From 04cdc29b007594a0e58ffef0c9dd87df3ea595ea Mon Sep 17 00:00:00 2001
From: Mark Wielaard <mark@klomp.org>
Date: Wed, 14 Oct 2020 06:11:34 -0400
Subject: [PATCH] arm64 VEX frontend and backend support for
Iop_M{Add,Sub}F{32,64}
The arm64 frontend used to implement the scalar fmadd, fmsub, fnmadd
and fnmsub iinstructions into separate addition/substraction and
multiplication instructions, which caused rounding issues.
This patch turns them into Iop_M{Add,Sub}F{32,64} instructions
(with some arguments negated). And the backend now emits fmadd or fmsub
instructions.
Alexandra Hajkova <ahajkova@redhat.com> added tests and fixed up the
implementation to make sure rounding (and sign) are correct now.
https://bugs.kde.org/show_bug.cgi?id=426014
---
VEX/priv/guest_arm64_toIR.c | 58 ++++++++---
VEX/priv/host_arm64_defs.c | 136 +++++++++++++++++++++++++-
VEX/priv/host_arm64_defs.h | 30 ++++++
VEX/priv/host_arm64_isel.c | 39 ++++++++
none/tests/arm64/Makefile.am | 6 +-
none/tests/arm64/fmadd_sub.c | 98 +++++++++++++++++++
none/tests/arm64/fmadd_sub.stderr.exp | 0
none/tests/arm64/fmadd_sub.stdout.exp | 125 +++++++++++++++++++++++
none/tests/arm64/fmadd_sub.vgtest | 3 +
9 files changed, 479 insertions(+), 16 deletions(-)
create mode 100644 none/tests/arm64/fmadd_sub.c
create mode 100644 none/tests/arm64/fmadd_sub.stderr.exp
create mode 100644 none/tests/arm64/fmadd_sub.stdout.exp
create mode 100644 none/tests/arm64/fmadd_sub.vgtest
diff --git a/VEX/priv/guest_arm64_toIR.c b/VEX/priv/guest_arm64_toIR.c
index 556b85a6a..d242d43c0 100644
--- a/VEX/priv/guest_arm64_toIR.c
+++ b/VEX/priv/guest_arm64_toIR.c
@@ -286,6 +286,12 @@ static IRExpr* triop ( IROp op, IRExpr* a1, IRExpr* a2, IRExpr* a3 )
return IRExpr_Triop(op, a1, a2, a3);
}
+static IRExpr* qop ( IROp op, IRExpr* a1, IRExpr* a2,
+ IRExpr* a3, IRExpr* a4 )
+{
+ return IRExpr_Qop(op, a1, a2, a3, a4);
+}
+
static IRExpr* loadLE ( IRType ty, IRExpr* addr )
{
return IRExpr_Load(Iend_LE, ty, addr);
@@ -532,6 +538,22 @@ static IROp mkADDF ( IRType ty ) {
}
}
+static IROp mkFMADDF ( IRType ty ) {
+ switch (ty) {
+ case Ity_F32: return Iop_MAddF32;
+ case Ity_F64: return Iop_MAddF64;
+ default: vpanic("mkFMADDF");
+ }
+}
+
+static IROp mkFMSUBF ( IRType ty ) {
+ switch (ty) {
+ case Ity_F32: return Iop_MSubF32;
+ case Ity_F64: return Iop_MSubF64;
+ default: vpanic("mkFMSUBF");
+ }
+}
+
static IROp mkSUBF ( IRType ty ) {
switch (ty) {
case Ity_F32: return Iop_SubF32;
@@ -14368,30 +14390,40 @@ Bool dis_AdvSIMD_fp_data_proc_3_source(/*MB_OUT*/DisResult* dres, UInt insn)
where Fx=Dx when sz=1, Fx=Sx when sz=0
-----SPEC------ ----IMPL----
- fmadd a + n * m a + n * m
- fmsub a + (-n) * m a - n * m
- fnmadd (-a) + (-n) * m -(a + n * m)
- fnmsub (-a) + n * m -(a - n * m)
+ fmadd a + n * m fmadd (a, n, m)
+ fmsub a + (-n) * m fmsub (a, n, m)
+ fnmadd (-a) + (-n) * m fmadd (-a, -n, m)
+ fnmsub (-a) + n * m fmadd (-a, n, m)
+
+ Note Iop_MAdd/SubF32/64 take arguments in the order: rm, N, M, A
*/
Bool isD = (ty & 1) == 1;
UInt ix = (bitO1 << 1) | bitO0;
IRType ity = isD ? Ity_F64 : Ity_F32;
- IROp opADD = mkADDF(ity);
- IROp opSUB = mkSUBF(ity);
- IROp opMUL = mkMULF(ity);
+ IROp opFMADD = mkFMADDF(ity);
+ IROp opFMSUB = mkFMSUBF(ity);
