Implementation of Windows amd64_win target

This is an implementation of the Windows ABI. It supports most features
(struct passing/returning, varargs, env). TLS is not yet supported.

This patch does not actually port QBE to Windows, it only allows QBE to
generate correct asm to target Windows. As a result, testing is
accomplished on a Linux host, by using a cross-compiling toolchain, and
running the resulting binaries by using wine. See:

	TARGET=amd64_win tools/test.sh all

A few cross-platform tests were changed from 'long' to 'long long' in
driver code because long in C does not match the size of a QBE 'l' on
Windows.

1 files changed, 762 insertions, 0 deletions

diff --git a/amd64/winabi.c b/amd64/winabi.c
new file mode 100755
index 0000000..82829bc
--- /dev/null
+++ b/amd64/winabi.c
@@ -0,0 +1,762 @@
+#include "all.h"
+
+#include <stdbool.h>
+
+typedef enum ArgPassStyle {
+  APS_Invalid = 0,
+  APS_Register,
+  APS_InlineOnStack,
+  APS_CopyAndPointerInRegister,
+  APS_CopyAndPointerOnStack,
+  APS_VarargsTag,
+  APS_EnvTag,
+} ArgPassStyle;
+
+typedef struct ArgClass {
+  Typ* type;
+  ArgPassStyle style;
+  int align;
+  uint size;
+  int cls;
+  Ref ref;
+} ArgClass;
+
+typedef struct ExtraAlloc ExtraAlloc;
+struct ExtraAlloc {
+  Ins instr;
+  ExtraAlloc* link;
+};
+
+#define ALIGN_DOWN(n, a) ((n) & ~((a)-1))
+#define ALIGN_UP(n, a) ALIGN_DOWN((n) + (a)-1, (a))
+
+// Number of stack bytes required be reserved for the callee.
+#define SHADOW_SPACE_SIZE 32
+
+int amd64_winabi_rsave[] = {RCX,  RDX,   R8,    R9,    R10,   R11,   RAX,  XMM0,
+                            XMM1, XMM2,  XMM3,  XMM4,  XMM5,  XMM6,  XMM7, XMM8,
+                            XMM9, XMM10, XMM11, XMM12, XMM13, XMM14, -1};
+int amd64_winabi_rclob[] = {RBX, R12, R13, R14, R15, RSI, RDI, -1};
+
+MAKESURE(winabi_arrays_ok,
+         sizeof amd64_winabi_rsave == (NGPS_WIN + NFPS + 1) * sizeof(int) &&
+             sizeof amd64_winabi_rclob == (NCLR_WIN + 1) * sizeof(int));
+
+// layout of call's second argument (RCall)
+//
+// bit 0: rax returned
+// bit 1: xmm0 returned
+// bits 23: 0
+// bits 4567: rcx, rdx, r8, r9 passed
+// bits 89ab: xmm0,1,2,3 passed
+// bit c: env call (rax passed)
+// bits d..1f: 0
+
+bits amd64_winabi_retregs(Ref r, int p[2]) {
+  assert(rtype(r) == RCall);
+
+  bits b = 0;
+  int num_int_returns = r.val & 1;
+  int num_float_returns = r.val & 2;
+  if (num_int_returns == 1) {
+    b |= BIT(RAX);
+  } else {
+    b |= BIT(XMM0);
+  }
+  if (p) {
+    p[0] = num_int_returns;
+    p[1] = num_float_returns;
+  }
+  return b;
+}
+
+static uint popcnt(bits b) {
+  b = (b & 0x5555555555555555) + ((b >> 1) & 0x5555555555555555);
+  b = (b & 0x3333333333333333) + ((b >> 2) & 0x3333333333333333);
+  b = (b & 0x0f0f0f0f0f0f0f0f) + ((b >> 4) & 0x0f0f0f0f0f0f0f0f);
+  b += (b >> 8);
+  b += (b >> 16);
+  b += (b >> 32);
+  return b & 0xff;
+}
+
+bits amd64_winabi_argregs(Ref r, int p[2]) {
+  assert(rtype(r) == RCall);
+
+  // On SysV, these are counts. Here, a count isn't sufficient, we actually need
+  // to know which ones are in use because they're not necessarily contiguous.
