#!/bin/sh
# tms-clang -- M1 IR pipeline: clang IR, scale GEPs, optimize, normalize GEPs.
#
# Usage: tms-clang [-O0|-O2] [-I dir] -o OUT.ll IN.c
#
# -O2 (the default) scales, optimizes, then normalizes; -O0 skips the optimizer.
# The output is normalized IR for llc -march=tms34010; this wrapper does not
# invoke llc itself.
set -eu

ROOT=$(cd "$(dirname "$0")/.." && pwd)
TMS_CLANG=${TMS_CLANG:-$ROOT/build/llvm-build/bin/clang}
TMS_OPT=${TMS_OPT:-$ROOT/build/llvm-build/bin/opt}
TMS_BITADDR=${TMS_BITADDR:-$ROOT/tools/tms-bitaddr/tms-bitaddr}

OPTIMIZE=1
INCLUDES=
OUT=
IN=

usage() {
    echo "usage: $0 [-O0|-O2] [-I dir] -o OUT.ll IN.c" >&2
    exit 2
}

while [ $# -gt 0 ]; do
    case "$1" in
    -O0) OPTIMIZE=0; shift ;;
    -O2) OPTIMIZE=1; shift ;;
    -I) [ $# -ge 2 ] || usage; INCLUDES="$INCLUDES -I $2"; shift 2 ;;
    -I*) INCLUDES="$INCLUDES $1"; shift ;;
    -o) [ $# -ge 2 ] || usage; OUT=$2; shift 2 ;;
    -*) usage ;;
    *) [ -z "$IN" ] || usage; IN=$1; shift ;;
    esac
done

[ -n "$IN" ] || usage
[ -n "$OUT" ] || usage

TMP=$(mktemp -d "${TMPDIR:-/tmp}/tms-clang.XXXXXX")
trap 'rm -rf "$TMP"' EXIT HUP INT TERM

# shellcheck disable=SC2086
"$TMS_CLANG" --target=tms34010-unknown-none -O0 -Xclang -disable-O0-optnone \
    -S -emit-llvm -ffreestanding -fno-builtin -nostdinc $INCLUDES \
    -o "$TMP/step1.ll" "$IN"

"$TMS_BITADDR" --scale "$TMP/step1.ll" "$TMP/scaled.ll" 2>/dev/null

if [ "$OPTIMIZE" -eq 1 ]; then
    # This target has no i64 and no multiply-high, so IndVarSimplify's
    # exit-value replacement (which rewrites a loop's final value into a
    # closed-form i33 multiply, e.g. n*(n+1)/2 for sumto) produces IR that
    # llc cannot select.  -replexitval is an option of IndVarSimplify; disabling
    # only that rewrite keeps the loop, which is what ordinary -O2 code needs.
    # This is a driver policy, not a target TTI policy: opt runs without a
    # target machine here.
    # `-switch-to-lookup=false` is a driver policy, not a target feature: the
    # optimizer's switch-to-lookup-table transform creates a *new* byte-offset
    # GEP after --scale, and --normalize can only treat late GEP offsets as
    # bit offsets.  Such a GEP would index the table one bit per slot (an 8x
    # address bug).  The backend also has no jump-table lowering, so keeping
    # the compare/branch switch is correct.
    "$TMS_OPT" -passes='default<O2>' -replexitval=never \
        -switch-to-lookup=false -S "$TMP/scaled.ll" -o "$TMP/opt.ll"
    NORMALIZE_IN="$TMP/opt.ll"
else
    NORMALIZE_IN="$TMP/scaled.ll"
fi

mkdir -p "$(dirname "$OUT")"
"$TMS_BITADDR" --normalize "$NORMALIZE_IN" "$OUT" 2>/dev/null