llvm: codegen on Apple Silicon

constantine/math_compiler/impl_curves_ops_affine.nim

@@ -17,7 +17,7 @@ import
   std / typetraits # for distinctBase
 
 ## Section name used for `llvmInternalFnDef`
-const SectionName = "ctt.curves_affine"
+const SectionName = "ctt,curves_affine"
 
 type
   EcPointAff* {.borrow: `.`.} = distinct Array

constantine/math_compiler/impl_curves_ops_jacobian.nim

@@ -18,7 +18,7 @@ import
   std / typetraits # for distinctBase
 
 ## Section name used for `llvmInternalFnDef`
-const SectionName = "ctt.curves_jacobian"
+const SectionName = "ctt,curves_jacobian"
 
 type
   EcPointJac* {.borrow: `.`.} = distinct Array

constantine/math_compiler/impl_fields_dispatch.nim

@@ -14,17 +14,17 @@ import
 
 proc modadd*(asy: Assembler_LLVM, fd: FieldDescriptor, r, a, b, M: ValueRef) =
   case asy.backend
-  of {bkX86_64_Linux, bkAmdGpu}:
+  of {bkX86_64_Linux, bkArm64_MacOS, bkAmdGpu}:
     asy.modadd_sat(fd, r, a, b, M)
   of bkNvidiaPTX:
     asy.modadd_nvidia(fd, r, a, b, M)
 
 proc modsub*(asy: Assembler_LLVM, fd: FieldDescriptor, r, a, b, M: ValueRef) =
   case asy.backend
+  of {bkX86_64_Linux, bkArm64_MacOS, bkAmdGpu}:
+    asy.modsub_sat(fd, r, a, b, M)
   of bkNvidiaPTX:
     asy.modsub_nvidia(fd, r, a, b, M)
-  else:
-    doAssert false, "Unimplemented"
 
 proc mtymul*(asy: Assembler_LLVM, fd: FieldDescriptor, r, a, b, M: ValueRef, finalReduce = true) =
   case asy.backend

constantine/math_compiler/impl_fields_nvidia.nim

@@ -47,7 +47,7 @@ import
 #
 # We cannot use i256 on Nvidia target: https://github.com/llvm/llvm-project/blob/llvmorg-18.1.8/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp#L244-L276
 
-const SectionName = "ctt.fields"
+const SectionName = "ctt,fields"
 
 proc finalSubMayOverflow(asy: Assembler_LLVM, fd: FieldDescriptor, r, a, M: Array) =
   ## If a >= Modulus: r <- a-M

constantine/math_compiler/impl_fields_ops.nim

@@ -14,7 +14,7 @@ import
   ./impl_fields_dispatch
 
 ## Section name used for `llvmInternalFnDef`
-const SectionName = "ctt.impl_fields"
+const SectionName = "ctt,impl_fields"
 
 template declFieldOps*(asy: Assembler_LLVM, fd: FieldDescriptor): untyped {.dirty.} =
   ## This template can be used to make operations on `Field` elements

constantine/math_compiler/impl_fields_sat.nim

@@ -76,7 +76,7 @@ import
 # And implementing them with i256 / i384 is similarly tricky
 # https://github.com/llvm/llvm-project/issues/102868
 
-const SectionName = "ctt.fields"
+const SectionName = "ctt,fields"
 
 proc finalSubMayOverflow*(asy: Assembler_LLVM, fd: FieldDescriptor, rr, a, M, carry: ValueRef) =
   ## If a >= Modulus: r <- a-M

constantine/math_compiler/impl_msm_nvidia.nim

@@ -18,7 +18,7 @@ import
   std / typetraits # for distinctBase
 
 ## Section name used for `llvmInternalFnDef`
-const SectionName = "ctt.msm_nvidia"
+const SectionName = "ctt,msm_nvidia"
 
 proc msm*(asy: Assembler_LLVM, cd: CurveDescriptor, r, coefs, points: ValueRef,
           c, N: int) {.used.} =

constantine/math_compiler/ir.nim

@@ -55,6 +55,7 @@ type
     bkAmdGpu
     bkNvidiaPTX
     bkX86_64_Linux
+    bkArm64_MacOS
 