+  int int_passed = (r.val >> 4) & 15;
+  int float_passed = (r.val >> 8) & 15;
+  bool env_param = (r.val >> 12) & 1;
+
+  bits b = 0;
+  b |= (int_passed & 1) ? BIT(RCX) : 0;
+  b |= (int_passed & 2) ? BIT(RDX) : 0;
+  b |= (int_passed & 4) ? BIT(R8) : 0;
+  b |= (int_passed & 8) ? BIT(R9) : 0;
+  b |= (float_passed & 1) ? BIT(XMM0) : 0;
+  b |= (float_passed & 2) ? BIT(XMM1) : 0;
+  b |= (float_passed & 4) ? BIT(XMM2) : 0;
+  b |= (float_passed & 8) ? BIT(XMM3) : 0;
+  b |= env_param ? BIT(RAX) : 0;
+  if (p) {
+    // TODO: The only place this is used is live.c. I'm not sure what should be
+    // returned here wrt to using the same counter for int/float regs on win.
+    // For now, try the number of registers in use even though they're not
+    // contiguous.
+    p[0] = popcnt(int_passed);
+    p[1] = popcnt(float_passed);
+  }
+  return b;
+}
+
+typedef struct RegisterUsage {
+  // Counter for both int/float as they're counted together. Only if the bool's
+  // set in regs_passed is the given register *actually* needed for a value
+  // (i.e. needs to be saved, etc.).
+  int num_regs_passed;
+
+  // Indexed first by 0=int, 1=float, use KBASE(cls).
+  // Indexed second by register index in calling convention, so for integer,
+  // 0=RCX, 1=RDX, 2=R8, 3=R9, and for float XMM0, XMM1, XMM2, XMM3.
+  bool regs_passed[2][4];
+
+  bool rax_returned;
+  bool xmm0_returned;
+
+  // This is also used as where the va_start will start for varargs functions
+  // (there's no 'Oparv', so we need to keep track of a count here.)
+  int num_named_args_passed;
+
+  // This is set when classifying the arguments for a call (but not when
+  // classifying the parameters of a function definition).
+  bool is_varargs_call;
+
+  bool has_env;
+} RegisterUsage;
+
+static int register_usage_to_call_arg_value(RegisterUsage reg_usage) {
+  return (reg_usage.rax_returned << 0) |        //
+         (reg_usage.xmm0_returned << 1) |       //
+         (reg_usage.regs_passed[0][0] << 4) |   //
+         (reg_usage.regs_passed[0][1] << 5) |   //
+         (reg_usage.regs_passed[0][2] << 6) |   //
+         (reg_usage.regs_passed[0][3] << 7) |   //
+         (reg_usage.regs_passed[1][0] << 8) |   //
+         (reg_usage.regs_passed[1][1] << 9) |   //
+         (reg_usage.regs_passed[1][2] << 10) |  //
+         (reg_usage.regs_passed[1][3] << 11) |  //
+         (reg_usage.has_env << 12);
+}
+
+// Assigns the argument to a register if there's any left according to the
+// calling convention, and updates the regs_passed bools. Otherwise marks the
+// value as needing stack space to be passed.
+static void assign_register_or_stack(RegisterUsage* reg_usage,
+                                     ArgClass* arg,
+                                     bool is_float,
+                                     bool by_copy) {
+  if (reg_usage->num_regs_passed == 4) {
+    arg->style = by_copy ? APS_CopyAndPointerOnStack : APS_InlineOnStack;
+  } else {
+    reg_usage->regs_passed[is_float][reg_usage->num_regs_passed] = true;
+    ++reg_usage->num_regs_passed;
+    arg->style = by_copy ? APS_CopyAndPointerInRegister : APS_Register;
+  }
+  ++reg_usage->num_named_args_passed;
+}
+
+static bool type_is_by_copy(Typ* type) {
+  // Note that only these sizes are passed by register, even though e.g. a
+  // 5 byte struct would "fit", it still is passed by copy-and-pointer.
+  return type->isdark || (type->size != 1 && type->size != 2 &&
+                          type->size != 4 && type->size != 8);
+}
+
+// This function is used for both arguments and parameters.
