Bug 1773 – excessively long integer literal

An excessively long integer literal is accepted by the compiler, as show in this simple example, where a LargeInt struct instance is declared with a capacity of 256 bits: LargeInt!(256) v = 0xFFFF_FFFF_FFFF_FFFF_FFFF_FFFF_FFFF_FFFF; Under the hood, this assignment uses the opAssign overload I created for ulong. Obviously, ulong is only 64 bits in size, and only the lowest 64 bits of the "v" variable get set to 1, while the upper bits remain at 0. Also obvious is that a 256-bit integer literal is not supposed to be supported. The prioblem is that this code is accepted by the compiler, leading to beleive that all 256 bits are set to 1 when in reality they aren't. I experimented a little bit and found that 0xf_ffff_ffff_ffff_ffff also compiled, but 0x1_0000_0000_0000_0000 did not compile. This is with gdc 0.24, using dmd 1.020

I get: test.d(1): integer overflow for: ulong v = 0xFFFF_FFFF_FFFF_FFFF_FFFF_FFFF_FFFF_FFFF; I could not reproduce the problem (I don't have your definition of LargeInt()) with dmd 1.025.

Sorry, LargeInt is not necessary. I have found that the problem is only with dmd (or gdc) on linux: $ ./dmd Digital Mars D Compiler v1.025 ... $ cat t.d ulong v = 0xFFFF_FFFF_FFFF_FFFF_FFFF_FFFF_FFFF_FFFF; ulong v = 0x1_0000_0000_0000_0000; void main() {} $ ./dmd t.d t.d(2): integer overflow Line 1 was not rejected by the compiler. On Windows, both lines are rejected.

I have been able to characterize the problem further by running a piece of the lexer as a standalone C program: #include <stdio.h> #include <stdint.h> void main() { int cnt = 17; uint64_t n; int r = 16, d; printf("Walter's parsing 0xF_FFFF_FFFF_FFFF_FFFF\n"); n = 0; cnt = 17; while(cnt--) { d = 15; // 0xF_FFFF_FFFF_FFFF_FFFF; if (d >= r) break; printf("Condition value %llu < %llu\n",n * r + d,n); if (n * r + d < n) { printf("Overflow\n"); break; } n = n * r + d; printf("%0d: n = %llu\n",cnt,n); } printf("Walter's parsing 0x1_0000_0000_0000_0000\n"); n = 0; cnt = 17; while(cnt--) { if (cnt == 16) d = 1; // 0x1_0000_0000_0000_0000; else d = 0; if (d >= r) break; printf("Condition value %llu < %llu\n",n * r + d,n); if (n * r + d < n) { printf("Overflow\n"); break; } n = n * r + d; printf("%0d: n = %llu\n",cnt,n); } } In running this program which mimics the parsing of the two integer literals shown above, the problem becomes obvious. It is in the conditional. I do not know the solution yet.

Fixed dmd 1.026 and 2.010

$ cat t.d ulong v1 = 0xFFFF_FFFF_FFFF_FFFF_FFFF_FFFF_FFFF_FFFF; ulong v2 = 0x1_0000_0000_0000_0000; ulong v3 = 0x1_FFFF_FFFF_FFFF_FFFF; ulong v4 = 0x7_FFFF_FFFF_FFFF_FFFF; ulong v5 = 0x1_0000_FFFF_FFFF_FFFF; void main() {} $ ./dmd t.d t.d(1): integer overflow t.d(2): integer overflow t.d(5): integer overflow $ ./dmd -v Digital Mars D Compiler v1.026 v3 and v4 are also overflows but D does not see that. The problem is that any input stream that fills in the r*n+d will all 1s is going to be a problem. I was not able to come up with a solution different than the one explained in the strtol.c file which can be found everywhere on the internet, which uses a cutoff value between legal numbers and illegal numbers.

Fixed dmd 1.035 and 2.019