Codegen C - Fix a minor ABI issue for x86, fix emulated i128 a bit

1 files changed, 38 insertions, 9 deletions

diff --git a/src/trans/codegen_c.cpp b/src/trans/codegen_c.cpp
index 4eb3509c..2ab4ff35 100644
--- a/src/trans/codegen_c.cpp
+++ b/src/trans/codegen_c.cpp
@@ -259,7 +259,7 @@ namespace {
             {
                 m_of
                     << "typedef struct { char _d; } tUNIT;\n"
-                    << "typedef struct { char _d; } tBANG;\n"
+                    << "typedef char tBANG;\n"
                     << "typedef struct { char _d; } tTYPEID;\n"
                     ;
             }
@@ -417,8 +417,21 @@ namespace {
                     << "static inline int cmp128(uint128_t a, uint128_t b) { if(a.hi != b.hi) return a.hi < b.hi ? -1 : 1; if(a.lo != b.lo) return a.lo < b.lo ? -1 : 1; return 0; }\n"
                     << "static inline bool add128_o(uint128_t a, uint128_t b, uint128_t* o) { o->lo = a.lo + b.lo; o->hi = a.hi + b.hi + (o->lo < a.lo ? 1 : 0); return (o->hi >= a.hi); }\n"
                     << "static inline bool sub128_o(uint128_t a, uint128_t b, uint128_t* o) { o->lo = a.lo - b.lo; o->hi = a.hi - b.hi - (o->lo < a.lo ? 1 : 0); return (o->hi <= a.hi); }\n"
-                    << "static inline bool mul128_o(uint128_t a, uint128_t b, uint128_t* o) { abort(); }\n"
-                    << "static inline bool div128_o(uint128_t a, uint128_t b, uint128_t* q, uint128_t* r) { abort(); }\n"
+                    // Serial shift+add
+                    << "static inline bool mul128_o(uint128_t a, uint128_t b, uint128_t* o) {"
+                    <<  "bool of = false;"
+                    <<  "o->hi = 0; o->lo = 0;"
+                    <<  "for(int i=0;i<128;i++){"
+                    <<   "if((1ull<<(i&63))&(i>64?a.hi:a.lo)) of |= add128_o(*o, b, o);"
+                    <<   "b.hi = (b.hi << 1) | (b.lo >> 63);"
+                    <<   "b.lo = (b.lo << 1);"
+                    <<  "}"
+                    <<  "return of;"
+                    <<  "}\n"
+                    // TODO: Long division
+                    << "static inline bool div128_o(uint128_t a, uint128_t b, uint128_t* q, uint128_t* r) {"
+                    <<  "abort();"
+                    <<  "}\n"
                     << "static inline uint128_t add128(uint128_t a, uint128_t b) { uint128_t v; add128_o(a, b, &v); return v; }\n"
                     << "static inline uint128_t sub128(uint128_t a, uint128_t b) { uint128_t v; sub128_o(a, b, &v); return v; }\n"
                     << "static inline uint128_t mul128(uint128_t a, uint128_t b) { uint128_t v; mul128_o(a, b, &v); return v; }\n"
@@ -427,8 +440,8 @@ namespace {
                     << "static inline uint128_t and128(uint128_t a, uint128_t b) { uint128_t v = { a.lo & b.lo, a.hi & b.hi }; return v; }\n"
                     << "static inline uint128_t or128 (uint128_t a, uint128_t b) { uint128_t v = { a.lo | b.lo, a.hi | b.hi }; return v; }\n"
                     << "static inline uint128_t xor128(uint128_t a, uint128_t b) { uint128_t v = { a.lo ^ b.lo, a.hi ^ b.hi }; return v; }\n"
-                    << "static inline uint128_t shl128(uint128_t a, uint32_t b) { uint128_t v; if(b < 64) { v.lo = a.lo << b; v.hi = (a.hi << b) | (a.lo >> (64 - b)); } else { v.hi = a.lo << (b - 64); v.lo = 0; } return v; }\n"
-                    << "static inline uint128_t shr128(uint128_t a, uint32_t b) { uint128_t v; if(b < 64) { v.lo = (a.lo >> b)|(a.hi << (64 - b)); v.hi = a.hi >> b; } else { v.lo = a.hi >> (b - 64); v.hi = 0; } return v; }\n"
+                    << "static inline uint128_t shl128(uint128_t a, uint32_t b) { uint128_t v; if(b == 0) { return a; } else if(b < 64) { v.lo = a.lo << b; v.hi = (a.hi << b) | (a.lo >> (64 - b)); } else { v.hi = a.lo << (b - 64); v.lo = 0; } return v; }\n"
+                    << "static inline uint128_t shr128(uint128_t a, uint32_t b) { uint128_t v; if(b == 0) { return a; } else if(b < 64) { v.lo = (a.lo >> b)|(a.hi << (64 - b)); v.hi = a.hi >> b; } else { v.lo = a.hi >> (b - 64); v.hi = 0; } return v; }\n"
                     << "static inline uint128_t popcount128(uint128_t a) { uint128_t v = { __builtin_popcount(a.lo) + __builtin_popcount(a.hi), 0 }; return v; }\n"