+// begin_instr should either point at the first Oarg or Opar, and end_instr
+// should point past the last one (so to the Ocall for arguments, or to the
+// first 'real' instruction of the function for parameters).
+static void classify_arguments(RegisterUsage* reg_usage,
+                               Ins* begin_instr,
+                               Ins* end_instr,
+                               ArgClass* arg_classes,
+                               Ref* env) {
+  ArgClass* arg = arg_classes;
+  // For each argument, determine how it will be passed (int, float, stack)
+  // and update the `reg_usage` counts. Additionally, fill out arg_classes for
+  // each argument.
+  for (Ins* instr = begin_instr; instr < end_instr; ++instr, ++arg) {
+    switch (instr->op) {
+      case Oarg:
+      case Opar:
+        assign_register_or_stack(reg_usage, arg, KBASE(instr->cls),
+                                 /*by_copy=*/false);
+        arg->cls = instr->cls;
+        arg->align = 3;
+        arg->size = 8;
+        break;
+      case Oargc:
+      case Oparc: {
+        int typ_index = instr->arg[0].val;
+        Typ* type = &typ[typ_index];
+        bool by_copy = type_is_by_copy(type);
+        assign_register_or_stack(reg_usage, arg, /*is_float=*/false, by_copy);
+        arg->cls = Kl;
+        if (!by_copy && type->size <= 4) {
+          arg->cls = Kw;
+        }
+        arg->align = 3;
+        arg->size = type->size;
+        break;
+      }
+      case Oarge:
+        *env = instr->arg[0];
+        arg->style = APS_EnvTag;
+        reg_usage->has_env = true;
+        break;
+      case Opare:
+        *env = instr->to;
+        arg->style = APS_EnvTag;
+        reg_usage->has_env = true;
+        break;
+      case Oargv:
+        reg_usage->is_varargs_call = true;
+        arg->style = APS_VarargsTag;
+        break;
+    }
+  }
+
+  if (reg_usage->has_env && reg_usage->is_varargs_call) {
+    die("can't use env with varargs");
+  }
+
+  // During a varargs call, float arguments have to be duplicated to their
+  // associated integer register, so mark them as in-use too.
+  if (reg_usage->is_varargs_call) {
+    for (int i = 0; i < 4; ++i) {
+      if (reg_usage->regs_passed[/*float*/ 1][i]) {
+        reg_usage->regs_passed[/*int*/ 0][i] = true;
+      }
+    }
+  }
+}
+
+static bool is_integer_type(int ty) {
+  assert(ty >= 0 && ty < 4 && "expecting Kw Kl Ks Kd");
+  return KBASE(ty) == 0;
+}
+
+static Ref register_for_arg(int cls, int counter) {
+  assert(counter < 4);
+  if (is_integer_type(cls)) {
+    return TMP(amd64_winabi_rsave[counter]);
+  } else {
+    return TMP(XMM0 + counter);
+  }
+}
+
+static Ins* lower_call(Fn* func,
+                       Blk* block,
+                       Ins* call_instr,
+                       ExtraAlloc** pextra_alloc) {
+  // Call arguments are instructions. Walk through them to find the end of the
+  // call+args that we need to process (and return the instruction past the body
+  // of the instruction for continuing processing).
+  Ins* instr_past_args = call_instr - 1;
+  for (; instr_past_args >= block->ins; --instr_past_args) {
+    if (!isarg(instr_past_args->op)) {
+      break;
+    }
+  }
+  Ins* earliest_arg_instr = instr_past_args + 1;
+
+  // Don't need an ArgClass for the call itself, so one less than the total
+  // number of instructions we're dealing with.
+  uint num_args = call_instr - earliest_arg_instr;
+  ArgClass* arg_classes = alloc(num_args * sizeof(ArgClass));
+
+  RegisterUsage reg_usage = {0};
+  ArgClass ret_arg_class = {0};
+
+  // Ocall's two arguments are the the function to be called in 0, and, if the
+  // the function returns a non-basic type, then arg[1] is a reference to the
+  // type of the return. req checks if Refs are equal; `R` is 0.