 proc finalizeAssemblerLLVM(asy: Assembler_LLVM) =
   if not asy.isNil:
@@ -94,6 +95,8 @@ proc configure(asy: var Assembler_LLVM, backend: Backend) =
       "n16:32:64")
   of bkX86_64_Linux:
     asy.module.setTarget("x86_64-pc-linux-gnu")
+  of bkArm64_MacOS:
+    asy.module.setTarget("aarch64-apple-darwin")
 
   asy.datalayout = asy.module.getDataLayout()
   asy.psize = int32 asy.datalayout.getPointerSize()
@@ -505,7 +508,7 @@ proc defineGlobalConstant*(
   # This has the following benefits:
   # - They might all be loaded in memory if they share a cacheline
   # - If a section is unused, it can be garbage collected by the linker
-  g.setSection(cstring("ctt." & section & ".constants"))
+  g.setSection(cstring("ctt," & section)) # mach-o (MacOS), needs a comma
   return g
 
 # ############################################################

constantine/math_compiler/pub_curves_affine.nim

@@ -16,7 +16,7 @@ import
   ./impl_curves_ops_affine
 
 ## Section name used for `llvmInternalFnDef`
-const SectionName = "ctt.pub_curves_affine"
+const SectionName = "ctt,pub_curves_affine"
 
 proc genEcIsNeutralAff*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
   ## Generate a public elliptic curve point isNeutral proc
@@ -27,7 +27,7 @@ proc genEcIsNeutralAff*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
 
   let name = cd.name & "isNeutralAff"
   let ptrBool = pointer_t(asy.ctx.int1_t())
-  asy.llvmPublicFnDef(name, "ctt." & cd.name, asy.void_t, [ptrBool, cd.curveTyAff]):
+  asy.llvmPublicFnDef(name, "ctt," & cd.name, asy.void_t, [ptrBool, cd.curveTyAff]):
     let (r, a) = llvmParams
     asy.isNeutralAff_internal(cd, r, a)
     asy.br.retVoid()

constantine/math_compiler/pub_curves_jacobian.nim

@@ -20,7 +20,7 @@ import
 
 
 ## Section name used for `llvmInternalFnDef`
-const SectionName = "ctt.pub_curves_jacobian"
+const SectionName = "ctt,pub_curves_jacobian"
 
 proc genEcFromAffine*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
   ## Generate a public elliptic curve point `fromAffine` proc
@@ -30,7 +30,7 @@ proc genEcFromAffine*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
   ## and return the corresponding name to call it
 
   let name = cd.name & "_isNeutral"
-  asy.llvmPublicFnDef(name, "ctt." & cd.name, asy.void_t, [cd.curveTy, cd.curveTyAff]):
+  asy.llvmPublicFnDef(name, "ctt," & cd.name, asy.void_t, [cd.curveTy, cd.curveTyAff]):
     let (jac, aff) = llvmParams
     asy.fromAffine(cd, jac, aff)
     asy.br.retVoid()
@@ -46,7 +46,7 @@ proc genEcIsNeutral*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
 
   let name = cd.name & "_isNeutral"
   let ptrBool = pointer_t(asy.ctx.int1_t())
-  asy.llvmPublicFnDef(name, "ctt." & cd.name, asy.void_t, [ptrBool, cd.curveTy]):
+  asy.llvmPublicFnDef(name, "ctt," & cd.name, asy.void_t, [ptrBool, cd.curveTy]):
     let (r, a) = llvmParams
     asy.isNeutral(cd, r, a)
     asy.br.retVoid()
@@ -62,7 +62,7 @@ proc genEcSetNeutral*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
   ## It returns the corresponding name to call it
   let name = cd.name & "_setNeutral"
   let ptrBool = pointer_t(asy.ctx.int1_t())
-  asy.llvmPublicFnDef(name, "ctt." & cd.name, asy.void_t, [cd.curveTy]):
+  asy.llvmPublicFnDef(name, "ctt," & cd.name, asy.void_t, [cd.curveTy]):
     let r = llvmParams
     asy.setNeutral(cd, r)
     asy.br.retVoid()
@@ -79,7 +79,7 @@ proc genEcCcopy*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
   ## and return the corresponding name to call it
 