                     << "static inline uint128_t __builtin_bswap128(uint128_t v) { uint128_t rv = { __builtin_bswap64(v.hi), __builtin_bswap64(v.lo) }; return rv; }\n"
                     << "static inline uint128_t intrinsic_ctlz_u128(uint128_t v) {\n"
@@ -445,8 +458,24 @@ namespace {
                     << "static inline int cmp128s(int128_t a, int128_t b) { if(a.hi != b.hi) return (int64_t)a.hi < (int64_t)b.hi ? -1 : 1; if(a.lo != b.lo) return a.lo < b.lo ? -1 : 1; return 0; }\n"
                     << "static inline bool add128s_o(int128_t a, int128_t b, int128_t* o) { o->lo = a.lo + b.lo; o->hi = a.hi + b.hi + (o->lo < a.lo ? 1 : 0); return (o->hi >= a.hi); }\n"
                     << "static inline bool sub128s_o(int128_t a, int128_t b, int128_t* o) { o->lo = a.lo - b.lo; o->hi = a.hi - b.hi - (o->lo < a.lo ? 1 : 0); return (o->hi <= a.hi); }\n"
-                    << "static inline bool mul128s_o(int128_t a, int128_t b, int128_t* o) { abort(); }\n"
-                    << "static inline bool div128s_o(int128_t a, int128_t b, int128_t* q, int128_t* r) { abort(); }\n"
+                    << "static inline bool mul128s_o(int128_t a, int128_t b, int128_t* o) {"
+                    << " bool sgna = a.hi & (1ull<<63);"
+                    << " bool sgnb = b.hi & (1ull<<63);"
+                    << " if(sgna) { a.hi = ~a.hi; a.lo = ~a.lo; a.lo += 1; if(a.lo == 0) a.hi += 1; }"
+                    << " if(sgnb) { b.hi = ~b.hi; b.lo = ~b.lo; b.lo += 1; if(b.lo == 0) b.hi += 1; }"
+                    << " bool rv = mul128_o(*(uint128_t*)&a, *(uint128_t*)&b, (uint128_t*)o);"
+                    << " if(sgnb != sgnb) { o->hi = ~o->hi; o->lo = ~o->lo; o->lo += 1; if(o->lo == 0) o->hi += 1; }"
+                    << " return rv;"
+                    << " }\n"
+                    << "static inline bool div128s_o(int128_t a, int128_t b, int128_t* q, int128_t* r) {"
+                    << " bool sgna = a.hi & (1ull<<63);"
+                    << " bool sgnb = b.hi & (1ull<<63);"
+                    << " if(sgna) { a.hi = ~a.hi; a.lo = ~a.lo; a.lo += 1; if(a.lo == 0) a.hi += 1; }"
+                    << " if(sgnb) { b.hi = ~b.hi; b.lo = ~b.lo; b.lo += 1; if(b.lo == 0) b.hi += 1; }"
+                    << " bool rv = div128_o(*(uint128_t*)&a, *(uint128_t*)&b, (uint128_t*)q, (uint128_t*)r);"
+                    << " if(sgnb != sgnb) { r->hi = ~r->hi; r->lo = ~r->lo; r->lo += 1; if(r->lo == 0) r->hi += 1; }"
+                    << " return rv;"
+                    << " }\n"
                     << "static inline int128_t add128s(int128_t a, int128_t b) { int128_t v; add128s_o(a, b, &v); return v; }\n"
                     << "static inline int128_t sub128s(int128_t a, int128_t b) { int128_t v; sub128s_o(a, b, &v); return v; }\n"
                     << "static inline int128_t mul128s(int128_t a, int128_t b) { int128_t v; mul128s_o(a, b, &v); return v; }\n"
@@ -455,8 +484,8 @@ namespace {
                     << "static inline int128_t and128s(int128_t a, int128_t b) { int128_t v = { a.lo & b.lo, a.hi & b.hi }; return v; }\n"
                     << "static inline int128_t or128s (int128_t a, int128_t b) { int128_t v = { a.lo | b.lo, a.hi | b.hi }; return v; }\n"
                     << "static inline int128_t xor128s(int128_t a, int128_t b) { int128_t v = { a.lo ^ b.lo, a.hi ^ b.hi }; return v; }\n"
-                    << "static inline int128_t shl128s(int128_t a, uint32_t b) { int128_t v; if(b < 64) { v.lo = a.lo << b; v.hi = (a.hi << b) | (a.lo >> (64 - b)); } else { v.hi = a.lo << (b - 64); v.lo = 0; } return v; }\n"
-                    << "static inline int128_t shr128s(int128_t a, uint32_t b) { int128_t v; if(b < 64) { v.lo = (a.lo >> b)|(a.hi << (64 - b)); v.hi = a.hi >> b; } else { v.lo = a.hi >> (b - 64); v.hi = 0; } return v; }\n"
+                    << "static inline int128_t shl128s(int128_t a, uint32_t b) { int128_t v; if(b == 0) { return a; } else if(b < 64) { v.lo = a.lo << b; v.hi = (a.hi << b) | (a.lo >> (64 - b)); } else { v.hi = a.lo << (b - 64); v.lo = 0; } return v; }\n"
+                    << "static inline int128_t shr128s(int128_t a, uint32_t b) { int128_t v; if(b == 0) { return a; } else if(b < 64) { v.lo = (a.lo >> b)|(a.hi << (64 - b)); v.hi = a.hi >> b; } else { v.lo = a.hi >> (b - 64); v.hi = 0; } return v; }\n"
                     ;
             }
             else