[M3devel] m3_load/m3_store/bit_field_ref lose offset??
Jay
jay.krell at cornell.edu
Sat Dec 13 19:03:08 CET 2008
I've just spent a few hours debugging this.
The problem is strongly related to the declaration that the global segment is of size BIGGEST_ALIGNMENT.
gcc decides they can fit in a register..so then it doesn't bother adding the offset to the address, something like that.
When things work, it goes here:
op0 = adjust_address (op0, mode1, offset);
in expr.c. When things don't work, that line is skipped and then shortly below that:
if (mode1 != mode) return convert_to_mode (tmode, op0, unsignedp); return op0; <<
On systems where BIGGEST_ALIGNMENT fits in a register, e.g. apparently MIPS, counter-e.g. apparently many others, but not clearly all others. This determines whether the mode of the segment is SImode (32bit integer), DImode (64bit integer) or BLKmode (larger than any machine type?). BLKmode works. It works on 32bit MIPS because between mode1 and mode, one of them is SImode, one is DImode. On x86/AMD64/PowerPC systems, I expect it is BLKmode, since biggest alignment is set very large -- 128. I don't know why it works for SPARC.
PERHAPS I can put together a C or treelang example and get more (gcc developers) eyes on it, but I doubt it.
The following hack works around it:
C:\dev2\cm3.2\m3-sys\m3cc\gcc\gcc\m3cg>cvs -z3 diff parse.cIndex: parse.c===================================================================RCS file: /usr/cvs/cm3/m3-sys/m3cc/gcc/gcc/m3cg/parse.c,vretrieving revision 1.77diff -r1.77 parse.c2865c2865< = m3_build_type (T_struct, BIGGEST_ALIGNMENT, BIGGEST_ALIGNMENT);---> = m3_build_type (T_struct, BIGGEST_ALIGNMENT * 2, BIGGEST_ALIGNMENT);
The repro is very easy.
Just add this to m3core:
MODULE Main2 EXPORTS Main;IMPORT RTIO;
VAR a := "a";
BEGIN RTIO.PutText(a);END Main2.
The load of a will be from MM_Main2 offset 0, instead of 104 or thereabouts.
I'll commit this shortly.
- Jay
From: jay.krell at cornell.eduTo: hosking at cs.purdue.eduDate: Thu, 13 Nov 2008 00:04:32 +0000CC: m3devel at elegosoft.comSubject: Re: [M3devel] m3_load/m3_store/bit_field_ref lose offset??
> Jay, do you have access to an AMD64_DARWIN box? No, sorry, I don't yet and not sure when I will.Unless it is easy enough to run on non-Apple hardware.I'll have to debug it (m3cg) then. - Jay
CC: m3devel at elegosoft.comFrom: hosking at cs.purdue.eduTo: jay.krell at cornell.eduSubject: Re: [M3devel] m3_load/m3_store/bit_field_ref lose offset??Date: Wed, 12 Nov 2008 20:46:53 +0000Jay, do you have access to an AMD64_DARWIN box? I did get things working on that 64-bit platform. Or is it an endianness issue you are dealing with. In any case, I will not have a lot of bandwidth for doing much in the way of tackling this unless it becomes critical path for me soon.
