v2: use a new arm_arch_v7ve_neon, fix use of DImode in output_move_neon
The valid offset range of LDRD in arm_legitimate_index_p is increased to
-1024..1020 if NEON is enabled since VALID_NEON_DREG_MODE includes DImode.
Fix this by moving the LDRD check earlier.
Passes bootstrap & regress, OK for commit?
gcc:
PR target/115153
* config/arm/arm.cc (arm_legitimate_index_p): Move LDRD case before
NEON.
(thumb2_legitimate_index_p): Update comments.
(output_move_neon): Use DFmode for vldr/vstr.
* lib/target-supports.exp: Add arm_arch_v7ve_neon target support.
gcc/testsuite:
PR target/115153
* gcc.target/arm/pr115153.c: Add new test.
---
diff --git a/gcc/config/arm/arm.cc b/gcc/config/arm/arm.cc
index
ea0c963a4d67ecd70e1571624e84dfe46d757df9..7dec0254f5a953050c9c52aa297fad7f3dfb6c74
100644
--- a/gcc/config/arm/arm.cc
+++ b/gcc/config/arm/arm.cc
@@ -8852,6 +8852,28 @@ arm_legitimate_index_p (machine_mode mode, rtx index,
RTX_CODE outer,
&& INTVAL (index) > -1024
&& (INTVAL (index) & 3) == 0);
+ if (arm_address_register_rtx_p (index, strict_p)
+ && (GET_MODE_SIZE (mode) <= 4))
+ return 1;
+
+ /* This handles DFmode only if !TARGET_HARD_FLOAT. */
+ if (mode == DImode || mode == DFmode)
+ {
+ if (code == CONST_INT)
+ {
+ HOST_WIDE_INT val = INTVAL (index);
+
+ /* Assume we emit ldrd or 2x ldr if !TARGET_LDRD.
+ If vldr is selected it uses arm_coproc_mem_operand. */
+ if (TARGET_LDRD)
+ return val > -256 && val < 256;
+ else
+ return val > -4096 && val < 4092;
+ }
+
+ return TARGET_LDRD && arm_address_register_rtx_p (index, strict_p);
+ }
+
/* For quad modes, we restrict the constant offset to be slightly less
than what the instruction format permits. We do this because for
quad mode moves, we will actually decompose them into two separate
@@ -8864,7 +8886,7 @@ arm_legitimate_index_p (machine_mode mode, rtx index,
RTX_CODE outer,
&& (INTVAL (index) & 3) == 0);
/* We have no such constraint on double mode offsets, so we permit the
- full range of the instruction format. */
+ full range of the instruction format. Note DImode is included here. */
if (TARGET_NEON && VALID_NEON_DREG_MODE (mode))
return (code == CONST_INT
&& INTVAL (index) < 1024
@@ -8877,27 +8899,6 @@ arm_legitimate_index_p (machine_mode mode, rtx index,
RTX_CODE outer,
&& INTVAL (index) > -1024
&& (INTVAL (index) & 3) == 0);
- if (arm_address_register_rtx_p (index, strict_p)
- && (GET_MODE_SIZE (mode) <= 4))
- return 1;
-
- if (mode == DImode || mode == DFmode)
- {
- if (code == CONST_INT)
- {
- HOST_WIDE_INT val = INTVAL (index);
-
- /* Assume we emit ldrd or 2x ldr if !TARGET_LDRD.
- If vldr is selected it uses arm_coproc_mem_operand. */
- if (TARGET_LDRD)
- return val > -256 && val < 256;
- else
- return val > -4096 && val < 4092;
- }
-
- return TARGET_LDRD && arm_address_register_rtx_p (index, strict_p);
- }
-
if (GET_MODE_SIZE (mode) <= 4
&& ! (arm_arch4
&& (mode == HImode
@@ -9000,7 +9001,7 @@ thumb2_legitimate_index_p (machine_mode mode, rtx index,
int strict_p)
&& (INTVAL (index) & 3) == 0);
/* We have no such constraint on double mode offsets, so we permit the
- full range of the instruction format. */
+ full range of the instruction format. Note DImode is included here. */
if (TARGET_NEON && VALID_NEON_DREG_MODE (mode))
return (code == CONST_INT
&& INTVAL (index) < 1024
@@ -9011,6 +9012,7 @@ thumb2_legitimate_index_p (machine_mode mode, rtx index,
int strict_p)
&& (GET_MODE_SIZE (mode) <= 4))
return 1;
+ /* This handles DImode if !TARGET_NEON, and DFmode if !TARGET_VFP_BASE. */
if (mode == DImode || mode == DFmode)
{
if (code == CONST_INT)
@@ -20854,7 +20856,7 @@ output_move_neon (rtx *operands)
/* We're only using DImode here because it's a convenient
size. */
ops[0] = gen_rtx_REG (DImode, REGNO (reg) + 2 * i);
- ops[1] = adjust_address (mem, DImode, 8 * i);
+ ops[1] = adjust_address (mem, DFmode, 8 * i);
if (reg_overlap_mentioned_p (ops[0], mem))
{
gcc_assert (overlap == -1);
@@ -20872,7 +20874,7 @@ output_move_neon (rtx *operands)
if (overlap != -1)
{
ops[0] = gen_rtx_REG (DImode, REGNO (reg) + 2 * overlap);
- ops[1] = adjust_address (mem, SImode, 8 * overlap);
+ ops[1] = adjust_address (mem, DFmode, 8 * overlap);
if (TARGET_HAVE_MVE && LABEL_REF_P (addr))
sprintf (buff, "v%sr.32\t%%P0, %%1", load ? "ld" : "st");
else
diff --git a/gcc/testsuite/gcc.target/arm/pr115153.c
b/gcc/testsuite/gcc.target/arm/pr115153.c
new file mode 100644
index
0000000000000000000000000000000000000000..80b57acf87ec667123146873afab1cd4a581e7f5
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/pr115153.c
@@ -0,0 +1,16 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -marm" } */
+/* { dg-require-effective-target arm_arch_v7ve_neon_ok } */
+/* { dg-add-options arm_arch_v7ve_neon } */
+/* { dg-final { check-function-bodies "**" "" "" } } */
+
+/*
+** f1:
+** add r0, r0, #256
+** ldrd r0, r1, \[r0\]
+** bx lr
+*/
+long long f1 (long long *p)
+{
+ return __atomic_load_n (p + 32, __ATOMIC_RELAXED);
+}
diff --git a/gcc/testsuite/lib/target-supports.exp
b/gcc/testsuite/lib/target-supports.exp
index
5c0a3dade222777402d2a44e63cb5be305341471..20ea318bba65ff7685d1378bb898a15cb6c4f00b
100644
--- a/gcc/testsuite/lib/target-supports.exp
+++ b/gcc/testsuite/lib/target-supports.exp
@@ -5518,6 +5518,8 @@ foreach { armfunc armflag armdefs } {
v7em "-march=armv7e-m+fp -mthumb" __ARM_ARCH_7EM__
v7ve "-march=armv7ve+fp -marm"
"__ARM_ARCH_7A__ && __ARM_FEATURE_IDIV"
+ v7ve_neon "-march=armv7ve+simd -mfpu=auto -mfloat-abi=softfp"
+ "__ARM_ARCH_7A__ && __ARM_FEATURE_IDIV && __ARM_NEON__"
v8a "-march=armv8-a+simd" __ARM_ARCH_8A__
v8a_hard "-march=armv8-a+simd -mfpu=auto -mfloat-abi=hard"
__ARM_ARCH_8A__
v8_1a "-march=armv8.1-a+simd" __ARM_ARCH_8A__
