jit: Gen stencils for x64.

fubark · Dec 8, 2023 · 08cee6b · 08cee6b
1 parent 0b68f7c
commit 08cee6b
Show file tree

Hide file tree

Showing 5 changed files with 197 additions and 3 deletions.
diff --git a/src/jit/stencils.zig → src/jit/a64_stencils.zig b/src/jit/stencils.zig → src/jit/a64_stencils.zig
diff --git a/src/jit/gen-stencils.cy → src/jit/gen-stencils-a64.cy b/src/jit/gen-stencils.cy → src/jit/gen-stencils-a64.cy
@@ -2,7 +2,7 @@ import llvm '../tools/llvm.cy'
 import os
 
 --|
---| Takes stencils.o and generates stencils.zig.
+--| Takes stencils.o and generates a64_stencils.zig
 --|
 
 var out = ''
@@ -20,19 +20,31 @@ var buf = cbuf.toArray(0, size)
 
 var llBin = llvm.CreateBinary(llBuf, none, outMsg)
 
+var binType = llvm.BinaryGetType(llBin)
+if binType != llvm.BinaryTypeMachO64L:
+    throw error.UnexpectedObjectFormat
+
 -- Find text section.
 my codeBuf = none
 var llSectIter = llvm.ObjectFileCopySectionIterator(llBin)
 while llvm.ObjectFileIsSectionIteratorAtEnd(llBin, llSectIter) == 0:
     var cname = llvm.GetSectionName(llSectIter)
+    if cname.addr() == 0:
+        llvm.MoveToNextSection(llSectIter)
+        continue
+
     var name = cname.fromCstr(0).decode()
+
     if name == '__text':
         var ccodeBuf = llvm.GetSectionContents(llSectIter)
         var size = llvm.GetSectionSize(llSectIter)
         codeBuf = ccodeBuf.toArray(0, size)
         break
     llvm.MoveToNextSection(llSectIter)
 
+if codeBuf == none:
+    throw error.MissingTextSection
+
 var llSymIter = llvm.ObjectFileCopySymbolIterator(llBin)
 
 type Sym object:
@@ -51,6 +63,7 @@ while llvm.ObjectFileIsSymbolIteratorAtEnd(llBin, llSymIter) == 0:
     var cname = llvm.GetSymbolName(llSymIter)
     var name = cname.fromCstr(0).decode()
     var addr = llvm.GetSymbolAddress(llSymIter)
+    -- Size is missing, calculate by sorting symbols and using their address.
     var sym = [Sym name: name, addr: addr]
     syms.append(sym)
     symMap[name] = sym
@@ -85,6 +98,7 @@ for syms -> sym, i:
 
     out += 'pub const $(sym.name[1..]) = [_]u8{ $(bytes.join(', ')) };\n'
 
+-- llSectIter is already at text section.
 -- Visit relocation entries and record them as Zig constants.
 var llRelocIter = llvm.GetRelocations(llSectIter)
 while llvm.IsRelocationIteratorAtEnd(llSectIter, llRelocIter) == 0:
@@ -119,4 +133,4 @@ while llvm.IsRelocationIteratorAtEnd(llSectIter, llRelocIter) == 0:
 
     llvm.MoveToNextRelocation(llRelocIter)
 