Antony Hosking | Associate Professor | Computer Science | Purdue University
305 N. University Street | West Lafayette | IN 47907 | USA
Office +1 765 494 6001 | Mobile +1 765 427 5484
On 12 Nov 2008, at 07:57, Jay wrote:
Target.Little_endian is used to drive some target-specificity in the IR. I guess, after all, that is why you have to specify it. The IR for RTLinker.m3 is identical for MIPS64_OPENBSD and SPARC64_SOLARIS, once I set them both to unaligned_procedures, which they both need. (Tangent, I ran m3cgcat on the .ms file a few times instead of the .mc file; that makes it crash.) I can't claim SPARC64_SOLARIS is any paragon of working and correct though, by far.I should also be able to compare with SPARC64_OPENBSD, I forgot about that, and didn't have an m3cg.I only ran SPARC64_SOLARIS far enough to discover it needed unaligned procedures.I think I should fork linux-common and solaris-common off of the minimal 29k openbsd-common, rather than worry about all the stuff in solaris-2-x and linux-libc, over 100k each. Then SPARC64_SOLARIS could/would be in better shape probably. MIPS64_OPENBSD cm3 gets further with unaligned_procedures, along with the parse.c fix -- it gets to a bus error in RunMainBody.Again an alignment problem, though I think maybe memory memory trashing.The fault occurs here: desc_offset := init_depth * ADRSIZE (InitDesc); desc := init_stack + desc_offset; init_depth := desc.low_link; << InitDesc = RECORD module : RT0.ModulePtr; low_link : INTEGER; END;desc_offset appears to not be a multiple of 4. I will look into that. Long story short -- consider again the parse.c change?I admit I still haven't debugged it (or rather, the code that consumes the data produced by it). - Jay
From: jay.krell at cornell.eduTo: hosking at cs.purdue.eduCC: m3devel at elegosoft.comSubject: RE: [M3devel] m3_load/m3_store/bit_field_ref lose offset??Date: Sun, 9 Nov 2008 11:10:54 +0000I agree. I am suspicious too. I will investigate more. Thanks. - Jay
From: hosking at cs.purdue.eduTo: jay.krell at cornell.eduDate: Sun, 9 Nov 2008 10:01:58 +0000CC: m3devel at elegosoft.comSubject: Re: [M3devel] m3_load/m3_store/bit_field_ref lose offset??
Aren't the integer values supposed to be written endian-independently in the IR? I am suspicious about these differences.
On 9 Nov 2008, at 01:25, Jay wrote:
Thanks for the tip/direction. Uncertain -- m3cgcat between AMD64_LINUX and MIPS64_OPENBSD are "the same" but for two systematic diffs: - AMD64_LINUX has a bunch of added calls to check_nil. This is a "bug" in it, inoptimal but ok code, because "first readable address" is 0. It should probably be 4k. I should write a program that iterates starting at 0 to find the value. (or better yet, cm3 should have a -configure option that does this? -- for configuration values for which there is a safe default; in general, imho, the targets should be auto configured via C, assuming there is a native development environment available). - some integers are different (again, systematically) AMD64_LINUX: init_int 656 13 Int.64 init_chars 664 " typecode " init_int 680 2 Int.64 init_var 688 v.1 0 init_int 696 13 Int.64 init_chars 704 " typeid " init_int 720 2 Int.64 init_var 728 v.1 0 init_int 736 13 Int.64 init_chars 744 "RunMainBody: " init_int 760 2 Int.64 init_var 768 v.1 0 init_int 776 19 Int.64 init_chars 784 "RunMainBody: exec: " init_int 808 2 Int.64 init_var 816 v.1 0 init_int 824 17 Int.64 init_chars 832 "ExpandInitStack: "MIPS64_OPENBSD: init_chars 664 " typecode " init_int 680 8796093022208 Int.64 init_var 688 v.1 0 init_int 696 13 Int.64 init_chars 704 " typeid " init_int 720 8796093022208 Int.64 init_var 728 v.1 0 init_int 736 13 Int.64 init_chars 744 "RunMainBody: " init_int 760 8796093022208 Int.64 init_var 768 v.1 0 init_int 776 19 Int.64 init_chars 784 "RunMainBody: exec: " init_int 808 8796093022208 Int.64 init_var 816 v.1 0This 8796093022208 is 80000000000 aka 0000_0800_0000_0000, where little endian has 2. I don't know what these numbers represent.I should look at other systems -- preferably a working 64 bit big endian system, if there are any. Could be flags/type with text literals? (Btw, I always leave the TextLiteral.i3 change on my system, lowering the maximum literal length to around 4gig even on 64 bit targets, so 32 bit can bootstrap 64 bit...need to fix the compiler here to do target arithmetic instead of host arithmetic...) Thanks, - Jay
From: hosking at cs.purdue.eduTo: jay.krell at cornell.eduDate: Sat, 8 Nov 2008 22:55:54 +0000CC: m3devel at elegosoft.comSubject: Re: [M3devel] m3_load/m3_store/bit_field_ref lose offset??Have you checked to make sure that the pre-gcc IR makes sense? Does it see these values as 32-bit or 64-bit? The reason I ask is that there may be something wrong in your m3middle target specs that cause bad IR. Once we know the IR looks right then we can try and fix the problem.
