Message ID | CANeU7Q=irkzkaKdaZT-0POrhj73ac3X8Bk-MW+jvFGtr1h5z0w@mail.gmail.com (mailing list archive) |
---|---|
State | Superseded, archived |
Headers | show |
Hi Chris, On 13 August 2017 at 05:18, Christopher Li <sparse@chrisli.org> wrote: > This is actually easier than I though. > It is also very simple to reason about. > > If you are not looking at the size field at all. > It is exactly the same as before, except we > create a few more pseudo for the same value > that have different size. > > The size just tag along the constant pseudo. > Which did not change any way. > I assume that you have not yet applied this approach to Sparse-LLVM and tested that it works for var arg calls? I would suggest completing the changes so that it can be tested thoroughly. I personally don't have a preference one way or the other which approach we use. The push approach has the benefit that it makes clear what arguments are being passed in a call. Previously this information was not revealed in the Sparse linearized dump. It would be a shame to lose that information - perhaps you could still dump the arguments along with a call? Second benefit to me is that I have already incorporated this into dmrC so changing it will mean rework for me. BTW one could also have a more restricted 'push' solution that is specific to var arg calls - i.e. only capture the arguments when calling a var arg function, and avoid putting the push instructions into the linear stream. Regards Dibyendu -- To unsubscribe from this list: send the line "unsubscribe linux-sparse" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Mon, Aug 14, 2017 at 1:11 PM, Dibyendu Majumdar <mobile@majumdar.org.uk> wrote: > I assume that you have not yet applied this approach to Sparse-LLVM > and tested that it works for var arg calls? I would suggest completing > the changes so that it can be tested thoroughly. No. this is just one night's hack to make the plumbing part work. Provide a size for the VALUE pseudo. I haven't have a chance to integrate into sparse-llvm. > The push approach has the benefit that it makes clear what arguments > are being passed in a call. Previously this information was not > revealed in the Sparse linearized dump. It would be a shame to lose > that information - perhaps you could still dump the arguments along > with a call? Are you sure? Here is what I have for RC5: ./test-linearize cse.c .L414: load.64 %r1052 <- 0[insn] call.64 %r1053 <- try_to_cse, %arg1, %r1044(last), %r1052 store.64 %r1053 -> 0[insn] phisrc.64 %phi474 <- %r1053 br .L413 You can see exactly what is the calling arguments. There is an pseudo_list for the arguments. struct /* call */ { pseudo_t func; struct pseudo_list *arguments; <============== struct symbol *fntype; }; The only thing you want, which is not there is the VALUE pseudo_size. I add that in this patch. I haven't take a deep look into the push instruction patch yet. I image this is less impact on the IR and get the job done. With less redundant information in the IR. The instruction stream is more compact as well. Chris -- To unsubscribe from this list: send the line "unsubscribe linux-sparse" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Chris, On 14 August 2017 at 18:23, Christopher Li <sparse@chrisli.org> wrote: > On Mon, Aug 14, 2017 at 1:11 PM, Dibyendu Majumdar > <mobile@majumdar.org.uk> wrote: >> The push approach has the benefit that it makes clear what arguments >> are being passed in a call. Previously this information was not >> revealed in the Sparse linearized dump. It would be a shame to lose >> that information - perhaps you could still dump the arguments along >> with a call? > > Are you sure? > Here is what I have for RC5: > ./test-linearize cse.c > > .L414: > load.64 %r1052 <- 0[insn] > call.64 %r1053 <- try_to_cse, %arg1, %r1044(last), %r1052 > store.64 %r1053 -> 0[insn] > phisrc.64 %phi474 <- %r1053 > br .L413 > > You can see exactly what is the calling arguments. There is an pseudo_list > for the arguments. > Ah okay - I don't see that in dmrC output. It must have been removed as part of the push change. Regards Dibyendu -- To unsubscribe from this list: send the line "unsubscribe linux-sparse" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/flow.c b/flow.c index 6b2c879..fa5d31c 100644 --- a/flow.c +++ b/flow.c @@ -517,7 +517,7 @@ found: if (!local) return 0; check_access(insn); - convert_load_instruction(insn, value_pseudo(0)); + convert_load_instruction(insn, value_pseudo(insn->type, 