+  bool il_has_struct_return = !req(call_instr->arg[1], R);
+  bool is_struct_return = false;
+  if (il_has_struct_return) {
+    Typ* ret_type = &typ[call_instr->arg[1].val];
+    is_struct_return = type_is_by_copy(ret_type);
+    if (is_struct_return) {
+      assign_register_or_stack(&reg_usage, &ret_arg_class, /*is_float=*/false,
+                               /*by_copy=*/true);
+    }
+    ret_arg_class.size = ret_type->size;
+  }
+  Ref env = R;
+  classify_arguments(&reg_usage, earliest_arg_instr, call_instr, arg_classes,
+                     &env);
+
+  // We now know which arguments are on the stack and which are in registers, so
+  // we can allocate the correct amount of space to stash the stack-located ones
+  // into.
+  uint stack_usage = 0;
+  for (uint i = 0; i < num_args; ++i) {
+    ArgClass* arg = &arg_classes[i];
+    // stack_usage only accounts for pushes that are for values that don't have
+    // enough registers. Large struct copies are alloca'd separately, and then
+    // only have (potentially) 8 bytes to add to stack_usage here.
+    if (arg->style == APS_InlineOnStack) {
+      if (arg->align > 4) {
+        err("win abi cannot pass alignments > 16");
+      }
+      stack_usage += arg->size;
+    } else if (arg->style == APS_CopyAndPointerOnStack) {
+      stack_usage += 8;
+    }
+  }
+  stack_usage = ALIGN_UP(stack_usage, 16);
+
+  // Note that here we're logically 'after' the call (due to emitting
+  // instructions in reverse order), so we're doing a negative stack
+  // allocation to clean up after the call.
+  Ref stack_size_ref =
+      getcon(-(int64_t)(stack_usage + SHADOW_SPACE_SIZE), func);
+  emit(Osalloc, Kl, R, stack_size_ref, R);
+
+  ExtraAlloc* return_pad = NULL;
+  if (is_struct_return) {
+    return_pad = alloc(sizeof(ExtraAlloc));
+    Ref ret_pad_ref = newtmp("abi.ret_pad", Kl, func);
+    return_pad->instr =
+        (Ins){Oalloc8, Kl, ret_pad_ref, {getcon(ret_arg_class.size, func)}};
+    return_pad->link = (*pextra_alloc);
+    *pextra_alloc = return_pad;
+    reg_usage.rax_returned = true;
+    emit(Ocopy, call_instr->cls, call_instr->to, TMP(RAX), R);
+  } else {
+    if (il_has_struct_return) {
+      // In the case that at the IL level, a struct return was specified, but as
+      // far as the calling convention is concerned it's not actually by
+      // pointer, we need to store the return value into an alloca because
+      // subsequent IL will still be treating the function return as a pointer.
+      ExtraAlloc* return_copy = alloc(sizeof(ExtraAlloc));
+      return_copy->instr =
+          (Ins){Oalloc8, Kl, call_instr->to, {getcon(8, func)}};
+      return_copy->link = (*pextra_alloc);
+      *pextra_alloc = return_copy;
+      Ref copy = newtmp("abi.copy", Kl, func);
+      emit(Ostorel, Kl, R, copy, call_instr->to);
+      emit(Ocopy, Kl, copy, TMP(RAX), R);
+      reg_usage.rax_returned = true;
+    } else if (is_integer_type(call_instr->cls)) {
+      // Only a basic type returned from the call, integer.
+      emit(Ocopy, call_instr->cls, call_instr->to, TMP(RAX), R);
+      reg_usage.rax_returned = true;
+    } else {
+      // Basic type, floating point.
+      emit(Ocopy, call_instr->cls, call_instr->to, TMP(XMM0), R);
+      reg_usage.xmm0_returned = true;
+    }
+  }
+
+  // Emit the actual call instruction. There's no 'to' value by this point
+  // because we've lowered it into register manipulation (that's the `R`),
+  // arg[0] of the call is the function, and arg[1] is register usage is
+  // documented as above (copied from SysV).
+  emit(Ocall, call_instr->cls, R, call_instr->arg[0],
+       CALL(register_usage_to_call_arg_value(reg_usage)));
+
+  if (!req(R, env)) {
+    // If there's an env arg to be passed, it gets stashed in RAX.
