[PATCH] Fold x86 bzhi with zero last operand (PR rtl-optimization/87817)
Jakub Jelinek
jakub@redhat.com
Wed Nov 14 12:21:00 GMT 2018
On Wed, Nov 14, 2018 at 09:29:15AM +0100, Richard Biener wrote:
> So where do the zero_extracts come from? Can we somehow avoid
> zero-sized bit-extracts at expansion time by folding them to zero?
The following patch implements the missing folding.
Note, as I've mentioned, it is just an optimization, as if we are unlucky
enough, a constant could be propagated to the insn only at RTL optimization
time. So the other patch is needed, even when this patch also fixes the
bmi2-bzhi-2.c testcase execution.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2018-11-14 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/87817
* config/i386/i386.c (ix86_fold_builtin): For _bzhi_u{32,64} if
last argument has low 8 bits clear, fold to 0.
* gcc.target/i386/bmi2-bzhi-3.c (main): Add a couple of new tests.
--- gcc/config/i386/i386.c.jj 2018-11-14 00:55:50.199357949 +0100
+++ gcc/config/i386/i386.c 2018-11-14 10:28:29.492491865 +0100
@@ -32671,6 +32671,8 @@ ix86_fold_builtin (tree fndecl, int n_ar
unsigned int idx = tree_to_uhwi (args[1]) & 0xff;
if (idx >= TYPE_PRECISION (TREE_TYPE (args[0])))
return args[0];
+ if (idx == 0)
+ return build_int_cst (TREE_TYPE (TREE_TYPE (fndecl)), 0);
if (!tree_fits_uhwi_p (args[0]))
break;
unsigned HOST_WIDE_INT res = tree_to_uhwi (args[0]);
--- gcc/testsuite/gcc.target/i386/bmi2-bzhi-3.c.jj 2016-10-31 13:28:07.961422421 +0100
+++ gcc/testsuite/gcc.target/i386/bmi2-bzhi-3.c 2018-11-14 10:38:40.352417418 +0100
@@ -58,7 +58,11 @@ main ()
link_error ();
if (_bzhi_u32 (c, 32) != c
|| _bzhi_u32 (c, 64) != c
- || _bzhi_u32 (c, 255) != c)
+ || _bzhi_u32 (c, 255) != c
+ || _bzhi_u32 (c, 544) != c
+ || _bzhi_u32 (c, 0) != 0
+ || _bzhi_u32 (c, 256) != 0
+ || _bzhi_u32 (c, 1024) != 0)
link_error ();
#ifdef __x86_64__
if (f21 () != 0 || f22 (-1ULL) != 0
@@ -70,7 +74,11 @@ main ()
|| f33 () != -1ULL || f34 (-1ULL) != -1ULL)
link_error ();
if (_bzhi_u64 (d, 64) != d
- || _bzhi_u64 (d, 255) != d)
+ || _bzhi_u64 (d, 255) != d
+ || _bzhi_u64 (d, 576) != d
+ || _bzhi_u64 (d, 0) != 0
+ || _bzhi_u64 (d, 256) != 0
+ || _bzhi_u64 (d, 512) != 0)
link_error ();
#endif
return 0;
Jakub
More information about the Gcc-patches
mailing list