-os.writeFile('$(curDir)/stencils.zig', out)
+os.writeFile('$(curDir)/a64_stencils.zig', out)
diff --git a/src/jit/gen-stencils-x64.cy b/src/jit/gen-stencils-x64.cy
@@ -0,0 +1,147 @@
+import llvm '../tools/llvm.cy'
+import os
+
+--|
+--| Takes stencils.o and generates x64_stencils.zig
+--|
+
+var out = ''
+var curDir = os.dirName(#ModUri)
+
+my outLLBuf = llvm.ffi.new(.voidPtr)
+var outMsg = llvm.ffi.new(.charPtr)
+if llvm.CreateMemoryBufferWithContentsOfFile(os.cstr('$(curDir)/stencils.o'), outLLBuf, outMsg) != 0:
+    throw error.Unexpected
+
+var llBuf = outLLBuf.get(0, .voidPtr)
+var cbuf = llvm.GetBufferStart(llBuf)
+var size = llvm.GetBufferSize(llBuf)
+var buf = cbuf.toArray(0, size)
+
+var llBin = llvm.CreateBinary(llBuf, none, outMsg)
+
+var binType = llvm.BinaryGetType(llBin)
+if binType != llvm.BinaryTypeELF64L:
+    throw error.UnexpectedObjectFormat
+
+-- Find text section.
+my codeBuf = none
+var llSectIter = llvm.ObjectFileCopySectionIterator(llBin)
+while llvm.ObjectFileIsSectionIteratorAtEnd(llBin, llSectIter) == 0:
+    var cname = llvm.GetSectionName(llSectIter)
+    if cname.addr() == 0:
+        llvm.MoveToNextSection(llSectIter)
+        continue
+
+    var name = cname.fromCstr(0).decode()
+
+    if name == '.text':
+        var ccodeBuf = llvm.GetSectionContents(llSectIter)
+        var size = llvm.GetSectionSize(llSectIter)
+        codeBuf = ccodeBuf.toArray(0, size)
+        break
+    llvm.MoveToNextSection(llSectIter)
+
+if codeBuf == none:
+    throw error.MissingTextSection
+
+var llSymIter = llvm.ObjectFileCopySymbolIterator(llBin)
+
+type Sym object:
+    var name string
+    var addr int
+    var code array
+
+-- First pass accumulates the unordered symbols.
+var syms = []
+var symMap = [:]
+while llvm.ObjectFileIsSymbolIteratorAtEnd(llBin, llSymIter) == 0:
+    if llvm.GetSectionContainsSymbol(llSectIter, llSymIter) == 0:
+        -- Not in text section, skip.
+        llvm.MoveToNextSymbol(llSymIter)
+        continue
+
+    var cname = llvm.GetSymbolName(llSymIter)
+    var name = cname.fromCstr(0).decode()
+    if name == '.text':
+        llvm.MoveToNextSymbol(llSymIter)
+        continue
+
+    var addr = llvm.GetSymbolAddress(llSymIter)
+    var size = llvm.GetSymbolSize(llSymIter)
+    var code = codeBuf[addr..addr+size] as array
+    var sym = [Sym name: name, addr: addr, code: code]
+    syms.append(sym)
+    symMap[name] = sym
+
+    llvm.MoveToNextSymbol(llSymIter)
+
+-- Seek to relocation section.
+llSectIter = llvm.ObjectFileCopySectionIterator(llBin)
+while llvm.ObjectFileIsSectionIteratorAtEnd(llBin, llSectIter) == 0:
+    var cname = llvm.GetSectionName(llSectIter)
+    if cname.addr() == 0:
+        llvm.MoveToNextSection(llSectIter)
+        continue
+
+    var name = cname.fromCstr(0).decode()
+    if name == '.rela.text':
+        break
+
+    llvm.MoveToNextSection(llSectIter)
+
+-- Visit relocation entries and record them as Zig constants.
+var llRelocIter = llvm.GetRelocations(llSectIter)
+while llvm.IsRelocationIteratorAtEnd(llSectIter, llRelocIter) == 0:
+    var symRef = llvm.GetRelocationSymbol(llRelocIter)
+    var csymName = llvm.GetSymbolName(symRef)
+    var symName = csymName.fromCstr(0).decode()
+
+    var offset = llvm.GetRelocationOffset(llRelocIter)
+    var relocType = llvm.GetRelocationType(llRelocIter)
+    var cname = llvm.GetRelocationTypeName(llRelocIter)
+    var name = cname.fromCstr(0).decode()
+    var cvalue = llvm.GetRelocationValueString(llRelocIter)
+    var value = cname.fromCstr(0).decode()
+
+    var instOffset = 0
+    var R_X86_64_PLT32 = 4
+    if relocType == R_X86_64_PLT32:
+        instOffset = 1
+
+    -- Find relevant func sym.
+    my found = none
+    for syms -> sym, i:
+        if offset >= sym.addr:
+            if i < syms.len()-1 and offset >= syms[i+1].addr:
+                continue
+            found = sym
+            break
+
+    if found == none:
+        throw error.MissingSym
+
+    var roffset = (offset - instOffset) - found.addr
+    if symName.startsWith('cont'):
+        -- Remove code after continuation.
+        found.code = found.code[0..roffset]
+
+        -- Skip continuations.
+        llvm.MoveToNextRelocation(llRelocIter)
+        continue
+
+    out += 'pub const $(found.name)_$(symName) = $(roffset);\n'
+
+    llvm.MoveToNextRelocation(llRelocIter)
+
+-- After continuations are removed, gen sym's code.
+for syms -> sym, i:
+    print '$(sym.name) $(sym.addr) $(sym.code.fmt(.x))'
+
+    var bytes = []
+    for sym.code -> byte:
+        bytes.append('0x$(byte.fmt(.x, [pad: `0`, width: 2]))')
+
+    out += 'pub const $(sym.name) = [_]u8{ $(bytes.join(', ')) };\n'
+
+os.writeFile('$(curDir)/x64_stencils.zig', out)
diff --git a/src/jit/stencils.c b/src/jit/stencils.c
@@ -5,6 +5,9 @@
 // clang-wasm: clang -c stencil.c -I../ -O2 --target=wasm32 -mtail-call -o stencils.o
 // zig: zig cc -c stencils.c -fdouble-square-bracket-attributes -O2 -I../ -o stencils.o
 