0)); return 1; } diff --git a/linearize.c b/linearize.c index ba76397..2aa3acb 100644 --- a/linearize.c +++ b/linearize.c @@ -785,22 +785,25 @@ static pseudo_t symbol_pseudo(struct entrypoint *ep, struct symbol *sym) return pseudo; } -pseudo_t value_pseudo(long long val) +pseudo_t value_pseudo(struct symbol *type, long long val) { #define MAX_VAL_HASH 64 static struct pseudo_list *prev[MAX_VAL_HASH]; int hash = val & (MAX_VAL_HASH-1); struct pseudo_list **list = prev + hash; + int size = type ? type->bit_size : value_size(val); pseudo_t pseudo; + FOR_EACH_PTR(*list, pseudo) { - if (pseudo->value == val) + if (pseudo->value == val && pseudo->size == size) return pseudo; } END_FOR_EACH_PTR(pseudo); pseudo = __alloc_pseudo(0); pseudo->type = PSEUDO_VAL; pseudo->value = val; + pseudo->size = size; add_pseudo(list, pseudo); /* Value pseudos have neither nr, usage nor def */ @@ -954,10 +957,10 @@ static pseudo_t linearize_store_gen(struct entrypoint *ep, unsigned long long mask = (1ULL << size) - 1; if (shift) { - store = add_binary_op(ep, ad->source_type, OP_SHL, value, value_pseudo(shift)); + store = add_binary_op(ep, ad->source_type, OP_SHL, value, value_pseudo(ctype, shift)); mask <<= shift; } - orig = add_binary_op(ep, ad->source_type, OP_AND, orig, value_pseudo(~mask)); + orig = add_binary_op(ep, ad->source_type, OP_AND, orig, value_pseudo(ctype, ~mask)); store = add_binary_op(ep, ad->source_type, OP_OR, orig, store); } add_store(ep, ad, store); @@ -1002,7 +1005,7 @@ static pseudo_t linearize_load_gen(struct entrypoint *ep, struct access_data *ad pseudo_t new = add_load(ep, ad); if (ctype->bit_offset) { - pseudo_t shift = value_pseudo(ctype->bit_offset); + pseudo_t shift = value_pseudo(ctype, ctype->bit_offset); pseudo_t newval = add_binary_op(ep, ad->source_type, OP_LSR, new, shift); new = newval; } @@ -1034,7 +1037,7 @@ static pseudo_t linearize_inc_dec(struct entrypoint *ep, struct expression *expr return VOID; old = linearize_load_gen(ep, &ad); - one = value_pseudo(expr->op_value); + one = value_pseudo(expr->ctype, expr->op_value); new = add_binary_op(ep, expr->ctype, op, old, one); linearize_store_gen(ep, new, &ad); finish_address_gen(ep, &ad); @@ -1073,7 +1076,7 @@ static pseudo_t linearize_regular_preop(struct entrypoint *ep, struct expression case '+': return pre; case '!': { - pseudo_t zero = value_pseudo(0); + pseudo_t zero = value_pseudo(expr->ctype, 0); return add_binary_op(ep, expr->ctype, OP_SET_EQ, pre, zero); } case '~': @@ -1165,7 +1168,7 @@ static inline pseudo_t add_convert_to_bool(struct entrypoint *ep, pseudo_t src, if (is_bool_type(type)) return src; - zero = value_pseudo(0); + zero = value_pseudo(type, 0); op = OP_SET_NE; return add_binary_op(ep, &bool_ctype, op, src, zero); } @@ -1591,7 +1594,7 @@ pseudo_t linearize_expression(struct entrypoint *ep, struct expression *expr) return add_symbol_address(ep, expr->symbol); case EXPR_VALUE: - return value_pseudo(expr->value); + return value_pseudo(expr->ctype, expr->value); case EXPR_STRING: case EXPR_FVALUE: case EXPR_LABEL: return add_setval(ep, expr->ctype, expr); @@ -1681,7 +1684,7 @@ static pseudo_t linearize_one_symbol(struct entrypoint *ep, struct symbol *sym) ad.result_type = sym; ad.source_type = base_type(sym); ad.address = symbol_pseudo(ep, sym); - linearize_store_gen(ep, value_pseudo(0), &ad); + linearize_store_gen(ep, value_pseudo(sym, 0), &ad); } value = linearize_initializer(ep, sym->initializer, &ad); diff --git a/linearize.h b/linearize.h index bac82d7..fd8e00d 100644 --- a/linearize.h +++ b/linearize.h @@ -32,7 +32,10 @@ struct pseudo { int nr; enum pseudo_type type; struct pseudo_user_list *users; - struct ident *ident; + union { + struct ident *ident; + int size; /* OP_SETVAL only */ + }; union { struct symbol *sym; struct instruction *def; @@ -333,7 +336,8 @@ extern void insert_branch(struct basic_block *bb, struct instruction *br, struct pseudo_t alloc_phi(struct basic_block *source, pseudo_t pseudo, int size); pseudo_t alloc_pseudo(struct instruction *def); -pseudo_t value_pseudo(long long val); +pseudo_t value_pseudo(struct symbol *type, long long val); +unsigned int value_size(long long value); struct entrypoint *linearize_symbol(struct symbol *sym); int unssa(struct entrypoint *ep); diff --git a/memops.c b/memops.c index aeacdf5..6a795c1 100644 --- a/memops.c +++ b/memops.c @@ -127,7 +127,7 @@ static void simplify_loads(struct basic_block *bb) if (!dominators) { if (local) { assert(pseudo->type != PSEUDO_ARG); - convert_load_instruction(insn, value_pseudo(0)); + convert_load_instruction(insn, value_pseudo(insn->type, 0)); } goto next_load; } diff --git a/simplify.c b/simplify.c index 2bc86f5..1e926e7 100644 --- a/simplify.c +++ b/simplify.c @@ -352,7 +352,7 @@ static int replace_with_pseudo(struct instruction *insn, pseudo_t pseudo) return REPEAT_CSE; } -static unsigned int value_size(long long value) +unsigned int value_size(long long value) { value >>= 8;
