Bug 17523 – Sporadic ICEs with inline asm

With a 64-bit dmd.exe (as built using win64.mak), some inline asm instructions can randomly cause an ICE. For example, consider the given file: ////// test.d ///// void fun() { asm { fstp ST(0); } } /////////////////// Compiling the file as usual (dmd.exe -c test.d) can result in any of the following results: 63% chance: the file compiles successfully 25% chance: "Internal error: ddmd\backend\cod3.c 6869" 6% chance: "FLunde Internal error: ddmd\backend\cod3.c 5527" 6% chance: "FLunde Internal error: ddmd\backend\cod3.c 6757" This affects building Druntime and Phobos - building them will likely fail with one of the above errors, or (less probably) succeed. Introduced in https://github.com/dlang/dmd/pull/6379

My advice revert the PR. It's not like iasm is a bottleneck in most cases.

Cannot duplicate with the Win32 build of DMD.

(In reply to Walter Bright from comment #2) > Cannot duplicate with the Win32 build of DMD. Some iasm test. It fails on my linux system sometimes. There is definitely something wrong. Maybe try compiling the code about 100-200 times ?

(In reply to Walter Bright from comment #2) > Cannot duplicate with the Win32 build of DMD. It is not reproduced with a 32-bit dmd.exe; you must make a Win64 build as the issue description says.

25% chance: "Internal error: ddmd\backend\cod3.c 6869" 6% chance: "FLunde Internal error: ddmd\backend\cod3.c 5527" 6% chance: "FLunde Internal error: ddmd\backend\cod3.c 6757" These lines are nowhere near asserts in the current HEAD. Can you try with HEAD?

(In reply to uplink.coder from comment #3) > Some iasm test. It fails on my linux system sometimes. > There is definitely something wrong. > Maybe try compiling the code about 100-200 times ? I tried a 64 bit dmd build on Linux, several thousand times. No errors.

Ran it under valgrind on Linux. No errors.

(In reply to Walter Bright from comment #5) > 25% chance: "Internal error: ddmd\backend\cod3.c 6869" > 6% chance: "FLunde Internal error: ddmd\backend\cod3.c 5527" > 6% chance: "FLunde Internal error: ddmd\backend\cod3.c 6757" > > These lines are nowhere near asserts in the current HEAD. Can you try with > HEAD? At the time I couldn't due to issue 17522. Here are the line numbers with HEAD: FLunde Internal error: ddmd\backend\cod3.c 5500 FLunde Internal error: ddmd\backend\cod3.c 6730 Internal error: ddmd\backend\cod3.c 6842

Also tried memset'ing the stack variables opnd1 .. opnd4 to 0xFF at the close of asmSemantic(). No errors. There is no host dependent code in iasm.d. So this is a bit mystifying.

(In reply to Vladimir Panteleev from comment #8) > At the time I couldn't due to issue 17522. 17522 has been fixed. > Here are the line numbers with HEAD: > > FLunde Internal error: ddmd\backend\cod3.c 5500 > FLunde Internal error: ddmd\backend\cod3.c 6730 > Internal error: ddmd\backend\cod3.c 6842 These suggest a bad opcode in code.Iop. At the end of asmSemantic(), I suggest inserting: printf("asmcode: %02x\n", s.asmcode.Iop); and at the beginning of pinholeopt() before L1: c->print();