+// For x86_64 only the linux target seems to omit the function call prelude:
+// clang -c stencils.c -o stencils.o -I../ -O2 -target x86_64-linux 
+
 Value hostFunc(VM* vm, const Value* args, u8 nargs);
 void zDumpJitSection(VM* vm, Value* fp, u64 chunkId, u64 irIdx, u8* startPc, u8* endPc);
 void cont(Value* fp);
@@ -118,4 +121,8 @@ void release(VM* vm, Value* fp, Value val) {
         }
     }
     [[clang::musttail]] return cont3(vm, fp, val);
-}
+}
+
+// void test(VM* vm, Value* fp, Value a, Value b) {
+//     // [[clang::musttail]] return cont4(vm+1, fp, a, b);
+// }
diff --git a/src/jit/x64_stencils.zig b/src/jit/x64_stencils.zig
@@ -0,0 +1,26 @@
+pub const divInt_divByZero = 60;
+pub const callHost_hostFunc = 41;
+pub const callHost_interrupt5 = 75;
+pub const stringTemplate_zAllocStringTemplate2 = 38;
+pub const stringTemplate_interrupt5 = 71;
+pub const dumpJitSection_zDumpJitSection = 29;
+pub const release_zFreeObject = 49;
+pub const addFloat = [_]u8{ 0x66, 0x48, 0x0f, 0x6e, 0xc2, 0x66, 0x48, 0x0f, 0x6e, 0xc9, 0xf2, 0x0f, 0x58, 0xc8, 0x66, 0x48, 0x0f, 0x7e, 0xca };
+pub const subFloat = [_]u8{ 0x66, 0x48, 0x0f, 0x6e, 0xc2, 0x66, 0x48, 0x0f, 0x6e, 0xc9, 0xf2, 0x0f, 0x5c, 0xc1, 0x66, 0x48, 0x0f, 0x7e, 0xc2 };
+pub const mulFloat = [_]u8{ 0x66, 0x48, 0x0f, 0x6e, 0xc2, 0x66, 0x48, 0x0f, 0x6e, 0xc9, 0xf2, 0x0f, 0x59, 0xc8, 0x66, 0x48, 0x0f, 0x7e, 0xca };
+pub const divFloat = [_]u8{ 0x66, 0x48, 0x0f, 0x6e, 0xc2, 0x66, 0x48, 0x0f, 0x6e, 0xc9, 0xf2, 0x0f, 0x5e, 0xc1, 0x66, 0x48, 0x0f, 0x7e, 0xc2 };
+pub const addInt = [_]u8{ 0x48, 0x01, 0xca, 0x48, 0xb8, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x48, 0x21, 0xc2, 0x48, 0xb8, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xfe, 0x7f, 0x48, 0x09, 0xc2 };
+pub const subInt = [_]u8{ 0x48, 0x29, 0xca, 0x48, 0xb8, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x48, 0x21, 0xc2, 0x48, 0xb8, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xfe, 0x7f, 0x48, 0x09, 0xc2 };
