This is the mail archive of the
binutils@sourceware.org
mailing list for the binutils project.
[RFA/ARM 17/21] Add support for 16-/64-bit Floating Point conversions.
- From: Matthew Gretton-Dann <matthew dot gretton-dann at arm dot com>
- To: binutils at sourceware dot org
- Date: Thu, 23 Aug 2012 16:04:24 +0100
- Subject: [RFA/ARM 17/21] Add support for 16-/64-bit Floating Point conversions.
- References: <cover.1345733341.git.matthew.gretton-dann@arm.com>
ARMv8 extends VCVT[TB] to support converting to/from double precision
floats.
This patch adds support for this, in the process it also fixes bugs in
the support for half-precision <-> single-precision conversions.
gas/ChangeLog:
2012-08-23 Matthew Gretton-Dann <matthew.gretton-dann@arm.com>
* config/tc-arm.c (el_type_type_check): Add handling for 16-bit
floating point types.
(do_neon_cvttb_2): New function.
(do_neon_cvttb_1): Likewise.
(do_neon_cvtb): Refactor to use do_neon_cvttb_1.
(do_neon_cvtt): Likewise.
gas/testsuite/ChangeLog:
2012-08-23 Matthew Gretton-Dann <matthew.gretton-dann@arm.com>
* gas/arm/armv8-a+fp.d: Update testcase.
* gas/arm/armv8-a+fp.s: Likewise.
* gas/arm/half-prec-vfpv3.s: Likewise.
opcodes/ChangeLog:
2012-08-23 Matthew Gretton-Dann <matthew.gretton-dann@arm.com>
* arm-dis.c (coprocessor_opcodes): Add support for HP/DP
conversions.
diff --git a/gas/config/tc-arm.c b/gas/config/tc-arm.c
index 3c8bd60..6d8d18f 100644
--- a/gas/config/tc-arm.c
+++ b/gas/config/tc-arm.c
@@ -12832,7 +12832,7 @@ el_type_of_type_chk (enum neon_el_type *type, unsigned *size,
if ((mask & (N_S8 | N_U8 | N_I8 | N_8 | N_P8)) != 0)
*size = 8;
- else if ((mask & (N_S16 | N_U16 | N_I16 | N_16 | N_P16)) != 0)
+ else if ((mask & (N_S16 | N_U16 | N_I16 | N_16 | N_F16 | N_P16)) != 0)
*size = 16;
else if ((mask & (N_S32 | N_U32 | N_I32 | N_32 | N_F32)) != 0)
*size = 32;
@@ -12851,7 +12851,7 @@ el_type_of_type_chk (enum neon_el_type *type, unsigned *size,
*type = NT_untyped;
else if ((mask & (N_P8 | N_P16)) != 0)
*type = NT_poly;
- else if ((mask & (N_F32 | N_F64)) != 0)
+ else if ((mask & (N_F16 | N_F32 | N_F64)) != 0)
*type = NT_float;
else
return FAIL;
@@ -14888,31 +14888,63 @@ do_neon_cvtm (void)
}
static void
-do_neon_cvtb (void)
+do_neon_cvttb_2 (bfd_boolean t, bfd_boolean to, bfd_boolean is_double)
{
- inst.instruction = 0xeb20a40;
+ if (is_double)
+ mark_feature_used (&fpu_vfp_ext_armv8);
- /* The sizes are attached to the mnemonic. */