This is actually easier than I though. It is also very simple to reason about. If you are not looking at the size field at all. It is exactly the same as before, except we create a few more pseudo for the same value that have different size. The size just tag along the constant pseudo. Which did not change any way. Chris --------------------------------------------- Currently value pseudo does not have size. This create a problem pointed out by Dibyendu. When using LLVM, calling varidic function with constant value, there is no where to find the size. Linus give out two suggestions. One is give pseudo a size. The other one is the push instruction. This is the implemnation of the first suggestion. The model is actual very simple. The pseudo is exactly as before if you are not looking at the size. There is a size at create time, which tag alone with it. I might not get all the type and cast of the constant right, we can fix it later if we found some case I did not cover. Testign done: - Test suite passed. - Kernel compile get the exact same output as the RC5 at similar time. Reported-by: Dibyendu Majumdar <mobile@majumdar.org.uk> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Christopher Li <sparse@chrisli.org> --- flow.c | 2 +- linearize.c | 23 +++++++++++++---------- linearize.h | 8 ++++++-- memops.c | 2 +- simplify.c | 20 ++++++++++---------- 5 files changed, 31 insertions(+), 24 deletions(-) if (!value) @@ -398,7 +398,7 @@ static int simplify_asr(struct instruction *insn, pseudo_t pseudo, long long val if (value >= size) { warning(insn->pos, "right shift by bigger than source value"); - return replace_with_pseudo(insn, value_pseudo(0)); + return replace_with_pseudo(insn, value_pseudo(insn->type, 0)); } if (!value) return replace_with_pseudo(insn, pseudo); @@ -508,7 +508,7 @@ static int simplify_constant_rightside(struct instruction *insn) case OP_SUB: if (value) { insn->opcode = OP_ADD; - insn->src2 = value_pseudo(-value); + insn->src2 = value_pseudo(insn->type, -value); return REPEAT_CSE; } /* Fall through */ @@ -525,7 +525,7 @@ static int simplify_constant_rightside(struct instruction *insn) case OP_MODU: case OP_MODS: if (value == 1) - return replace_with_pseudo(insn, value_pseudo(0)); + return replace_with_pseudo(insn, value_pseudo(insn->type, 0)); return 0; case OP_DIVU: case OP_DIVS: @@ -686,7 +686,7 @@ static int simplify_constant_binop(struct instruction *insn) } res &= bits; - replace_with_pseudo(insn, value_pseudo(res)); + replace_with_pseudo(insn, value_pseudo(insn->type, res)); return REPEAT_CSE; } @@ -700,14 +700,14 @@ static int simplify_binop_same_args(struct instruction *insn, pseudo_t arg) warning(insn->pos, "self-comparison always evaluates to false"); case OP_SUB: case OP_XOR: - return replace_with_pseudo(insn, value_pseudo(0)); + return replace_with_pseudo(insn, value_pseudo(insn->type, 0)); case OP_SET_EQ: case OP_SET_LE: case OP_SET_GE: case OP_SET_BE: case OP_SET_AE: if (Wtautological_compare) warning(insn->pos, "self-comparison always evaluates to true"); - return replace_with_pseudo(insn, value_pseudo(1)); + return replace_with_pseudo(insn, value_pseudo(insn->type, 1)); case OP_AND: case OP_OR: @@ -716,7 +716,7 @@ static int simplify_binop_same_args(struct instruction *insn, pseudo_t arg) case OP_AND_BOOL: case OP_OR_BOOL: remove_usage(arg, &insn->src2); - insn->src2 = value_pseudo(0); + insn->src2 = value_pseudo(insn->type, 0); insn->opcode = OP_SET_NE; return REPEAT_CSE; @@ -819,7 +819,7 @@ static int simplify_constant_unop(struct instruction *insn) mask = 1ULL << (insn->size-1); res &= mask | (mask-1); - replace_with_pseudo(insn, value_pseudo(res)); + replace_with_pseudo(insn, value_pseudo(insn->type, res)); return REPEAT_CSE; } @@ -952,7 +952,7 @@ static int simplify_cast(struct instruction *insn) if (constant(src)) { int sign = orig_type->ctype.modifiers & MOD_SIGNED; long long val = get_cast_value(src->value, orig_size, size, sign); - src = value_pseudo(val); + src = value_pseudo(orig_type, val); goto simplify; }