+pub const mulInt = [_]u8{ 0x48, 0x0f, 0xaf, 0xd1, 0x48, 0xb8, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x48, 0x21, 0xc2, 0x48, 0xb8, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xfe, 0x7f, 0x48, 0x09, 0xc2 };
+pub const divInt = [_]u8{ 0x49, 0x89, 0xc8, 0x49, 0xc1, 0xe0, 0x10, 0x74, 0x33, 0x49, 0xc1, 0xf8, 0x10, 0x48, 0xc1, 0xe2, 0x10, 0x48, 0xc1, 0xfa, 0x10, 0x48, 0x89, 0xd0, 0x48, 0x99, 0x49, 0xf7, 0xf8, 0x49, 0xb8, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x49, 0x21, 0xc0, 0x48, 0xba, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xfe, 0x7f, 0x4c, 0x09, 0xc2 };
+pub const lessInt = [_]u8{ 0x48, 0xc1, 0xe2, 0x10, 0x48, 0xc1, 0xfa, 0x10, 0x49, 0x89, 0xc8, 0x49, 0xc1, 0xe0, 0x10, 0x49, 0xc1, 0xf8, 0x10, 0x49, 0xb9, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0xfc, 0x7f, 0x49, 0x8d, 0x41, 0x01, 0x4c, 0x39, 0xc2, 0x49, 0x0f, 0x4d, 0xc1, 0x48, 0x89, 0xc2 };
+pub const intPair = [_]u8{ 0x48, 0xc1, 0xe2, 0x10, 0x48, 0xc1, 0xfa, 0x10, 0x48, 0xc1, 0xe1, 0x10, 0x48, 0xc1, 0xf9, 0x10 };
+pub const isTrue = [_]u8{ 0x48, 0xb8, 0x00, 0x00, 0x00, 0x00, 0x07, 0x00, 0xff, 0xff, 0x48, 0x21, 0xd0, 0x48, 0xb9, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0xfc, 0x7f, 0x4c, 0x8d, 0x41, 0x01, 0x45, 0x31, 0xc9, 0x4c, 0x39, 0xc2, 0x41, 0x0f, 0x94, 0xc1, 0x49, 0xb8, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xfc, 0x7f, 0x45, 0x31, 0xd2, 0x4c, 0x39, 0xc2, 0x41, 0x0f, 0x95, 0xc2, 0x48, 0x39, 0xc8, 0x45, 0x0f, 0x45, 0xca, 0x41, 0x0f, 0xb6, 0xd1 };
+pub const lessIntCFlag = [_]u8{ 0x48, 0x39, 0xca, 0x0f };
+pub const call = [_]u8{ 0x48, 0x83, 0xc6, 0x28 };
+pub const callHost = [_]u8{ 0x41, 0x57, 0x41, 0x56, 0x41, 0x55, 0x41, 0x54, 0x53, 0x48, 0x89, 0xcb, 0x49, 0x89, 0xd6, 0x49, 0x89, 0xf7, 0x49, 0x89, 0xfd, 0x49, 0xbc, 0xff, 0xff, 0x00, 0x00, 0x02, 0x00, 0xfc, 0x7f, 0x48, 0x89, 0x77, 0x20, 0x0f, 0xb6, 0xd3, 0x4c, 0x89, 0xf6, 0xe8, 0x00, 0x00, 0x00, 0x00, 0x4c, 0x89, 0xef, 0x4c, 0x89, 0xfe, 0x4c, 0x89, 0xf2, 0x48, 0x89, 0xd9, 0x4c, 0x39, 0xe0, 0x75, 0x11, 0x4d, 0x89, 0xe0, 0x5b, 0x41, 0x5c, 0x41, 0x5d, 0x41, 0x5e, 0x41, 0x5f, 0xe9, 0x00, 0x00, 0x00, 0x00, 0x49, 0x89, 0xc0, 0x5b, 0x41, 0x5c, 0x41, 0x5d, 0x41, 0x5e, 0x41, 0x5f };