+    emit(Ocopy, Kl, TMP(RAX), env, R);
+  }
+
+  if (reg_usage.is_varargs_call) {
+    // Any float arguments need to be duplicated to integer registers. This is
+    // required by the calling convention so that dumping to shadow space can be
+    // done without a prototype and for varargs.
+#define DUP_IF_USED(index, floatreg, intreg)        \
+  if (reg_usage.regs_passed[/*float*/ 1][index]) {  \
+    emit(Ocast, Kl, TMP(intreg), TMP(floatreg), R); \
+  }
+    DUP_IF_USED(0, XMM0, RCX);
+    DUP_IF_USED(1, XMM1, RDX);
+    DUP_IF_USED(2, XMM2, R8);
+    DUP_IF_USED(3, XMM3, R9);
+#undef DUP_IF_USED
+  }
+
+  int reg_counter = 0;
+  if (is_struct_return) {
+    Ref first_reg = register_for_arg(Kl, reg_counter++);
+    emit(Ocopy, Kl, first_reg, return_pad->instr.to, R);
+  }
+
+  // This is where we actually do the load of values into registers or into
+  // stack slots.
+  Ref arg_stack_slots = newtmp("abi.args", Kl, func);
+  uint slot_offset = SHADOW_SPACE_SIZE;
+  ArgClass* arg = arg_classes;
+  for (Ins* instr = earliest_arg_instr; instr != call_instr; ++instr, ++arg) {
+    switch (arg->style) {
+      case APS_Register: {
+        Ref into = register_for_arg(arg->cls, reg_counter++);
+        if (instr->op == Oargc) {
+          // If this is a small struct being passed by value. The value in the
+          // instruction in this case is a pointer, but it needs to be loaded
+          // into the register.
+          emit(Oload, arg->cls, into, instr->arg[1], R);
+        } else {
+          // Otherwise, a normal value passed in a register.
+          emit(Ocopy, instr->cls, into, instr->arg[0], R);
+        }
+        break;
+      }
+      case APS_InlineOnStack: {
+        Ref slot = newtmp("abi.off", Kl, func);
+        if (instr->op == Oargc) {
+          // This is a small struct, so it's not passed by copy, but the
+          // instruction is a pointer. So we need to copy it into the stack
+          // slot. (And, remember that these are emitted backwards, so store,
+          // then load.)
+          Ref smalltmp = newtmp("abi.smalltmp", arg->cls, func);
+          emit(Ostorel, Kl, R, smalltmp, slot);
+          emit(Oload, arg->cls, smalltmp, instr->arg[1], R);
+        } else {
+          // Stash the value into the stack slot.
+          emit(Ostorel, Kl, R, instr->arg[0], slot);
+        }
+        emit(Oadd, Kl, slot, arg_stack_slots, getcon(slot_offset, func));
+        slot_offset += arg->size;
+        break;
+      }
+      case APS_CopyAndPointerInRegister:
+      case APS_CopyAndPointerOnStack: {
+        // Alloca a space to copy into, and blit the value from the instr to the
+        // copied location.
+        ExtraAlloc* arg_copy = alloc(sizeof(ExtraAlloc));
+        Ref copy_ref = newtmp("abi.copy", Kl, func);
+        arg_copy->instr =
+            (Ins){Oalloc8, Kl, copy_ref, {getcon(arg->size, func)}};
+        arg_copy->link = (*pextra_alloc);
+        *pextra_alloc = arg_copy;
+        emit(Oblit1, 0, R, INT(arg->size), R);
+        emit(Oblit0, 0, R, instr->arg[1], copy_ref);
+
+        // Now load the pointer into the correct register or stack slot.
+        if (arg->style == APS_CopyAndPointerInRegister) {
+          Ref into = register_for_arg(arg->cls, reg_counter++);
+          emit(Ocopy, Kl, into, copy_ref, R);
+        } else {
+          assert(arg->style == APS_CopyAndPointerOnStack);
+          Ref slot = newtmp("abi.off", Kl, func);
+          emit(Ostorel, Kl, R, copy_ref, slot);
+          emit(Oadd, Kl, slot, arg_stack_slots, getcon(slot_offset, func));
+          slot_offset += 8;
+        }
+        break;
+      }
+      case APS_EnvTag:
+      case APS_VarargsTag:
+        // Nothing to do here, see right before the call for reg dupe.
