fused multiply-add for doubles
+__target_intrinsic(hlsl)
+__target_intrinsic(glsl)
 double fma(double a, double b, double c);
-__generic<let N : int> vector<double, N> fma(vector<double, N> a, vector<double, N> b, vector<double, N> c);
-__generic<let N : int, let M : int> matrix<double,N,M> fma(matrix<double,N,M> a, matrix<double,N,M> b, matrix<double,N,M> c);
+
+__generic<let N : int>
+__target_intrinsic(hlsl)
+__target_intrinsic(glsl)
+vector<double, N> fma(vector<double, N> a, vector<double, N> b, vector<double, N> c)
+{
+    VECTOR_MAP_TRINARY(double, N, fma, a, b, c);
+}
+
+__generic<let N : int, let M : int>
+__target_intrinsic(hlsl)
+matrix<double, N, M> fma(matrix<double, N, M> a, matrix<double, N, M> b, matrix<double, N, M> c)
+{
+    MATRIX_MAP_TRINARY(double, N, M, fma, a, b, c);
+}
 
 // Floating point remainder of x/y
 __generic<T : __BuiltinFloatingPointType>
@@ -1425,7 +1440,6 @@ __generic<T : __BuiltinFloatingPointType>
 __target_intrinsic(hlsl)
 __target_intrinsic(cpu)
 __target_intrinsic(cuda)
-//__target_intrinsic(glsl, "(!(isinf($0) || isnan($0)))")
 bool isfinite(T x)
 {
     return !(isinf(x) || isnan(x));
@@ -1433,7 +1447,6 @@ bool isfinite(T x)
 
 __generic<T : __BuiltinFloatingPointType, let N : int>
 __target_intrinsic(hlsl)
-//__target_intrinsic(glsl, "(!(isinf($0) || isnan($0)))")
 vector<bool, N> isfinite(vector<T, N> x)
 {
     VECTOR_MAP_UNARY(bool, N, isfinite, x);
@@ -1488,18 +1501,16 @@ matrix<bool, N, M> isnan(matrix<T, N, M> x)
 
 __generic<T : __BuiltinFloatingPointType>
 __target_intrinsic(hlsl)
-__target_intrinsic(glsl, "($0 * pow(2.0f, $1))")
-T ldexp(T x, T exp);
-/*{
+T ldexp(T x, T exp)
+{
     return x * exp2(exp);
-}*/
+}
 
 __generic<T : __BuiltinFloatingPointType, let N : int>
 __target_intrinsic(hlsl)
-__target_intrinsic(glsl, "($0 * pow(2.0f, $1))")
 vector<T, N> ldexp(vector<T, N> x, vector<T, N> exp)
 {
-    VECTOR_MAP_BINARY(T, N, ldexp, x, exp);
+    return x * exp2(exp);
 }
 
 __generic<T : __BuiltinFloatingPointType, let N : int, let M : int>
@@ -1522,17 +1533,17 @@ T length(vector<T, N> x)
 __generic<T : __BuiltinFloatingPointType>
 __target_intrinsic(hlsl)
 __target_intrinsic(glsl, mix)
-T lerp(T x, T y, T s);
-/*{
-    return x * (1 - s) + y * s;
-}*/
+T lerp(T x, T y, T s)
+{
+    return x * (T(1.0f) - s) + y * s;
+}
 
 __generic<T : __BuiltinFloatingPointType, let N : int>
 __target_intrinsic(hlsl)
 __target_intrinsic(glsl, mix)
 vector<T, N> lerp(vector<T, N> x, vector<T, N> y, vector<T, N> s)
 {
-    VECTOR_MAP_TRINARY(T, N, lerp, x, y, s);
+    return x * (T(1.0f) - s) + y * s;
 }
 
 __generic<T : __BuiltinFloatingPointType, let N : int, let M : int>
@@ -1543,7 +1554,14 @@ matrix<T,N,M> lerp(matrix<T,N,M> x, matrix<T,N,M> y, matrix<T,N,M> s)
 }
 
 // Legacy lighting function (obsolete)
-float4 lit(float n_dot_l, float n_dot_h, float m);
+__target_intrinsic(hlsl)
+float4 lit(float n_dot_l, float n_dot_h, float m)
+{
+    let ambient = 1.0f;
+    let diffuse = max(n_dot_l, 0.0f);
+    let specular = step(0.0f, n_dot_l) * max(n_dot_h * m, 0.0f);
+    return float4(ambient, diffuse, specular, 1.0f);
+}
 
