Hi!
After clarification from Intel folks, I've committed following ABI tweak
- the boundary at which large simdlen values become unsupported is whether
the return type (or characteristic type if the former is void) can be passed
as a vector in the __regcall calling convention (thus, fits into 8 [xyz]mm*
regs for 32-bit code and 16 [xyz]mm* regs for 64-bit code).
Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk.
2016-04-12 Jakub Jelinek <ja...@redhat.com>
* config/i386/i386.c (ix86_simd_clone_compute_vecsize_and_simdlen):
Bump the upper SIMDLEN limits, so that if the return type or
characteristic type if the return type is void can be passed in
all available SSE2/AVX/AVX2/AVX512-F registers, the SIMDLEN is
allowed.
* c-c++-common/cilk-plus/SE/ef_error2.c (func2): Use vectorlength
128 instead of 32.
--- gcc/config/i386/i386.c.jj 2016-04-07 23:31:35.000000000 +0200
+++ gcc/config/i386/i386.c 2016-04-12 18:26:28.412795125 +0200
@@ -53761,7 +53761,7 @@ ix86_simd_clone_compute_vecsize_and_simd
if (clonei->simdlen
&& (clonei->simdlen < 2
- || clonei->simdlen > 128
+ || clonei->simdlen > 1024
|| (clonei->simdlen & (clonei->simdlen - 1)) != 0))
{
warning_at (DECL_SOURCE_LOCATION (node->decl), 0,
@@ -53867,21 +53867,28 @@ ix86_simd_clone_compute_vecsize_and_simd
clonei->simdlen /= GET_MODE_BITSIZE (TYPE_MODE (base_type));
}
else if (clonei->simdlen > 16)
- switch (clonei->vecsize_int)
- {
- case 512:
- /* For AVX512-F, support VLEN up to 128. */
- break;
- case 256:
- /* For AVX2, support VLEN up to 32. */
- if (clonei->simdlen <= 32)
- break;
- /* FALLTHRU */
- default:
- /* Otherwise, support VLEN up to 16. */
- warning_at (DECL_SOURCE_LOCATION (node->decl), 0,
- "unsupported simdlen %d", clonei->simdlen);
- return 0;
+ {
+ /* For compatibility with ICC, use the same upper bounds
+ for simdlen. In particular, for CTYPE below, use the return type,
+ unless the function returns void, in that case use the characteristic
+ type. If it is possible for given SIMDLEN to pass CTYPE value
+ in registers (8 [XYZ]MM* regs for 32-bit code, 16 [XYZ]MM* regs
+ for 64-bit code), accept that SIMDLEN, otherwise warn and don't
+ emit corresponding clone. */
+ tree ctype = ret_type;
+ if (TREE_CODE (ret_type) == VOID_TYPE)
+ ctype = base_type;
+ int cnt = GET_MODE_BITSIZE (TYPE_MODE (ctype)) * clonei->simdlen;
+ if (SCALAR_INT_MODE_P (TYPE_MODE (ctype)))
+ cnt /= clonei->vecsize_int;
+ else
+ cnt /= clonei->vecsize_float;
+ if (cnt > (TARGET_64BIT ? 16 : 8))
+ {
+ warning_at (DECL_SOURCE_LOCATION (node->decl), 0,
+ "unsupported simdlen %d", clonei->simdlen);
+ return 0;
+ }
}
return ret;
}
--- gcc/testsuite/c-c++-common/cilk-plus/SE/ef_error2.c.jj 2013-12-19
08:25:31.664476591 +0100
+++ gcc/testsuite/c-c++-common/cilk-plus/SE/ef_error2.c 2016-04-12
20:50:45.849737800 +0200
@@ -1,8 +1,8 @@
/* { dg-do compile { target { i?86-*-* x86_64-*-* } } } */
/* { dg-options "-fcilkplus -Wall" } */
-__attribute__((vector (vectorlength(32))))
-//#pragma omp simd simdlen (32)
+__attribute__((vector (vectorlength(128))))
+//#pragma omp simd simdlen (128)
int func2 (int x, int y) /* { dg-warning "unsupported simdlen" } */
{
return (x+y);
Jakub