+        break;
+      case APS_Invalid:
+        die("unreachable");
+    }
+  }
+
+  if (stack_usage) {
+    // The last (first in call order) thing we do is allocate the the stack
+    // space we're going to fill with temporaries.
+    emit(Osalloc, Kl, arg_stack_slots,
+         getcon(stack_usage + SHADOW_SPACE_SIZE, func), R);
+  } else {
+    // When there's no usage for temporaries, we can add this into the other
+    // alloca, but otherwise emit it separately (not storing into a reference)
+    // so that it doesn't get removed later for being useless.
+    emit(Osalloc, Kl, R, getcon(SHADOW_SPACE_SIZE, func), R);
+  }
+
+  return instr_past_args;
+}
+
+static void lower_block_return(Fn* func, Blk* block) {
+  int jmp_type = block->jmp.type;
+
+  if (!isret(jmp_type) || jmp_type == Jret0) {
+    return;
+  }
+
+  // Save the argument, and set the block to be a void return because once it's
+  // lowered it's handled by the the register/stack manipulation.
+  Ref ret_arg = block->jmp.arg;
+  block->jmp.type = Jret0;
+
+  RegisterUsage reg_usage = {0};
+
+  if (jmp_type == Jretc) {
+    Typ* type = &typ[func->retty];
+    if (type_is_by_copy(type)) {
+      assert(rtype(func->retr) == RTmp);
+      emit(Ocopy, Kl, TMP(RAX), func->retr, R);
+      emit(Oblit1, 0, R, INT(type->size), R);
+      emit(Oblit0, 0, R, ret_arg, func->retr);
+    } else {
+      emit(Oload, Kl, TMP(RAX), ret_arg, R);
+    }
+    reg_usage.rax_returned = true;
+  } else {
+    int k = jmp_type - Jretw;
+    if (is_integer_type(k)) {
+      emit(Ocopy, k, TMP(RAX), ret_arg, R);
+      reg_usage.rax_returned = true;
+    } else {
+      emit(Ocopy, k, TMP(XMM0), ret_arg, R);
+      reg_usage.xmm0_returned = true;
+    }
+  }
+  block->jmp.arg = CALL(register_usage_to_call_arg_value(reg_usage));
+}
+
+static void lower_vastart(Fn* func,
+                          RegisterUsage* param_reg_usage,
+                          Ref valist) {
+  assert(func->vararg);
+  // In varargs functions:
+  // 1. the int registers are already dumped to the shadow stack space;
+  // 2. any parameters passed in floating point registers have
+  //    been duplicated to the integer registers
+  // 3. we ensure (later) that for varargs functions we're always using an rbp
+  //    frame pointer.
+  // So, the ... argument is just indexed past rbp by the number of named values
+  // that were actually passed.
+
+  Ref offset = newtmp("abi.vastart", Kl, func);
+  emit(Ostorel, Kl, R, offset, valist);
+
+  // *8 for sizeof(u64), +16 because the return address and rbp have been pushed
+  // by the time we get to the body of the function.
+  emit(Oadd, Kl, offset, TMP(RBP),
+       getcon(param_reg_usage->num_named_args_passed * 8 + 16, func));
+}
+
+static void lower_vaarg(Fn* func, Ins* vaarg_instr) {
+  // va_list is just a void** on winx64, so load the pointer, then load the
+  // argument from that pointer, then increment the pointer to the next arg.
+  // (All emitted backwards as usual.)
+  Ref inc = newtmp("abi.vaarg.inc", Kl, func);
+  Ref ptr = newtmp("abi.vaarg.ptr", Kl, func);
+  emit(Ostorel, Kl, R, inc, vaarg_instr->arg[0]);
+  emit(Oadd, Kl, inc, ptr, getcon(8, func));
+  emit(Oload, vaarg_instr->cls, vaarg_instr->to, ptr, R);
+  emit(Oload, Kl, ptr, vaarg_instr->arg[0], R);
+}
+
+static void lower_args_for_block(Fn* func,
+                                 Blk* block,
+                                 RegisterUsage* param_reg_usage,
+                                 ExtraAlloc** pextra_alloc) {
+  // global temporary buffer used by emit. Reset to the end, and predecremented
+  // when adding to it.