 // Base-e logarithm
 __generic<T : __BuiltinFloatingPointType>
@@ -1606,14 +1624,25 @@ matrix<T,N,M> log2(matrix<T,N,M> x)
 
 // multiply-add
 
+__generic<T : __BuiltinArithmeticType>
+__target_intrinsic(hlsl)
 __target_intrinsic(glsl, fma)
-__generic<T : __BuiltinArithmeticType> T mad(T mvalue, T avalue, T bvalue);
+T mad(T mvalue, T avalue, T bvalue);
 
+__generic<T : __BuiltinArithmeticType, let N : int>
+__target_intrinsic(hlsl)
 __target_intrinsic(glsl, fma)
-__generic<T : __BuiltinArithmeticType, let N : int> vector<T,N> mad(vector<T,N> mvalue, vector<T,N> avalue, vector<T,N> bvalue);
+vector<T, N> mad(vector<T, N> mvalue, vector<T, N> avalue, vector<T, N> bvalue)
+{
+    VECTOR_MAP_TRINARY(T, N, mad, mvalue, avalue, bvalue);
+}
 
-__target_intrinsic(glsl, fma)
-__generic<T : __BuiltinArithmeticType, let N : int, let M : int> matrix<T,N,M> mad(matrix<T,N,M> mvalue, matrix<T,N,M> avalue, matrix<T,N,M> bvalue);
+__generic<T : __BuiltinArithmeticType, let N : int, let M : int>
+__target_intrinsic(hlsl)
+matrix<T, N, M> mad(matrix<T, N, M> mvalue, matrix<T, N, M> avalue, matrix<T, N, M> bvalue)
+{
+    MATRIX_MAP_TRINARY(T, N, M, mad, mvalue, avalue, bvalue);
+}
 
 // maximum
 __generic<T : __BuiltinArithmeticType>
@@ -1677,32 +1706,79 @@ matrix<T,N,M> modf(matrix<T,N,M> x, out matrix<T,N,M> ip)
 }
 
 // msad4 (whatever that is)
-uint4 msad4(uint reference, uint2 source, uint4 accum);
+__target_intrinsic(hlsl)
+uint4 msad4(uint reference, uint2 source, uint4 accum)
+{
+    int4 bytesRef = (reference >> uint4(24, 16, 8, 0)) & 0xFF;
+    int4 bytesX   = (source.x  >> uint4(24, 16, 8, 0)) & 0xFF;
+    int4 bytesY   = (source.y  >> uint4(24, 16, 8, 0)) & 0xFF;
+
+    uint4 mask = bytesRef == 0 ? 0 : 0xFFFFFFFFu;
+
+    uint4 result = accum;
+    result += mask.x & abs(bytesRef - int4(bytesX.x,           bytesY.y, bytesY.z, bytesY.w));
+    result += mask.y & abs(bytesRef - int4(bytesX.x, bytesX.y,           bytesY.z, bytesY.w));
+    result += mask.z & abs(bytesRef - int4(bytesX.x, bytesX.y, bytesX.z,           bytesY.w));
+    result += mask.w & abs(bytesRef - int4(bytesX.x, bytesX.y, bytesX.z, bytesX.w));
+    return result;
+}
 
 // General inner products
 
 // scalar-scalar
-__generic<T : __BuiltinArithmeticType> T mul(T x, T y);
+__generic<T : __BuiltinArithmeticType>
+__intrinsic_op($(kIROp_Mul))
+T mul(T x, T y);
 
 // scalar-vector and vector-scalar
-__generic<T : __BuiltinArithmeticType, let N : int> vector<T,N> mul(vector<T,N> x, T y);
-__generic<T : __BuiltinArithmeticType, let N : int> vector<T,N> mul(T x, vector<T,N> y);
+__generic<T : __BuiltinArithmeticType, let N : int>
+__intrinsic_op($(kIROp_Mul))
+vector<T, N> mul(vector<T, N> x, T y);
+
+__generic<T : __BuiltinArithmeticType, let N : int>
+__intrinsic_op($(kIROp_Mul))
+vector<T, N> mul(T x, vector<T, N> y);
 
