Bug 13485 – FP wrong-code with -O

Test https://github.com/D-Programming-Language/dmd/pull/3992

Pasting in code here: --- //Optimization check F foo(F)(F c, F d) { c += d; c += d; return c; } void test0() { alias F = float; enum F d = (cast(F)(2)) ^^ (F.max_exp - 1); assert(foo(-d, d) == d); } void test1() { alias F = double; enum F d = (cast(F)(2)) ^^ (F.max_exp - 1); assert(foo(-d, d) == d); } void test2() { alias F = real; enum F d = (cast(F)(2)) ^^ (F.max_exp - 1); assert(foo(-d, d) == d); } void main() { test0(); test1(); test2(); import core.stdc.stdio; printf("Success\n"); } ---

FYI, this looks more like an x87 problem, rather than compiler. And it's not just D that is affected by this. For instance, this bug in gcc: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=323 Though they did eventually fix it in implementing C99-conformant (kinda) excess precision. http://marc.info/?l=gcc-patches&m=122576450613005&w=2

No, (In reply to Iain Buclaw from comment #2) > FYI, this looks more like an x87 problem, rather than compiler. And it's > not just D that is affected by this. > > For instance, this bug in gcc: > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=323 > > > Though they did eventually fix it in implementing C99-conformant (kinda) > excess precision. > > http://marc.info/?l=gcc-patches&m=122576450613005&w=2 I have looked into disassembler code. It is Not problem with x87 0. dmd do this optimization d+=d and c} 1. float and double on 64bit are using xmm registers even in dmd. 2. It works fine with LDC.

No, (In reply to Iain Buclaw from comment #2) > FYI, this looks more like an x87 problem, rather than compiler. And it's > not just D that is affected by this. > > For instance, this bug in gcc: > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=323 > > > Though they did eventually fix it in implementing C99-conformant (kinda) > excess precision. > > http://marc.info/?l=gcc-patches&m=122576450613005&w=2 I am sorry for previous not finished comment. I have looked into disassembler code. It is Not problem with x87 0. dmd do this optimization {d+=d; c+=d} But this is correct only when flags like -ffast-math was enabled. 1. float and double on 64bit are using xmm registers even in dmd. 2. It works fine with LDC.

For example current std.math.log implementation (from CEPHES library if I am not wrong) needs IEEE754 standard. If there are bugs like current then log function can be incorrect. log2 contains code: z = x - 0.5; z -= 0.5; This should not be optimized. Am sorry for my English.

(In reply to Илья Ярошенко from comment #5) > For example current std.math.log implementation (from CEPHES library if I am > not wrong) needs IEEE754 standard. > > If there are bugs like current then log function can be incorrect. > > log2 contains code: > > z = x - 0.5; > z -= 0.5; > > This should not be optimized. > DMD does this: return yl2x(x, LN2); Which translates to: fyl2x ST1(x), ST(LN2)

> DMD does this: > > return yl2x(x, LN2); > > Which translates to: > > fyl2x ST1(x), ST(LN2) My dump: dmd -O -m64 -c test7.d (fails with -run) objdump -d test7.o > test7.asm 0000000000000000 <_D5test710__T3fooTfZ3fooFNaNbNiNfffZf>: 0: 55 push %rbp 1: 48 8b ec mov %rsp,%rbp 4: f3 0f 10 e0 movss %xmm0,%xmm4 //xmm4 = xmm0 = d 8: f3 0f 10 d9 movss %xmm1,%xmm3 //xmm3 = xmm1 = c c: f3 0f 58 c4 addss %xmm4,%xmm0 //d+=d (incorrect) 10: f3 0f 58 d8 addss %xmm0,%xmm3 //c+=d 14: f3 0f 10 c3 movss %xmm3,%xmm0 // 18: 5d pop %rbp 19: c3 retq 1a: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1) And without optimization: dmd -m64 -c test7.d (Success with -run) objdump -d test7.o > test7.asm 0000000000000000 <_D5test710__T3fooTfZ3fooFNaNbNiNfffZf>: 0: 55 push %rbp 1: 48 8b ec mov %rsp,%rbp 4: 48 83 ec 10 sub $0x10,%rsp 8: f3 0f 11 4d f8 movss %xmm1,-0x8(%rbp) d: f3 0f 10 4d f8 movss -0x8(%rbp),%xmm1 12: f3 0f 58 c8 addss %xmm0,%xmm1 //c+=d 16: f3 0f 11 4d f8 movss %xmm1,-0x8(%rbp) 1b: f3 0f 10 55 f8 movss -0x8(%rbp),%xmm2 20: f3 0f 58 d0 addss %xmm0,%xmm2 //c+=d 24: f3 0f 11 55 f8 movss %xmm2,-0x8(%rbp) 29: f3 0f 10 45 f8 movss -0x8(%rbp),%xmm0 2e: c9 leaveq 2f: c3 retq

Please test your original code against this pull. https://github.com/D-Programming-Language/dmd/pull/4394

Commits pushed to master at https://github.com/D-Programming-Language/dmd https://github.com/D-Programming-Language/dmd/commit/5b10f22e044a89b66bbceebb00dd342987b9d29b Fix Issue 13485 - FP wrong-code with -O https://github.com/D-Programming-Language/dmd/commit/527b3cfc1b018feecd4ee8c037dda35f105ed853 Merge pull request #4394 from yebblies/issue13485 Issue 13485 - FP wrong-code with -O

Commits pushed to https://github.com/D-Programming-Language/dmd https://github.com/D-Programming-Language/dmd/commit/5b10f22e044a89b66bbceebb00dd342987b9d29b Fix Issue 13485 - FP wrong-code with -O https://github.com/D-Programming-Language/dmd/commit/527b3cfc1b018feecd4ee8c037dda35f105ed853 Merge pull request #4394 from yebblies/issue13485