-miamcu specifies how parameters are passed to functions and how value
is returned from a function. Inside function, we can use instructions
supported by -march=XXX. -miamcu -march=haswell can generate x87, SSE
and AVX instructions as long as the IA MCU psABI is followed. But since
Lakemont processor doesn't support x87 instructions, we shouldn't
generate x87 instructions with -march=lakemont. This patch separates
code generation from -miamcu and makes -march=lakemont not to generate
x87 instructions.
gcc/
PR target/67963
PR target/67985
* common/config/i386/i386-common.c (ix86_handle_option): Remove
OPT_miamcu handling.
* config/i386/i386.c (PTA_NO_80387): New macro.
(processor_alias_table): Add PTA_NO_80387 to lakemont.
(ix86_option_override_internal): Update MASK_80387 from
PTA_NO_80387. Don't warn x87/MMX/SSE/AVX for -miamcu. Warn
SSE math only if 80387 is supported. Don't change
MASK_FLOAT_RETURNS.
(ix86_valid_target_attribute_tree): Enable FPMATH_387 only if
80387 is supported.
* config/i386/i386.h (TARGET_FLOAT_RETURNS_IN_80387): True only
if TARGET_80387 is true and TARGET_IAMCU is false.
(TARGET_FLOAT_RETURNS_IN_80387_P): True only if TARGET_80387_P
is true and TARGET_IAMCU_P is false.
df10fba...
by
rguenth <rguenth@138bc75d-0d04-0410-961f-82ee72b054a4>
2015-10-19 Richard Biener <email address hidden>
PR tree-optimization/67975
* tree-cfg.h (extract_true_false_controlled_edges): Declare.
* tree-cfg.c (extract_true_false_controlled_edges): Split out
core worker from ...
* tree-ssa-loop-im.c (extract_true_false_args_from_phi): ... here.
* tree-ssa-sccvn.c (vn_phi_compute_hash): Hash number of args
instead of block number for PHIs with two or one args.
(vn_phi_eq): Compare edge predicates of PHIs that are in different
blocks.
b6ec4d7...
by
hjl <hjl@138bc75d-0d04-0410-961f-82ee72b054a4>
Use GET_MODE_BITSIZE to get vector natural alignment
Since GET_MODE_ALIGNMENT is defined by psABI and the biggest alignment
is 4 byte for IA MCU psABI, we should use GET_MODE_BITSIZE for IA MCU
psABI to get vector natural alignment to check misaligned vector move.
* config/i386/i386.c (ix86_expand_vector_move): Use
GET_MODE_BITSIZE for IA MCU psABI to get vector natural
alignment.
00a7b29...
by
hjl <hjl@138bc75d-0d04-0410-961f-82ee72b054a4>
Don't leak ISA to __attribute__ ((target("arch=XXX")))
When processing __attribute__ ((target("arch=XXX"))), we should clear
the ISA bits in x_ix86_isa_flags first to avoid leaking ISA from
command line.
gcc/
PR target/67995
* config/i386/i386.c (ix86_valid_target_attribute_tree): If
arch= is set, clear all bits in x_ix86_isa_flags, except for
ISA_64BIT, ABI_64, ABI_X32, and CODE16.
PR middle-end/68002
* common.opt (fkeep-static-functions): New option.
* doc/invoke.texi: Document it.
* cgraphunit.c (cgraph_node::finalize_function): Use it.
cd5ad70...
by
rsandifo <rsandifo@138bc75d-0d04-0410-961f-82ee72b054a4>
Remove undefined behaviour from builtins-20.c
builtins-20.c had:
if (cos((y*=2, -fabs(tan(x/-y)))) != cos((y*=2,tan(x/y))))
link_error ();
which is undefined behaviour. The test expected that y had the same
value in x/y and x/-y, but gimplification actually implements the
"obvious" interpretation, multiplying y by 2, using it for one cos call,
then multiplying it by 2 again and using it for the other cos call.
The file has other (valid) tests that side-effects don't block
optimisation, such as:
if (cosf((y*=3, -x)) != cosf((y*=3,x)))
link_error ();
so this patch simply removes this instance.
Tested on x86_64-linux-gnu, aarch64-linux-gnu and arm-linux-gnueabi.