 // scalar-matrix and matrix-scalar
-__generic<T : __BuiltinArithmeticType, let N : int, let M :int> matrix<T,N,M> mul(matrix<T,N,M> x, T y);
-__generic<T : __BuiltinArithmeticType, let N : int, let M :int> matrix<T,N,M> mul(T x, matrix<T,N,M> y);
+__generic<T : __BuiltinArithmeticType, let N : int, let M :int>
+__intrinsic_op($(kIROp_Mul))
+matrix<T, N, M> mul(matrix<T, N, M> x, T y);
+
+__generic<T : __BuiltinArithmeticType, let N : int, let M :int>
+__intrinsic_op($(kIROp_Mul))
+matrix<T, N, M> mul(T x, matrix<T, N, M> y);
 
 // vector-vector (dot product)
-__generic<T : __BuiltinArithmeticType, let N : int> __intrinsic_op(dot) T mul(vector<T,N> x, vector<T,N> y);
+__generic<T : __BuiltinArithmeticType, let N : int>
+__target_intrinsic(hlsl)
+__target_intrinsic(glsl, "dot")
+T mul(vector<T, N> x, vector<T, N> y)
+{
+    return dot(x, y);
+}
+
+${{{{
+// TODO: The following functions could conceivably be defined
+// in the stdlib for the benefit of targets without direct
+// support for matrices, but the use of `__intrinsic_op` to
+// map them to a dedicated IR instruction interferes with
+// that choice.
+}}}}
 
 // vector-matrix
-__generic<T : __BuiltinArithmeticType, let N : int, let M : int> __intrinsic_op(mulVectorMatrix) vector<T,M> mul(vector<T,N> x, matrix<T,N,M> y);
+__generic<T : __BuiltinArithmeticType, let N : int, let M : int>
+__intrinsic_op(mulVectorMatrix)
+vector<T,M> mul(vector<T,N> x, matrix<T,N,M> y);
 
 // matrix-vector
-__generic<T : __BuiltinArithmeticType, let N : int, let M : int> __intrinsic_op(mulMatrixVector) vector<T,N> mul(matrix<T,N,M> x, vector<T,M> y);
+__generic<T : __BuiltinArithmeticType, let N : int, let M : int>
+__intrinsic_op(mulMatrixVector)
+vector<T,N> mul(matrix<T,N,M> x, vector<T,M> y);
 
 // matrix-matrix
-__generic<T : __BuiltinArithmeticType, let R : int, let N : int, let C : int> __intrinsic_op(mulMatrixMatrix) matrix<T,R,C> mul(matrix<T,R,N> x, matrix<T,N,C> y);
+__generic<T : __BuiltinArithmeticType, let R : int, let N : int, let C : int>
+__intrinsic_op(mulMatrixMatrix)
+matrix<T,R,C> mul(matrix<T,R,N> x, matrix<T,N,C> y);
 
 // noise (deprecated)
 
@@ -1753,10 +1829,13 @@ int NonUniformResourceIndex(int index)
 }
 
 // Normalize a vector
-__generic<T : __BuiltinFloatingPointType, let N : int> vector<T,N> normalize(vector<T,N> x);
-/*{
+__generic<T : __BuiltinFloatingPointType, let N : int>
+__target_intrinsic(hlsl)
+__target_intrinsic(glsl)
+vector<T,N> normalize(vector<T,N> x)
+{
     return x / length(x);
-}*/
+}
 
 // Raise to a power
 __generic<T : __BuiltinFloatingPointType>
@@ -1856,31 +1935,33 @@ void ProcessTriTessFactorsMin(
 __generic<T : __BuiltinFloatingPointType>
 __target_intrinsic(hlsl)
 __target_intrinsic(glsl)
-T radians(T x);
+T radians(T x)
+{
+    return x * (T.getPi() / T(180.0f));
+}
 
 __generic<T : __BuiltinFloatingPointType, let N : int>
 __target_intrinsic(hlsl)
 __target_intrinsic(glsl)
 vector<T, N> radians(vector<T, N> x)
 {
-    VECTOR_MAP_UNARY(T, N, radians, x);
+    return x * (T.getPi() / T(180.0f));
 }
 
 __generic<T : __BuiltinFloatingPointType, let N : int, let M : int>
 __target_intrinsic(hlsl)
 matrix<T, N, M> radians(matrix<T, N, M> x)
 {
-    MATRIX_MAP_UNARY(T, N, M, radians, x);
+    return x * (T.getPi() / T(180.0f));
 }
 
 // Approximate reciprocal
 __generic<T : __BuiltinFloatingPointType>
 __target_intrinsic(hlsl)
-__target_intrinsic(glsl, "1.0/($0)")
-T rcp(T x);
-/*{
-    return T(1) / x;
-}*/
+T rcp(T x)
+{
+    return T(1.0) / x;
+}
 