- if (inst.vectype.el[0].type != NT_invtype
- && inst.vectype.el[0].size == 16)
- inst.instruction |= 0x00010000;
+ encode_arm_vfp_reg (inst.operands[0].reg,
+ (is_double && !to) ? VFP_REG_Dd : VFP_REG_Sd);
+ encode_arm_vfp_reg (inst.operands[1].reg,
+ (is_double && to) ? VFP_REG_Dm : VFP_REG_Sm);
+ inst.instruction |= to ? 0x10000 : 0;
+ inst.instruction |= t ? 0x80 : 0;
+ inst.instruction |= is_double ? 0x100 : 0;
+ do_vfp_cond_or_thumb ();
+}
- /* Programmer's syntax: the sizes are attached to the operands. */
- else if (inst.operands[0].vectype.type != NT_invtype
- && inst.operands[0].vectype.size == 16)
- inst.instruction |= 0x00010000;
+static void
+do_neon_cvttb_1 (bfd_boolean t)
+{
+ enum neon_shape rs = neon_select_shape (NS_FF, NS_FD, NS_DF, NS_NULL);
- encode_arm_vfp_reg (inst.operands[0].reg, VFP_REG_Sd);
- encode_arm_vfp_reg (inst.operands[1].reg, VFP_REG_Sm);
- do_vfp_cond_or_thumb ();
+ if (rs == NS_NULL)
+ return;
+ else if (neon_check_type (2, rs, N_F16, N_F32 | N_VFP).type != NT_invtype)
+ {
+ inst.error = NULL;
+ do_neon_cvttb_2 (t, /*to=*/TRUE, /*is_double=*/FALSE);
+ }
+ else if (neon_check_type (2, rs, N_F32 | N_VFP, N_F16).type != NT_invtype)
+ {
+ inst.error = NULL;
+ do_neon_cvttb_2 (t, /*to=*/FALSE, /*is_double=*/FALSE);
+ }
+ else if (neon_check_type (2, rs, N_F16, N_F64 | N_VFP).type != NT_invtype)
+ {
+ inst.error = NULL;
+ do_neon_cvttb_2 (t, /*to=*/TRUE, /*is_double=*/TRUE);
+ }
+ else if (neon_check_type (2, rs, N_F64 | N_VFP, N_F16).type != NT_invtype)
+ {
+ inst.error = NULL;
+ do_neon_cvttb_2 (t, /*to=*/FALSE, /*is_double=*/TRUE);
+ }
+ else
+ return;
+}
+
+static void
+do_neon_cvtb (void)
+{
+ do_neon_cvttb_1 (FALSE);
}
static void
do_neon_cvtt (void)
{
- do_neon_cvtb ();
- inst.instruction |= 0x80;
+ do_neon_cvttb_1 (TRUE);
}
static void
@@ -18958,8 +18990,8 @@ static const struct asm_opcode insns[] =
nCEF(vcvt, _vcvt, 3, (RNSDQ, RNSDQ, oI32z), neon_cvt),
nCEF(vcvtr, _vcvt, 2, (RNSDQ, RNSDQ), neon_cvtr),
- nCEF(vcvtb, _vcvt, 2, (RVS, RVS), neon_cvtb),
- nCEF(vcvtt, _vcvt, 2, (RVS, RVS), neon_cvtt),
+ NCEF(vcvtb, eb20a40, 2, (RVSD, RVSD), neon_cvtb),
+ NCEF(vcvtt, eb20a40, 2, (RVSD, RVSD), neon_cvtt),
/* NOTE: All VMOV encoding is special-cased! */
diff --git a/gas/testsuite/gas/arm/armv8-a+fp.d b/gas/testsuite/gas/arm/armv8-a+fp.d
index e478411..bb52e0a 100644
--- a/gas/testsuite/gas/arm/armv8-a+fp.d
+++ b/gas/testsuite/gas/arm/armv8-a+fp.d
@@ -50,6 +50,14 @@ Disassembly of section .text:
0[0-9a-f]+ <[^>]+> feb91b41 vrintn.f64.f64 d1, d1
0[0-9a-f]+ <[^>]+> fefaeb6e vrintp.f64.f64 d30, d30
0[0-9a-f]+ <[^>]+> fefbfb6f vrintm.f64.f64 d31, d31
+0[0-9a-f]+ <[^>]+> eeb30bc0 vcvtt.f16.f64 s0, d0
+0[0-9a-f]+ <[^>]+> eef30b60 vcvtb.f16.f64 s1, d16
+0[0-9a-f]+ <[^>]+> eeb3fbcf vcvtt.f16.f64 s30, d15
+0[0-9a-f]+ <[^>]+> eef3fb6f vcvtb.f16.f64 s31, d31
+0[0-9a-f]+ <[^>]+> eeb20bc0 vcvtt.f64.f16 d0, s0
+0[0-9a-f]+ <[^>]+> eef20b60 vcvtb.f64.f16 d16, s1
+0[0-9a-f]+ <[^>]+> eeb2fbcf vcvtt.f64.f16 d15, s30
+0[0-9a-f]+ <[^>]+> eef2fb6f vcvtb.f64.f16 d31, s31
0[0-9a-f]+ <[^>]+> fe00 0a00 vseleq.f32 s0, s0, s0
0[0-9a-f]+ <[^>]+> fe50 0aa0 vselvs.f32 s1, s1, s1
0[0-9a-f]+ <[^>]+> fe2f fa0f vselge.f32 s30, s30, s30
@@ -96,3 +104,11 @@ Disassembly of section .text:
0[0-9a-f]+ <[^>]+> feb9 1b41 vrintn.f64.f64 d1, d1
0[0-9a-f]+ <[^>]+> fefa eb6e vrintp.f64.f64 d30, d30
0[0-9a-f]+ <[^>]+> fefb fb6f vrintm.f64.f64 d31, d31
+0[0-9a-f]+ <[^>]+> eeb3 0bc0 vcvtt.f16.f64 s0, d0
+0[0-9a-f]+ <[^>]+> eef3 0b60 vcvtb.f16.f64 s1, d16
+0[0-9a-f]+ <[^>]+> eeb3 fbcf vcvtt.f16.f64 s30, d15
+0[0-9a-f]+ <[^>]+> eef3 fb6f vcvtb.f16.f64 s31, d31
+0[0-9a-f]+ <[^>]+> eeb2 0bc0 vcvtt.f64.f16 d0, s0
+0[0-9a-f]+ <[^>]+> eef2 0b60 vcvtb.f64.f16 d16, s1
+0[0-9a-f]+ <[^>]+> eeb2 fbcf vcvtt.f64.f16 d15, s30
+0[0-9a-f]+ <[^>]+> eef2 fb6f vcvtb.f64.f16 d31, s31
diff --git a/gas/testsuite/gas/arm/armv8-a+fp.s b/gas/testsuite/gas/arm/armv8-a+fp.s
index 10a391a..f99302f 100644
--- a/gas/testsuite/gas/arm/armv8-a+fp.s
+++ b/gas/testsuite/gas/arm/armv8-a+fp.s
@@ -50,6 +50,14 @@
vrintn.f64.f64 d1, d1
vrintp.f64.f64 d30, d30
vrintm.f64.f64 d31, d31
+ vcvtt.f16.f64 s0, d0
+ vcvtb.f16.f64 s1, d16
+ vcvtt.f16.f64 s30, d15
+ vcvtb.f16.f64 s31, d31
+ vcvtt.f64.f16 d0, s0
+ vcvtb.f64.f16 d16, s1
+ vcvtt.f64.f16 d15, s30
+ vcvtb.f64.f16 d31, s31
.thumb
vseleq.f32 s0, s0, s0
@@ -98,3 +106,11 @@
vrintn.f64.f64 d1, d1
vrintp.f64.f64 d30, d30
vrintm.f64.f64 d31, d31
+ vcvtt.f16.f64 s0, d0
+ vcvtb.f16.f64 s1, d16
+ vcvtt.f16.f64 s30, d15
+ vcvtb.f16.f64 s31, d31
+ vcvtt.f64.f16 d0, s0
+ vcvtb.f64.f16 d16, s1
+ vcvtt.f64.f16 d15, s30
+ vcvtb.f64.f16 d31, s31
diff --git a/gas/testsuite/gas/arm/half-prec-vfpv3.s b/gas/testsuite/gas/arm/half-prec-vfpv3.s
index acd1508..d658807 100644
--- a/gas/testsuite/gas/arm/half-prec-vfpv3.s
+++ b/gas/testsuite/gas/arm/half-prec-vfpv3.s
@@ -1,20 +1,20 @@
.text
- vcvtt.f32.f32 s0, s1
- vcvtteq.f32.f32 s2, s3
