On Tue, May 14, 2013 at 10:39:13AM +0200, Jakub Jelinek wrote:
> When trying with -O2 -mno-avx:
> #ifndef __AVX__
> #pragma GCC push_options
> #pragma GCC target("avx")
> #define __DISABLE_AVX__
> #endif
> typedef float __v8sf __attribute__ ((__vector_size__ (32)));
> typedef float __m256 __attribute__ ((__vector_size__ (32), __may_alias__));
> extern __inline __m256 __attribute__((__gnu_inline__, __always_inline__,
> __artificial__))
> _mm256_and_ps (__m256 __A, __m256 __B) { return (__m256)
> __builtin_ia32_andps256 ((__v8sf)__A, (__v8sf)__B); }
> #ifdef __DISABLE_AVX__
> #pragma GCC pop_options
> #undef __DISABLE_AVX__
> #endif
> __m256 a, b, c;
> void __attribute__((target ("avx")))
> foo (void)
> {
> a = _mm256_and_ps (b, c);
> }
> we get bogus errors and ICE:
> tty2.c: In function '_mm256_and_ps':
> tty2.c:9:1: note: The ABI for passing parameters with 32-byte alignment has
> changed in GCC 4.6
> tty2.c: In function 'foo':
> tty2.c:9:82: error: '__builtin_ia32_andps256' needs isa option -m32
> tty2.c:9:82: internal compiler error: in emit_move_insn, at expr.c:3486
> 0x77a3d2 emit_move_insn(rtx_def*, rtx_def*)
> ../../gcc/expr.c:3485
> (I have added "1 ||" instead of your generate_builtins into i386.c
> (def_builtin)), that just shows that target attribute/pragma support still
> has very severe issues that need to be fixed, instead of papered around.
>
> Note, we ICE on:
> #pragma GCC target ("mavx")
> That should be fixed too.
Ok, I had a brief look at the above two issues.
The first testcase has the problem that the ix86_previous_fndecl cache
gets out of date. When set_cfun is called on _mm256_and_ps (with the
implicit avx attribute), then ix86_previous_fndecl is set to _mm256_and_ps,
TARGET_AVX is set to true, target reinited. Then set_cfun is called
with NULL, we don't do anything. Later on #pragma GCC pop_options appears,
sets !TARGET_AVX (as that is the new target_option_current_node).
Next foo is being parsed, avx attribute is noticed, the same target node
is used for it, but when set_cfun is called for foo, ix86_previous_fndecl's
target node is the same as foo's and so we don't do cl_target_restore_option
at all, so !TARGET_AVX remains, while it should be set. That is the reason
for the bogus inform etc. Fixed by resetting the ix86_previous_fndecl cache
on any #pragma GCC target below. The #pragma GCC target ("mavx") is also
fixed below. The patch also includes the "1 ||" to enable building all
builtins. We still ICE with:
#0 fancy_abort (file=0x11d8fad "../../gcc/expr.c", line=316,
function=0x11dada3 "convert_move") at ../../gcc/diagnostic.c:1180
#1 0x0000000000771c39 in convert_move (to=0x7ffff1b2df00, from=0x7ffff1b314e0,
unsignedp=0) at ../../gcc/expr.c:316
#2 0x000000000078009f in store_expr (exp=0x7ffff19ab390,
target=0x7ffff1b2df00, call_param_p=0, nontemporal=false) at
../../gcc/expr.c:5300
#3 0x000000000077eba1 in expand_assignment (to=0x7ffff1b35090,
from=0x7ffff19ab390, nontemporal=false) at ../../gcc/expr.c:5025
on the first testcase. We don't ICE say on:
#ifndef __AVX__
#pragma GCC push_options
#pragma GCC target("avx")
#define __DISABLE_AVX__
#endif
typedef float __v8sf __attribute__ ((__vector_size__ (32)));
typedef float __m256 __attribute__ ((__vector_size__ (32), __may_alias__));
extern __inline __m256 __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
_mm256_and_ps (__m256 __A, __m256 __B) { return (__m256)
__builtin_ia32_andps256 ((__v8sf)__A, (__v8sf)__B); }
#ifdef __DISABLE_AVX__
#pragma GCC pop_options
#undef __DISABLE_AVX__
#endif
__m256 a[10], b[10], c[10];
void __attribute__((target ("avx")))
foo (void)
{
a[0] = _mm256_and_ps (b[0], c[0]);
}
The problem is that in the first testcase, the VAR_DECL c (guess also b and
a) have TYPE_MODE (TREE_TYPE (c)) == V8SFmode (this is dynamic, for vector
types TYPE_MODE is a function call), but DECL_MODE (c) is BLKmode
(it has been laid out while -mno-avx has been the current) and also
DECL_RTL which is a mem:BLK. Guess expr.c would need to special case
TREE_STATIC or DECL_EXTERNAL VAR_DECLs with vector type, if they have
DECL_MODE BLKmode, but TYPE_MODE some vector type, just adjust the MEM
to the desired mode?