Here is the output with the changes you requested from 10 runs: asmcode: e0 code 0000019B245D80B8: nxt=0000000000000000 op=0xE0 flg=40 FLunde code 0000019B245D80F0: nxt=0000000000000000 op=0xC3 code 0000019B245D80F0: nxt=0000000000000000 op=0xC3 Internal error: ddmd\backend\cod3.c 6843 === Exited with status 1 === asmcode: 40 code 00000277E13B8EE8: nxt=0000000000000000 op=0x40 flg=40 code 00000277E13B8F20: nxt=0000000000000000 op=0xC3 code 00000277E13B8F20: nxt=0000000000000000 op=0xC3 === Exited with status 0 === asmcode: d0 code 000001E109C46C18: nxt=0000000000000000 op=0xD0 flg=40 rm=0xD8=3,3,0 code 000001E109C46C50: nxt=0000000000000000 op=0xC3 code 000001E109C46C50: nxt=0000000000000000 op=0xC3 === Exited with status 0 === asmcode: 60 code 000002285BE98B38: nxt=0000000000000000 op=0x60 flg=40 code 000002285BE98B70: nxt=0000000000000000 op=0xC3 code 000002285BE98B70: nxt=0000000000000000 op=0xC3 === Exited with status 0 === asmcode: c0 code 00000244E18AC278: nxt=0000000000000000 op=0xC0 flg=40 rm=0xD8=3,3,0 FLunde code 00000244E18AC2B0: nxt=0000000000000000 op=0xC3 code 00000244E18AC2B0: nxt=0000000000000000 op=0xC3 Internal error: ddmd\backend\cod3.c 6843 === Exited with status 1 === asmcode: 70 code 000002295B099D28: nxt=0000000000000000 op=0x70 flg=40 FLunde FLunde Internal error: ddmd\backend\cod3.c 5501 === Exited with status 1 === asmcode: 80 code 0000022F4FE57D38: nxt=0000000000000000 op=0x80 flg=40 rm=0xD8=3,3,0 FLunde code 0000022F4FE57D70: nxt=0000000000000000 op=0xC3 code 0000022F4FE57D70: nxt=0000000000000000 op=0xC3 Internal error: ddmd\backend\cod3.c 6843 === Exited with status 1 === asmcode: 50 code 000001E7F37DC868: nxt=0000000000000000 op=0x50 flg=40 code 000001E7F37DC8A0: nxt=0000000000000000 op=0xC3 code 000001E7F37DC8A0: nxt=0000000000000000 op=0xC3 === Exited with status 0 === asmcode: 60 code 00000292EE418E48: nxt=0000000000000000 op=0x60 flg=40 code 00000292EE418E80: nxt=0000000000000000 op=0xC3 code 00000292EE418E80: nxt=0000000000000000 op=0xC3 === Exited with status 0 === asmcode: 00 code 0000025538918CE8: nxt=0000000000000000 op=0x00 flg=40 rm=0xD8=3,3,0 code 0000025538918D20: nxt=0000000000000000 op=0xC3 code 0000025538918D20: nxt=0000000000000000 op=0xC3 === Exited with status 0 === If you're having trouble getting win64.mak to work, the latest version of Digger should be able to build a 64-bit DMD on any Windows system with no other dependencies. Run "digger checkout", edit the source files under work/repo/dmd/src/ddmd, and run "digger -c build.components.dmd.dmdModel=64 rebuild --model=32 --without=phobos --without=druntime --without=rdmd --without=phobos-includes" to build a 64-bit dmd.exe under work/result/bin.

Hmmm, the opcode should be DD: DD D8 fstp ST Which suggests the problem is confined to iasm.d

So the way forward is to follow the flow backwards to where c.Iop is being set and print that. I'd start with line 1384: pc.Iop = opcode; and insert: printf("test1: opcode = 0x%x\n", opcode); Ain't debugging fun? :-)

(In reply to Walter Bright from comment #13) > Ain't debugging fun? :-) Doing it through Bugzilla comments isn't very effective, though. Have you tried building a 64-bit DMD using win64.mak or Digger yet?

(In reply to Walter Bright from comment #13) > printf("test1: opcode = 0x%x\n", opcode); After adding this printf, the bug can no longer be reproduced. Since that pointed at a codegen bug in the host compiler, I tried building DMD with 2.074.1 (instead of Digger's default of 2.070.2 for Win64), and I can no longer reproduce the bug either. I'm going to bisect what caused the behaviour change, just to ensure that the bug disappeared because of a codegen fix, and not because some random change made it no longer manifest for that test case.

I haven't tried building Win64 dmd. I figured that would be time consuming :-)

(In reply to Vladimir Panteleev from comment #15) > Since that pointed at a codegen bug in the host compiler, I tried building > DMD with 2.074.1 (instead of Digger's default of 2.070.2 for Win64), and I > can no longer reproduce the bug either. > > I'm going to bisect what caused the behaviour change, just to ensure that > the bug disappeared because of a codegen fix, and not because some random > change made it no longer manifest for that test case. That's a good plan. There's a lot that has changed from 70 to 74.

(In reply to Vladimir Panteleev from comment #15) > I'm going to bisect what caused the behaviour change, just to ensure that > the bug disappeared because of a codegen fix, and not because some random > change made it no longer manifest for that test case. Done. Bisecting the version to bootstrap DMD really put Digger to the test, and I needed to make some upgrades :) The bug was fixed in this PR: https://github.com/dlang/dmd/pull/5924 It does seem to be a fix in the backend, but it fixes an ICE, not resulting codegen. Does that make sense to you?

> It does seem to be a fix in the backend, but it fixes an ICE, not resulting codegen. Does that make sense to you? Yes, because the compiler itself was corrupted by the Win64 build.