- vcvttne.f32.f32 s2, s3
- vcvttcs.f32.f32 s2, s3
- vcvttcc.f32.f32 s2, s3
- vcvttmi.f32.f32 s2, s3
- vcvttpl.f32.f32 s2, s3
- vcvttvs.f32.f32 s2, s3
- vcvttvc.f32.f32 s2, s3
- vcvtthi.f32.f32 s2, s3
- vcvttls.f32.f32 s2, s3
- vcvttge.f32.f32 s2, s3
- vcvttlt.f32.f32 s2, s3
- vcvttgt.f32.f32 s2, s3
- vcvttle.f32.f32 s2, s3
- vcvttal.f32.f32 s2, s3
+ vcvtt.f32.f16 s0, s1
+ vcvtteq.f32.f16 s2, s3
+ vcvttne.f32.f16 s2, s3
+ vcvttcs.f32.f16 s2, s3
+ vcvttcc.f32.f16 s2, s3
+ vcvttmi.f32.f16 s2, s3
+ vcvttpl.f32.f16 s2, s3
+ vcvttvs.f32.f16 s2, s3
+ vcvttvc.f32.f16 s2, s3
+ vcvtthi.f32.f16 s2, s3
+ vcvttls.f32.f16 s2, s3
+ vcvttge.f32.f16 s2, s3
+ vcvttlt.f32.f16 s2, s3
+ vcvttgt.f32.f16 s2, s3
+ vcvttle.f32.f16 s2, s3
+ vcvttal.f32.f16 s2, s3
vcvtt.f16.f32 s0, s1
vcvtteq.f16.f32 s2, s3
@@ -33,22 +33,22 @@
vcvttle.f16.f32 s2, s3
vcvttal.f16.f32 s2, s3
- vcvtb.f32.f32 s0, s1
- vcvtbeq.f32.f32 s2, s3
- vcvtbne.f32.f32 s2, s3
- vcvtbcs.f32.f32 s2, s3
- vcvtbcc.f32.f32 s2, s3
- vcvtbmi.f32.f32 s2, s3
- vcvtbpl.f32.f32 s2, s3
- vcvtbvs.f32.f32 s2, s3
- vcvtbvc.f32.f32 s2, s3
- vcvtbhi.f32.f32 s2, s3
- vcvtbls.f32.f32 s2, s3
- vcvtbge.f32.f32 s2, s3
- vcvtblt.f32.f32 s2, s3
- vcvtbgt.f32.f32 s2, s3
- vcvtble.f32.f32 s2, s3
- vcvtbal.f32.f32 s2, s3
+ vcvtb.f32.f16 s0, s1
+ vcvtbeq.f32.f16 s2, s3
+ vcvtbne.f32.f16 s2, s3
+ vcvtbcs.f32.f16 s2, s3
+ vcvtbcc.f32.f16 s2, s3
+ vcvtbmi.f32.f16 s2, s3
+ vcvtbpl.f32.f16 s2, s3
+ vcvtbvs.f32.f16 s2, s3
+ vcvtbvc.f32.f16 s2, s3
+ vcvtbhi.f32.f16 s2, s3
+ vcvtbls.f32.f16 s2, s3
+ vcvtbge.f32.f16 s2, s3
+ vcvtblt.f32.f16 s2, s3
+ vcvtbgt.f32.f16 s2, s3
+ vcvtble.f32.f16 s2, s3
+ vcvtbal.f32.f16 s2, s3
vcvtb.f16.f32 s0, s1
vcvtbeq.f16.f32 s2, s3
diff --git a/opcodes/arm-dis.c b/opcodes/arm-dis.c
index 60463d1..c6deac9 100644
--- a/opcodes/arm-dis.c
+++ b/opcodes/arm-dis.c
@@ -319,6 +319,8 @@ static const struct opcode32 coprocessor_opcodes[] =
{FPU_NEON_EXT_V1, 0x0e400b10, 0x0fd00f10, "vmov%c.8\t%16-19,7D[%5,6,21d], %12-15r"},
{FPU_NEON_EXT_V1, 0x0e500b10, 0x0f500f10, "vmov%c.%23?us8\t%12-15r, %16-19,7D[%5,6,21d]"},
/* Half-precision conversion instructions. */
+ {FPU_VFP_EXT_ARMV8, 0x0eb20b40, 0x0fbf0f50, "vcvt%7?tb%c.f64.f16\t%z1, %y0"},
+ {FPU_VFP_EXT_ARMV8, 0x0eb30b40, 0x0fbf0f50, "vcvt%7?tb%c.f16.f64\t%y1, %z0"},
{FPU_VFP_EXT_FP16, 0x0eb20a40, 0x0fbf0f50, "vcvt%7?tb%c.f32.f16\t%y1, %y0"},
{FPU_VFP_EXT_FP16, 0x0eb30a40, 0x0fbf0f50, "vcvt%7?tb%c.f16.f32\t%y1, %y0"},