Antony Hosking | Associate Professor | Computer Science | Purdue University
305 N. University Street | West Lafayette | IN 47907 | USA
Office +1 765 494 6001 | Mobile +1 765 427 5484
On 8 Nov 2008, at 15:45, Jay wrote:
truncated as usual...along with losing a lot of whitespace..Let's see...Tony, how about switching these #if's?I explain some of why below.static voidm3_store (tree v, int o, tree src_t, m3_type src_T, tree dst_t, m3_type dst_T){ tree val;#if 0 if (o != 0 || TREE_TYPE (v) != dst_t) { v = m3_build3 (BIT_FIELD_REF, dst_t, v, TYPE_SIZE (dst_t), bitsize_int (o)); }#else /* failsafe, but inefficient */ if (o != 0 || TREE_TYPE (v) != dst_t) { v = m3_build1 (ADDR_EXPR, t_addr, v); v = m3_build2 (PLUS_EXPR, t_addr, v, size_int (o / BITS_PER_UNIT)); v = m3_build1 (INDIRECT_REF, dst_t, m3_cast (build_pointer_type (dst_t), v)); }#endif TREE_THIS_VOLATILE (v) = 1; /* force this to avoid aliasing problems */ val = m3_cast (src_t, EXPR_REF (-1)); if (src_T != dst_T) { val = m3_build1 (CONVERT_EXPR, dst_t, val); } add_stmt (build2 (MODIFY_EXPR, dst_t, v, val)); EXPR_POP ();}m3_load (tree v, int o, tree src_t, m3_type src_T, tree dst_t, m3_type dst_T){#if 0 if (o != 0 || TREE_TYPE (v) != src_t) { v = m3_build3 (BIT_FIELD_REF, src_t, v, TYPE_SIZE (src_t), bitsize_int (o)); }#else /* failsafe, but inefficient */ if (o != 0 || TREE_TYPE (v) != src_t) { v = m3_build1 (ADDR_EXPR, t_addr, v); v = m3_build2 (PLUS_EXPR, t_addr, v, size_int (o / BITS_PER_UNIT)); v = m3_build1 (INDIRECT_REF, src_t, m3_cast (build_pointer_type (src_t), v)); }#endif TREE_THIS_VOLATILE (v) = 1; /* force this to avoid aliasing problems */ if (src_T != dst_T) { v = m3_build1 (CONVERT_EXPR, dst_t, v); } EXPR_PUSH (v);}I'm working on MIPS64_OPENBSD (OpenBSD on an SGI O2).In an early version it seems that many loads/stores of global variables is wrong.The "first" code affected is RTLinker_ExpandModuleTable.PROCEDURE ExpandModuleTable () = CONST InitialTableSize = 500; VAR new_mods: ADDRESS; n_bytes: INTEGER; BEGIN IF (modules = NIL) THEN (* first time... *) max_modules := InitialTableSize; modules := Cstdlib.malloc (InitialTableSize * BYTESIZE (RT0.ModulePtr)); IF (modules = NIL) THEN Cstdlib.abort (); END; ELSE n_bytes := max_modules * BYTESIZE (RT0.ModulePtr); new_mods := Cstdlib.malloc (n_bytes + n_bytes); IF (new_mods = NIL) THEN Cstdlib.abort (); END; EVAL Cstring.memcpy (new_mods, modules, n_bytes); Cstdlib.free (modules); modules := new_mods; INC (max_modules, max_modules); END; END ExpandModuleTable;It should read back modules as NIL.What happens is it reads back modules as not NIL,and then max_modules should also be zero, but is "large", leading to malloc to fail,and abort called.The output of cm3cg -y is reasonable.It indicates offset loads.What is incorrect is that in many loads and stores, but not all, theoffsets are lost.