+  curi = &insb[NIns];
+
+  lower_block_return(func, block);
+
+  if (block->nins) {
+    // Work backwards through the instructions, either copying them unchanged,
+    // or modifying as necessary.
+    for (Ins* instr = &block->ins[block->nins - 1]; instr >= block->ins;) {
+      switch (instr->op) {
+        case Ocall:
+          instr = lower_call(func, block, instr, pextra_alloc);
+          break;
+        case Ovastart:
+          lower_vastart(func, param_reg_usage, instr->arg[0]);
+          --instr;
+          break;
+        case Ovaarg:
+          lower_vaarg(func, instr);
+          --instr;
+          break;
+        case Oarg:
+        case Oargc:
+          die("unreachable");
+        default:
+          emiti(*instr);
+          --instr;
+          break;
+      }
+    }
+  }
+
+  // This it the start block, which is processed last. Add any allocas that
+  // other blocks needed.
+  bool is_start_block = block == func->start;
+  if (is_start_block) {
+    for (ExtraAlloc* ea = *pextra_alloc; ea; ea = ea->link) {
+      emiti(ea->instr);
+    }
+  }
+
+  // emit/emiti add instructions from the end to the beginning of the temporary
+  // global buffer. dup the final version into the final block storage.
+  block->nins = &insb[NIns] - curi;
+  idup(block, curi, block->nins);
+}
+
+static Ins* find_end_of_func_parameters(Blk* start_block) {
+  Ins* i;
+  for (i = start_block->ins; i < &start_block->ins[start_block->nins]; ++i) {
+    if (!ispar(i->op)) {
+      break;
+    }
+  }
+  return i;
+}
+
+// Copy from registers/stack into values.
+static RegisterUsage lower_func_parameters(Fn* func) {
+  // This is half-open, so end points after the last Opar.
+  Blk* start_block = func->start;
+  Ins* start_of_params = start_block->ins;
+  Ins* end_of_params = find_end_of_func_parameters(start_block);
+
+  size_t num_params = end_of_params - start_of_params;
+  ArgClass* arg_classes = alloc(num_params * sizeof(ArgClass));
+  ArgClass arg_ret = {0};
+
+  // global temporary buffer used by emit. Reset to the end, and predecremented
+  // when adding to it.
+  curi = &insb[NIns];
+
+  RegisterUsage reg_usage = {0};
+  if (func->retty >= 0) {
+    bool by_copy = type_is_by_copy(&typ[func->retty]);
+    if (by_copy) {
+      assign_register_or_stack(&reg_usage, &arg_ret, /*is_float=*/false,
+                               by_copy);
+      Ref ret_ref = newtmp("abi.ret", Kl, func);
+      emit(Ocopy, Kl, ret_ref, TMP(RCX), R);
+      func->retr = ret_ref;
+    }
+  }
+  Ref env = R;
+  classify_arguments(&reg_usage, start_of_params, end_of_params, arg_classes,
+                     &env);
+  func->reg = amd64_winabi_argregs(
+      CALL(register_usage_to_call_arg_value(reg_usage)), NULL);
+
+  // Copy from the registers or stack slots into the named parameters. Depending
+  // on how they're passed, they either need to be copied or loaded.
+  ArgClass* arg = arg_classes;
+  int reg_counter = 0;
+  uint slot_offset = SHADOW_SPACE_SIZE / 4 + 4;
+  for (Ins* instr = start_of_params; instr < end_of_params; ++instr, ++arg) {
+    switch (arg->style) {
+      case APS_Register: {
+        Ref from = register_for_arg(arg->cls, reg_counter++);
+        // If it's a struct at the IL level, we need to copy the register into
+        // an alloca so we have something to point at (same for InlineOnStack).