 __generic<T : __BuiltinFloatingPointType, let N : int>
 __target_intrinsic(hlsl)
@@ -1891,7 +1972,6 @@ vector<T, N> rcp(vector<T, N> x)
 
 __generic<T : __BuiltinFloatingPointType, let N : int, let M : int>
 __target_intrinsic(hlsl)
-// Note: GLSL doesn't define a vector `rcp`, so not intrinsic there
 matrix<T, N, M> rcp(matrix<T, N, M> x)
 {
     MATRIX_MAP_UNARY(T, N, M, rcp, x);
@@ -1899,27 +1979,35 @@ matrix<T, N, M> rcp(matrix<T, N, M> x)
 
 // Reflect incident vector across plane with given normal
 __generic<T : __BuiltinFloatingPointType, let N : int>
-vector<T,N> reflect(vector<T,N> i, vector<T,N> n);
-/*{
+__target_intrinsic(hlsl)
+__target_intrinsic(glsl)
+vector<T,N> reflect(vector<T,N> i, vector<T,N> n)
+{
     return i - T(2) * dot(n,i) * n;
-}*/
+}
 
 // Refract incident vector given surface normal and index of refraction
 __generic<T : __BuiltinFloatingPointType, let N : int>
-vector<T,N> refract(vector<T,N> i, vector<T,N> n, float eta);
-/*{
+__target_intrinsic(hlsl)
+__target_intrinsic(glsl)
+vector<T,N> refract(vector<T,N> i, vector<T,N> n, T eta)
+{
     let dotNI = dot(n,i);
     let k = T(1) - eta*eta*(T(1) - dotNI * dotNI);
-    if(k < 0) return vector<T,N>(T(0));
+    if(k < T(0)) return vector<T,N>(T(0));
     return eta * i - (eta * dotNI + sqrt(k)) * n;
-}*/
+}
 
 // Reverse order of bits
 __target_intrinsic(glsl, "bitfieldReverse")
 uint reversebits(uint value);
 
 __target_intrinsic(glsl, "bitfieldReverse")
-__generic<let N : int> vector<uint,N> reversebits(vector<uint,N> value);
+__generic<let N : int>
+vector<uint, N> reversebits(vector<uint, N> value)
+{
+    VECTOR_MAP_UNARY(uint, N, reversebits, value);
+}
 
 // Round-to-nearest
 __generic<T : __BuiltinFloatingPointType>
@@ -2073,7 +2161,13 @@ matrix<T, N, M> sinh(matrix<T, N, M> x)
 
 // Smooth step (Hermite interpolation)
 __generic<T : __BuiltinFloatingPointType>
-T smoothstep(T min, T max, T x);
+__target_intrinsic(hlsl)
+__target_intrinsic(glsl)
+T smoothstep(T min, T max, T x)
+{
+    let t = saturate((x - min) / (max - min));
+    return t * t * (T(3.0f) - (t + t));
+}
 
 __generic<T : __BuiltinFloatingPointType, let N : int>
 __target_intrinsic(hlsl)
@@ -2113,10 +2207,10 @@ matrix<T, N, M> sqrt(matrix<T, N, M> x)
 __generic<T : __BuiltinFloatingPointType>
 __target_intrinsic(hlsl)
 __target_intrinsic(glsl)
-T step(T y, T x);
-/*{
+T step(T y, T x)
+{
     return x < y ? T(0.0f) : T(1.0f);
-}*/
+}
 
 __generic<T : __BuiltinFloatingPointType, let N : int>
 __target_intrinsic(hlsl)
diff --git a/source/slang/slang-emit.cpp b/source/slang/slang-emit.cpp
index dcca7d25e..e613f5462 100644
--- a/source/slang/slang-emit.cpp
+++ b/source/slang/slang-emit.cpp
@@ -15,6 +15,7 @@
 #include "slang-ir-specialize.h"
 #include "slang-ir-specialize-resources.h"
 #include "slang-ir-ssa.h"
+#include "slang-ir-strip-witness-tables.h"
 #include "slang-ir-union.h"
 #include "slang-ir-validate.h"
 #include "slang-ir-wrap-structured-buffers.h"
@@ -441,6 +442,16 @@ Result linkAndOptimizeIR(
         break;
     }
 