(923) begin_procedure procedure RTLinker__ExpandModuleTableRTLinker__ExpandModuleTable(924) set_source_line source line 208(925) load_nil(926) store store (M3_AJWxb1_new_mods) offset 0x0 src_t 0xb dst_t 0xb(927) set_source_line source line 207(928) set_source_line source line 210(929) load_nil(930) load m3cg_load (MM_RTLinker): offset 0x440, convert 0xb -> 0xbHere is the incorrect code:.globl RTLinker__ExpandModuleTable.stabd 46,0,0.stabn 68,0,206,.LM108.LM108:.LFBB8:.set nomips16.ent RTLinker__ExpandModuleTableRTLinker__ExpandModuleTable:.frame $fp,64,$31 # vars= 32, regs= 3/0, args= 0, gp= 0.mask 0xd0000000,-8.fmask 0x00000000,0daddiu $sp,$sp,-64sd $31,56($sp)sd $fp,48($sp)sd $28,40($sp)move $fp,$splui $28,%hi(%neg(%gp_rel(RTLinker__ExpandModuleTable)))daddu $28,$28,$25daddiu $28,$28,%lo(%neg(%gp_rel(RTLinker__ExpandModuleTable))).LBB9:.stabn 68,0,208,.LM109.LM109:sd $0,16($fp).stabn 68,0,210,.LM110.LM110:dla $2,MM_RTLinkerld $2,0($2) << WRONGbne $2,$0,.L61.stabn 68,0,212,.LM111.LM111:dla $3,MM_RTLinkerld $2,128($3)andi $2,$2,0x0ori $2,$2,0x1f4sd $2,128($3).stabn 68,0,213,.LM112.LM112:li $4,4000 # 0xfa0jal mallocsd $2,0($fp)ld $4,0($fp)dla $3,MM_RTLinkerld $2,136($3)andi $2,$2,0x0or $2,$4,$2sd $2,136($3) << OK, offset seems low, but it agrees with elsewhere .stabn 68,0,214,.LM113.LM113:dla $2,MM_RTLinkerld $2,0($2) << WRONG bne $2,$0,.L65jal abortb .L65.L61:.stabn 68,0,216,.LM114.LM114:dla $2,MM_RTLinkerld $2,0($2) << WRONG dsll $2,$2,3sd $2,8($fp).stabn 68,0,217,.LM115.LM115:ld $3,8($fp)ld $2,8($fp)daddu $2,$3,$2move $4,$2jal mallocsd $2,0($fp)ld $2,0($fp)sd $2,16($fp).stabn 68,0,218,.LM116.LM116:ld $2,16($fp)bne $2,$0,.L64jal abort.L64:.stabn 68,0,219,.LM117.LM117:ld $4,16($fp)dla $2,MM_RTLinkerld $2,0($2) << WRONGld $3,8($fp)move $5,$2move $6,$3jal memcpysd $2,0($fp).stabn 68,0,220,.LM118.LM118:dla $2,MM_RTLinkerld $2,0($2) << WRONGmove $4,$2jal free.stabn 68,0,221,.LM119.LM119:ld $4,16($fp)dla $3,MM_RTLinkerld $2,136($3) << OKandi $2,$2,0x0or $2,$4,$2sd $2,136($3) << OK .stabn 68,0,222,.LM120.LM120:dla $2,MM_RTLinkerld $3,0($2) << WRONGdla $2,MM_RTLinkerld $2,0($2) << WRONGdaddu $4,$3,$2dla $3,MM_RTLinkerld $2,128($3) << OKandi $2,$2,0x0or $2,$2,$4sd $2,128($3) << OK.L65:.LBE9:.stabn 68,0,224,.LM121.LM121:move $sp,$fpld $31,56($sp)ld $fp,48($sp)ld $28,40($sp)daddiu $sp,$sp,64j $31.end RTLinker__ExpandModuleTableHere is corrected code:RTLinker__ExpandModuleTable:.frame $fp,64,$31 # vars= 32, regs= 3/0, args= 0, gp= 0.mask 0xd0000000,-8.fmask 0x00000000,0daddiu $sp,$sp,-64sd $31,56($sp)sd $fp,48($sp)sd $28,40($sp)move $fp,$splui $28,%hi(%neg(%gp_rel(RTLinker__ExpandModuleTable)))daddu $28,$28,$25daddiu $28,$28,%lo(%neg(%gp_rel(RTLinker__ExpandModuleTable))).LBB9:.stabn 68,0,208,.LM109.LM109:sd $0,16($fp).stabn 68,0,210,.LM110.LM110:dla $2,MM_RTLinkerdaddiu $2,$2,136 << CORRECT (offset seems low, but at least it isn't zero) ld $2,0($2)bne $2,$0,.L61.stabn 68,0,212,.LM111.LM111:dla $2,MM_RTLinkerdaddiu $3,$2,128 << CORRECT li $2,500 # 0x1f4sd $2,0($3).stabn 68,0,213,.LM112.LM112:li $4,4000 # 0xfa0jal mallocsd $2,0($fp)dla $2,MM_RTLinkerdaddiu $3,$2,136 << CORRECT ld $2,0($fp)sd $2,0($3).stabn 68,0,214,.LM113.LM113:dla $2,MM_RTLinkerdaddiu $2,$2,136 << CORRECT ld $2,0($2)bne $2,$0,.L65jal abortb .L65.L61:I still dump core with this change, but I get much further.I have not tested this change at all otherwise -- i.e. on working platforms.I have not debugged through the gcc/m3cg code to see whythe offsets are only sometimes lost, or why other platformshave no problem. Is MIPS's "global pointer" unusual these days?I know SPARC has similar, and all the non-x86, non-AMD64 NT platformshad similar, but that's essentially nothing now (I'm not counting IA64).I know Linux has PLT (procedure linkage table) and GOT (global offset table).Same thing?I haven't looked much at other MIPS ports yet.http://modula3.elegosoft.com/cgi-bin/cvsweb.cgi/cm3/m3-sys/m3cc/gcc/gcc/m3cg/parse.c.diff?r1=1.7;r2=1.8is what made it look about like it does -- in particularing using BIT_FIELD_REF.1.7:static void m3_load (v, o, src_t, src_T, dest_t, dest_T) /* GCC32OK */ tree v; int o; tree src_t, dest_t; m3_type src_T, dest_T;{ if (o == 0 && TREE_TYPE (v) == src_t) { EXPR_PUSH (v); } else { tree adr = m3_build1 (ADDR_EXPR, t_addr, v); if (o != 0) { adr = m3_build2 (PLUS_EXPR, t_addr, adr, size_int (o / BITS_PER_UNIT)); } EXPR_PUSH (m3_build1 (INDIRECT_REF, src_t, m3_cast (build_pointer_type (src_t), adr))); }#if 1 if (src_T != dest_T) { EXPR_REF (-1) = m3_build1 (CONVERT_EXPR, dest_t, EXPR_REF (-1)); } if (debug_vars) { const char *name = "noname"; if (v != 0 && DECL_NAME(v) != 0) { name = IDENTIFIER_POINTER(DECL_NAME(v)); } fprintf(stderr, " m3_load (%s): offset %d, convert %d -> %d\n", name, o, src_T, dest_T); }#else if (src_T != dest_T) { if (IS_INTEGER_TYPE(dest_T) && dest_t != t_int) { EXPR_REF (-1) = m3_build1 (CONVERT_EXPR, t_int, EXPR_REF (-1)); } else if (IS_WORD_TYPE(dest_T) && dest_t != t_word) { EXPR_REF (-1) = m3_build1 (CONVERT_EXPR, t_word, EXPR_REF (-1)); } else if (IS_INTEGER_TYPE(dest_T) || IS_WORD_TYPE(dest_T)) { /* okay */ } else { fatal_error("m3load: cannot convert types: src_t %d dest_t %d\n", src_T, dest_T); } }#endif}static voidm3_store (v, o, src_t, dest_t) /* GCC32OK */ tree v; int o; tree src_t, dest_t;{ tree lhs, rhs; if (TREE_TYPE (EXPR_REF (-1)) == src_t) { rhs = EXPR_REF (-1); } else { rhs = m3_cast (src_t, EXPR_REF (-1)); } if (o == 0 && TREE_TYPE (v) == dest_t) { lhs = v; } else { tree f = make_node (FIELD_DECL); TREE_TYPE (f) = dest_t; DECL_ALIGN (f) = TYPE_ALIGN (dest_t); DECL_SIZE (f) = TYPE_SIZE (dest_t); DECL_MODE (f) = TYPE_MODE (dest_t); DECL_FIELD_OFFSET (f) = size_int (o / BITS_PER_UNIT); DECL_FIELD_BIT_OFFSET (f) = bitsize_int (o % BITS_PER_UNIT); DECL_FIELD_CONTEXT (f) = TREE_TYPE (v); lhs = m3_build2 (COMPONENT_REF, dest_t, v, f); }#if 1 expand_assignment (lhs, m3_build1 (CONVERT_EXPR, dest_t, rhs), 0, 0);#else if (dest_t == NULL) { expand_assignment (lhs, rhs, 0, 0); } else { expand_assignment (lhs, m3_build1 (CONVERT_EXPR, dest_t, rhs), 0, 0); }#endif EXPR_POP ();}1.8:static voidm3_load (tree v, int o, tree src_t, m3_type src_T, tree dst_t, m3_type dst_T) /* GCC32OK */{ if (o == 0 && TREE_TYPE (v) == src_t) { EXPR_PUSH (v); } else { EXPR_PUSH (m3_build3 (BIT_FIELD_REF, src_t, v, TYPE_SIZE (src_t), bitsize_int (o))); } if (src_T != dst_T) { EXPR_REF (-1) = m3_build1 (CONVERT_EXPR, dst_t, EXPR_REF (-1)); } if (option_vars_trace) { const char *name = "noname"; if (v != 0 && DECL_NAME(v) != 0) { name = IDENTIFIER_POINTER(DECL_NAME(v)); } fprintf(stderr, " m3_load (%s): offset %d, convert %d -> %d\n", name, o, src_T, dst_T); }}static voidm3_store (tree v, int o, tree src_t, tree dst_t) /* GCC32OK */{ tree lhs, rhs; if (TREE_TYPE (EXPR_REF (-1)) == src_t) { rhs = EXPR_REF (-1); } else { rhs = m3_cast (src_t, EXPR_REF (-1)); } if (o == 0 && TREE_TYPE (v) == dst_t) { lhs = v; } else { lhs = m3_build3 (BIT_FIELD_REF, dst_t, v, TYPE_SIZE (dst_t), bitsize_int (o)); } if (src_t != dst_t) { rhs = m3_build1 (CONVERT_EXPR, dst_t, rhs); } expand_assignment (lhs, rhs, 0, 0); EXPR_POP ();}- Jay<AMD64_LINUX_RTLinker.mc><MIPS64_OPENBSD_RTLinker.mc>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://m3lists.elegosoft.com/pipermail/m3devel/attachments/20081213/d23773fa/attachment-0001.html>
More information about the M3devel
mailing list