   let name = cd.name & "_ccopy"
-  asy.llvmPublicFnDef(name, "ctt." & cd.name, asy.void_t, [cd.curveTy, cd.curveTy, asy.ctx.int1_t()]):
+  asy.llvmPublicFnDef(name, "ctt," & cd.name, asy.void_t, [cd.curveTy, cd.curveTy, asy.ctx.int1_t()]):
     let (a, b, c) = llvmParams
     asy.ccopy(cd, a, b, c)
     asy.br.retVoid()
@@ -94,7 +94,7 @@ proc genEcNeg*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
   ## and return the corresponding name to call it
 
   let name = cd.name & "_neg"
-  asy.llvmPublicFnDef(name, "ctt." & cd.name, asy.void_t, [cd.curveTy]):
+  asy.llvmPublicFnDef(name, "ctt," & cd.name, asy.void_t, [cd.curveTy]):
     let a = llvmParams
     asy.neg(cd, a)
     asy.br.retVoid()
@@ -109,7 +109,7 @@ proc genEcCneg*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
   ## Returns the name to call the kernel.
 
   let name = cd.name & "_neg"
-  asy.llvmPublicFnDef(name, "ctt." & cd.name, asy.void_t, [cd.curveTy, asy.ctx.int1_t()]):
+  asy.llvmPublicFnDef(name, "ctt," & cd.name, asy.void_t, [cd.curveTy, asy.ctx.int1_t()]):
     let (a, c) = llvmParams
     asy.cneg(cd, a, c)
     asy.br.retVoid()
@@ -125,7 +125,7 @@ proc genEcSum*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
   ##
   ## Returns the name of the produced kernel to call it.
   let name = cd.name & "_sum"
-  asy.llvmPublicFnDef(name, "ctt." & cd.name, asy.void_t, [cd.curveTy, cd.curveTy, cd.curveTy]):
+  asy.llvmPublicFnDef(name, "ctt," & cd.name, asy.void_t, [cd.curveTy, cd.curveTy, cd.curveTy]):
     let (ri, pi, qi) = llvmParams
     asy.sum(cd, ri, pi, qi)
     asy.br.retVoid()
@@ -139,7 +139,7 @@ proc genEcDouble*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
   ##
   ## Returns the name of the produced kernel to call it.
   let name = cd.name & "_double"
-  asy.llvmPublicFnDef(name, "ctt." & cd.name, asy.void_t, [cd.curveTy, cd.curveTy]):
+  asy.llvmPublicFnDef(name, "ctt," & cd.name, asy.void_t, [cd.curveTy, cd.curveTy]):
     let (ri, pi) = llvmParams
     asy.double(cd, ri, pi)
     asy.br.retVoid()
@@ -155,7 +155,7 @@ proc genEcMixedSum*(asy: Assembler_LLVM, cd: CurveDescriptor): string =
   ##
   ## Returns the name of the produced kernel to call it.
   let name = cd.name & "_mixedSum"
-  asy.llvmPublicFnDef(name, "ctt." & cd.name, asy.void_t, [cd.curveTy, cd.curveTy, cd.curveTyAff]):
+  asy.llvmPublicFnDef(name, "ctt," & cd.name, asy.void_t, [cd.curveTy, cd.curveTy, cd.curveTyAff]):
     let (ri, pi, qi) = llvmParams
     asy.mixedSum(cd, ri, pi, qi)
     asy.br.retVoid()
@@ -175,7 +175,7 @@ proc genEcMSM*(asy: Assembler_LLVM, cd: CurveDescriptor, c, N: int): string =
   let cT = array_t(cd.curveTyAff, N)
   let name = cd.name & "_msm_public_c_" & $c & "_N_" & $N
 
-  asy.llvmPublicFnDef(name, "ctt." & cd.name, asy.void_t, [cd.curveTy, fT, cT]):
+  asy.llvmPublicFnDef(name, "ctt," & cd.name, asy.void_t, [cd.curveTy, fT, cT]):
     let (ri, cs, ps) = llvmParams
     asy.msm(cd, ri, cs, ps, c, N)
     asy.br.retVoid()