+    // For all targets that don't support true dynamic dispatch through
+    // witness tables (that is all targets at present), we need
+    // to eliminate witness tables from the IR so that they
+    // don't keep symbols live that we don't actually need.
+    stripWitnessTables(irModule);
+#if 0
+    dumpIRIfEnabled(compileRequest, irModule, "AFTER STRIP WITNESS TABLES");
+#endif
+    validateIRModuleIfEnabled(compileRequest, irModule);
+
     // The resource-based specialization pass above
     // may create specialized versions of functions, but
     // it does not try to completely eliminate the original
diff --git a/source/slang/slang-ir-strip-witness-tables.cpp b/source/slang/slang-ir-strip-witness-tables.cpp
new file mode 100644
index 000000000..8536508ba
--- /dev/null
+++ b/source/slang/slang-ir-strip-witness-tables.cpp
@@ -0,0 +1,33 @@
+// slang-ir-strip-witness-tables.cpp
+#include "slang-ir-strip-witness-tables.h"
+
+#include "slang-ir.h"
+#include "slang-ir-insts.h"
+
+namespace Slang
+{
+
+void stripWitnessTables(IRModule* module)
+{
+    // Our goal here is to empty out any witness tables in
+    // the IR so that they don't keep other symbols alive
+    // further into compilation. Luckily we expect all
+    // witness tables to live directly at the global scope
+    // (or inside of a generic, which we can ignore for
+    // now because the emit logic also ignores generics),
+    // and there is a single function we can call to
+    // remove all of the content from the witness tables
+    // (since the key-value associations are stored as
+    // children of each table).
+
+    for( auto inst : module->getGlobalInsts() )
+    {
+        auto witnessTable = as<IRWitnessTable>(inst);
+        if(!witnessTable)
+            continue;
+
+        witnessTable->removeAndDeallocateAllDecorationsAndChildren();
+    }
+}
+
+}
\ No newline at end of file
diff --git a/source/slang/slang-ir-strip-witness-tables.h b/source/slang/slang-ir-strip-witness-tables.h
new file mode 100644
index 000000000..43bd0127d
--- /dev/null
+++ b/source/slang/slang-ir-strip-witness-tables.h
@@ -0,0 +1,10 @@
+// slang-ir-strip-witness-tables.cpp
+#pragma once
+
+namespace Slang
+{
+struct IRModule;
+
+    /// Strip the contents of all witness table instructions from the given IR `module`
+void stripWitnessTables(IRModule* module);
+}
\ No newline at end of file
diff --git a/source/slang/slang.vcxproj b/source/slang/slang.vcxproj
index 76cffe08b..2f55fffdc 100644
--- a/source/slang/slang.vcxproj
+++ b/source/slang/slang.vcxproj
@@ -230,6 +230,7 @@
     <ClInclude Include="slang-ir-specialize.h" />
     <ClInclude Include="slang-ir-ssa.h" />
     <ClInclude Include="slang-ir-string-hash.h" />
+    <ClInclude Include="slang-ir-strip-witness-tables.h" />
     <ClInclude Include="slang-ir-strip.h" />
     <ClInclude Include="slang-ir-type-set.h" />
     <ClInclude Include="slang-ir-union.h" />
@@ -313,6 +314,7 @@
     <ClCompile Include="slang-ir-specialize.cpp" />
     <ClCompile Include="slang-ir-ssa.cpp" />
     <ClCompile Include="slang-ir-string-hash.cpp" />
+    <ClCompile Include="slang-ir-strip-witness-tables.cpp" />
     <ClCompile Include="slang-ir-strip.cpp" />
     <ClCompile Include="slang-ir-type-set.cpp" />
     <ClCompile Include="slang-ir-union.cpp" />
diff --git a/source/slang/slang.vcxproj.filters b/source/slang/slang.vcxproj.filters
index 449f72069..7a0f465fe 100644
--- a/source/slang/slang.vcxproj.filters
+++ b/source/slang/slang.vcxproj.filters
@@ -1,4 +1,4 @@
-<?xml version="1.0" encoding="utf-8"?>
+<?xml version="1.0" encoding="utf-8"?>
 <Project ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
   <ItemGroup>
     <Filter Include="Header Files">
@@ -249,6 +249,9 @@
     <ClInclude Include="slang-visitor.h">
       <Filter>Header Files</Filter>
     </ClInclude>
+    <ClInclude Include="slang-ir-strip-witness-tables.h">
+      <Filter>Header Files</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="slang-check-conformance.cpp">
@@ -467,6 +470,9 @@
     <ClCompile Include="slang.cpp">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="slang-ir-strip-witness-tables.cpp">
+      <Filter>Source Files</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <None Include="..\core\core.natvis">