+        if (instr->op == Oparc) {
+          arg->ref = newtmp("abi", Kl, func);
+          emit(Ostorel, Kl, R, arg->ref, instr->to);
+          emit(Ocopy, instr->cls, arg->ref, from, R);
+          emit(Oalloc8, Kl, instr->to, getcon(arg->size, func), R);
+        } else {
+          emit(Ocopy, instr->cls, instr->to, from, R);
+        }
+        break;
+      }
+      case APS_InlineOnStack:
+        if (instr->op == Oparc) {
+          arg->ref = newtmp("abi", Kl, func);
+          emit(Ostorel, Kl, R, arg->ref, instr->to);
+          emit(Ocopy, instr->cls, arg->ref, SLOT(-slot_offset), R);
+          emit(Oalloc8, Kl, instr->to, getcon(arg->size, func), R);
+        } else {
+          emit(Ocopy, Kl, instr->to, SLOT(-slot_offset), R);
+        }
+        slot_offset += 2;
+        break;
+      case APS_CopyAndPointerOnStack:
+        emit(Oload, Kl, instr->to, SLOT(-slot_offset), R);
+        slot_offset += 2;
+        break;
+      case APS_CopyAndPointerInRegister: {
+        // Because this has to be a copy (that we own), it is sufficient to just
+        // copy the register to the target.
+        Ref from = register_for_arg(Kl, reg_counter++);
+        emit(Ocopy, Kl, instr->to, from, R);
+        break;
+      }
+      case APS_EnvTag:
+        break;
+      case APS_VarargsTag:
+      case APS_Invalid:
+        die("unreachable");
+    }
+  }
+
+  // If there was an `env`, it was passed in RAX, so copy it into the env ref.
+  if (!req(R, env)) {
+    emit(Ocopy, Kl, env, TMP(RAX), R);
+  }
+
+  int num_created_instrs = &insb[NIns] - curi;
+  int num_other_after_instrs = (int)(start_block->nins - num_params);
+  int new_total_instrs = num_other_after_instrs + num_created_instrs;
+  Ins* new_instrs = vnew(new_total_instrs, sizeof(Ins), PFn);
+  Ins* instr_p = icpy(new_instrs, curi, num_created_instrs);
+  icpy(instr_p, end_of_params, num_other_after_instrs);
+  start_block->nins = new_total_instrs;
+  start_block->ins = new_instrs;
+
+  return reg_usage;
+}
+
+// The main job of this function is to lower generic instructions into the
+// specific details of how arguments are passed, and parameters are
+// interpreted for win x64. A useful reference is
+// https://learn.microsoft.com/en-us/cpp/build/x64-calling-convention .
+//
+// Some of the major differences from SysV if you're comparing the code
+// (non-exhaustive):
+// - only 4 int and 4 float regs are used
+// - when an int register is assigned a value, its associated float register is
+//   left unused (and vice versa). i.e. there's only one counter as you assign
+//   arguments to registers.
+// - any structs that aren't 1/2/4/8 bytes in size are passed by pointer, not
+//   by copying them into the stack. So e.g. if you pass something like
+//   `struct { void*, int64_t }` by value, it first needs to be copied to
+//   another alloca (in order to maintain value semantics at the language
+//   level), then the pointer to that copy is treated as a regular integer
+//   argument (which then itself may *also* be copied to the stack in the case
+//   there's no integer register remaining.)
+// - when calling a varargs functions, floating point values must be duplicated
+//   integer registers. Along with the above restrictions, this makes varargs
+//   handling simpler for the callee than SysV.
+void amd64_winabi_abi(Fn* func) {
+  // The first thing to do is lower incoming parameters to this function.
+  RegisterUsage param_reg_usage = lower_func_parameters(func);
+
+  // This is the second larger part of the job. We walk all blocks, and rewrite
+  // instructions returns, calls, and handling of varargs into their win x64
+  // specific versions. Any other instructions are just passed through unchanged
+  // by using `emiti`.
+
+  // Skip over the entry block, and do it at the end so that our later
+  // modifications can add allocations to the start block. In particular, we
+  // need to add stack allocas for copies when structs are passed or returned by
+  // value.
+  ExtraAlloc* extra_alloc = NULL;
+  for (Blk* block = func->start->link; block; block = block->link) {
+    lower_args_for_block(func, block, &param_reg_usage, &extra_alloc);
+  }
+  lower_args_for_block(func, func->start, &param_reg_usage, &extra_alloc);
+
+  if (debug['A']) {
+    fprintf(stderr, "\n> After ABI lowering:\n");
+    printfn(func, stderr);
+  }
+}