+pub const end = [_]u8{ 0x0f, 0xb6, 0xc2, 0x48, 0x89, 0x87, 0xb0, 0x03, 0x00, 0x00 };
+pub const stringTemplate = [_]u8{ 0x41, 0x57, 0x41, 0x56, 0x41, 0x55, 0x41, 0x54, 0x53, 0x4d, 0x89, 0xc6, 0x48, 0x89, 0xcb, 0x49, 0x89, 0xd4, 0x49, 0x89, 0xf7, 0x49, 0x89, 0xfd, 0x41, 0x8d, 0x40, 0x01, 0x0f, 0xb6, 0xd0, 0x45, 0x0f, 0xb6, 0xc6, 0x4c, 0x89, 0xe6, 0xe8, 0x00, 0x00, 0x00, 0x00, 0x85, 0xd2, 0x74, 0x1d, 0x41, 0x89, 0xd0, 0x4c, 0x89, 0xef, 0x4c, 0x89, 0xfe, 0x4c, 0x89, 0xe2, 0x48, 0x89, 0xd9, 0x5b, 0x41, 0x5c, 0x41, 0x5d, 0x41, 0x5e, 0x41, 0x5f, 0xe9, 0x00, 0x00, 0x00, 0x00, 0x4c, 0x89, 0xef, 0x4c, 0x89, 0xfe, 0x48, 0x89, 0xc2, 0x48, 0x89, 0xd9, 0x4d, 0x89, 0xf0, 0x5b, 0x41, 0x5c, 0x41, 0x5d, 0x41, 0x5e, 0x41, 0x5f };
+pub const dumpJitSection = [_]u8{ 0x55, 0x41, 0x57, 0x41, 0x56, 0x41, 0x55, 0x41, 0x54, 0x53, 0x50, 0x4c, 0x89, 0xcb, 0x4d, 0x89, 0xc6, 0x49, 0x89, 0xcf, 0x49, 0x89, 0xd4, 0x49, 0x89, 0xf5, 0x48, 0x89, 0xfd, 0xe8, 0x00, 0x00, 0x00, 0x00, 0x48, 0x89, 0xef, 0x4c, 0x89, 0xee, 0x4c, 0x89, 0xe2, 0x4c, 0x89, 0xf9, 0x4d, 0x89, 0xf0, 0x49, 0x89, 0xd9, 0x48, 0x83, 0xc4, 0x08, 0x5b, 0x41, 0x5c, 0x41, 0x5d, 0x41, 0x5e, 0x41, 0x5f, 0x5d };
+pub const release = [_]u8{ 0x41, 0x57, 0x41, 0x56, 0x53, 0x48, 0x89, 0xd3, 0x49, 0x89, 0xf6, 0x49, 0x89, 0xff, 0x48, 0x89, 0xd0, 0x48, 0xc1, 0xe8, 0x32, 0x3d, 0xff, 0x3f, 0x00, 0x00, 0x72, 0x1a, 0x48, 0xbe, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x01, 0x00, 0x48, 0x21, 0xde, 0xff, 0x4e, 0x04, 0x75, 0x08, 0x4c, 0x89, 0xff, 0xe8, 0x00, 0x00, 0x00, 0x00, 0x4c, 0x89, 0xff, 0x4c, 0x89, 0xf6, 0x48, 0x89, 0xda, 0x5b, 0x41, 0x5e, 0x41, 0x5f };
+//pub const test = [_]u8{ 0x48, 0x81, 0xc7, 0xb8, 0x03, 0x00, 0x00 };