--- gcc/config/i386/i386-c.c.jj 2013-01-15 17:20:37.000000000 +0100
+++ gcc/config/i386/i386-c.c 2013-05-14 11:46:50.773806894 +0200
@@ -369,20 +369,23 @@ ix86_pragma_target_parse (tree args, tre
if (! args)
{
- cur_tree = ((pop_target)
- ? pop_target
- : target_option_default_node);
+ cur_tree = (pop_target ? pop_target : target_option_default_node);
cl_target_option_restore (&global_options,
TREE_TARGET_OPTION (cur_tree));
}
else
{
cur_tree = ix86_valid_target_attribute_tree (args);
- if (!cur_tree)
- return false;
+ if (!cur_tree || cur_tree == error_mark_node)
+ {
+ cl_target_option_restore (&global_options,
+ TREE_TARGET_OPTION (prev_tree));
+ return false;
+ }
}
target_option_current_node = cur_tree;
+ ix86_reset_previous_fndecl ();
/* Figure out the previous/current isa, arch, tune and the differences. */
prev_opt = TREE_TARGET_OPTION (prev_tree);
--- gcc/config/i386/i386-protos.h.jj 2013-04-03 08:28:50.000000000 +0200
+++ gcc/config/i386/i386-protos.h 2013-05-14 11:41:35.389638299 +0200
@@ -40,6 +40,8 @@ extern void ix86_output_addr_diff_elt (F
extern enum calling_abi ix86_cfun_abi (void);
extern enum calling_abi ix86_function_type_abi (const_tree);
+extern void ix86_reset_previous_fndecl (void);
+
#ifdef RTX_CODE
extern int standard_80387_constant_p (rtx);
extern const char *standard_80387_constant_opcode (rtx);
--- gcc/config/i386/i386.c.jj 2013-05-14 08:23:31.000000000 +0200
+++ gcc/config/i386/i386.c 2013-05-14 11:40:54.996867411 +0200
@@ -4559,6 +4559,13 @@ ix86_can_inline_p (tree caller, tree cal
/* Remember the last target of ix86_set_current_function. */
static GTY(()) tree ix86_previous_fndecl;
+/* Invalidate ix86_previous_fndecl cache. */
+void
+ix86_reset_previous_fndecl (void)
+{
+ ix86_previous_fndecl = NULL_TREE;
+}
+
/* Establish appropriate back-end context for processing the function
FNDECL. The argument might be NULL to indicate processing at top
level, outside of any function scope. */
@@ -26829,7 +26836,7 @@ def_builtin (HOST_WIDE_INT mask, const c
ix86_builtins_isa[(int) code].isa = mask;
mask &= ~OPTION_MASK_ISA_64BIT;
- if (mask == 0
+ if (/* HACK */ 1 || mask == 0
|| (mask & ix86_isa_flags) != 0
|| (lang_hooks.builtin_function
== lang_hooks.builtin_function_ext_scope))
Jakub