IROp opNEG = mkNEGF(ity);
IRTemp res = newTemp(ity);
IRExpr* eA = getQRegLO(aa, ity);
IRExpr* eN = getQRegLO(nn, ity);
IRExpr* eM = getQRegLO(mm, ity);
IRExpr* rm = mkexpr(mk_get_IR_rounding_mode());
- IRExpr* eNxM = triop(opMUL, rm, eN, eM);
switch (ix) {
- case 0: assign(res, triop(opADD, rm, eA, eNxM)); break;
- case 1: assign(res, triop(opSUB, rm, eA, eNxM)); break;
- case 2: assign(res, unop(opNEG, triop(opADD, rm, eA, eNxM))); break;
- case 3: assign(res, unop(opNEG, triop(opSUB, rm, eA, eNxM))); break;
- default: vassert(0);
+ case 0: /* FMADD */
+ assign(res, qop(opFMADD, rm, eN, eM, eA));
+ break;
+ case 1: /* FMSUB */
+ assign(res, qop(opFMSUB, rm, eN, eM, eA));
+ break;
+ case 2: /* FNMADD */
+ assign(res, qop(opFMADD, rm, unop(opNEG, eN), eM,
+ unop(opNEG,eA)));
+ break;
+ case 3: /* FNMSUB */
+ assign(res, qop(opFMADD, rm, eN, eM, unop(opNEG, eA)));
+ break;
+ default:
+ vassert(0);
}
putQReg128(dd, mkV128(0x0000));
putQRegLO(dd, mkexpr(res));
diff --git a/VEX/priv/host_arm64_defs.c b/VEX/priv/host_arm64_defs.c
index e4ef56986..13b497f60 100644
--- a/VEX/priv/host_arm64_defs.c
+++ b/VEX/priv/host_arm64_defs.c
@@ -546,6 +546,14 @@ static const HChar* showARM64FpBinOp ( ARM64FpBinOp op ) {
}
}
+static const HChar* showARM64FpTriOp ( ARM64FpTriOp op ) {
+ switch (op) {
+ case ARM64fpt_FMADD: return "fmadd";
+ case ARM64fpt_FMSUB: return "fmsub";
+ default: vpanic("showARM64FpTriOp");
+ }
+}
+
static const HChar* showARM64FpUnaryOp ( ARM64FpUnaryOp op ) {
switch (op) {
case ARM64fpu_NEG: return "neg ";
@@ -1154,6 +1162,28 @@ ARM64Instr* ARM64Instr_VBinS ( ARM64FpBinOp op,
i->ARM64in.VBinS.argR = argR;
return i;
}
+ARM64Instr* ARM64Instr_VTriD ( ARM64FpTriOp op,
+ HReg dst, HReg arg1, HReg arg2, HReg arg3 ) {
+ ARM64Instr* i = LibVEX_Alloc_inline(sizeof(ARM64Instr));
+ i->tag = ARM64in_VTriD;
+ i->ARM64in.VTriD.op = op;
+ i->ARM64in.VTriD.dst = dst;
+ i->ARM64in.VTriD.arg1 = arg1;
+ i->ARM64in.VTriD.arg2 = arg2;
+ i->ARM64in.VTriD.arg3 = arg3;
+ return i;
+}
+ARM64Instr* ARM64Instr_VTriS ( ARM64FpTriOp op,
+ HReg dst, HReg arg1, HReg arg2, HReg arg3 ) {
+ ARM64Instr* i = LibVEX_Alloc_inline(sizeof(ARM64Instr));
+ i->tag = ARM64in_VTriS;
+ i->ARM64in.VTriS.op = op;
+ i->ARM64in.VTriS.dst = dst;
+ i->ARM64in.VTriS.arg1 = arg1;
+ i->ARM64in.VTriS.arg2 = arg2;
+ i->ARM64in.VTriS.arg3 = arg3;
+ return i;
+}
ARM64Instr* ARM64Instr_VCmpD ( HReg argL, HReg argR ) {
ARM64Instr* i = LibVEX_Alloc_inline(sizeof(ARM64Instr));
i->tag = ARM64in_VCmpD;
@@ -1756,6 +1786,26 @@ void ppARM64Instr ( const ARM64Instr* i ) {
vex_printf(", ");
ppHRegARM64asSreg(i->ARM64in.VBinS.argR);
return;
+ case ARM64in_VTriD:
+ vex_printf("f%s ", showARM64FpTriOp(i->ARM64in.VTriD.op));
+ ppHRegARM64(i->ARM64in.VTriD.dst);
+ vex_printf(", ");
+ ppHRegARM64(i->ARM64in.VTriD.arg1);
+ vex_printf(", ");
+ ppHRegARM64(i->ARM64in.VTriD.arg2);
+ vex_printf(", ");
+ ppHRegARM64(i->ARM64in.VTriD.arg3);
+ return;
+ case ARM64in_VTriS:
+ vex_printf("f%s ", showARM64FpTriOp(i->ARM64in.VTriS.op));
+ ppHRegARM64asSreg(i->ARM64in.VTriS.dst);
+ vex_printf(", ");
+ ppHRegARM64asSreg(i->ARM64in.VTriS.arg1);
+ vex_printf(", ");
+ ppHRegARM64asSreg(i->ARM64in.VTriS.arg2);
+ vex_printf(", ");
+ ppHRegARM64asSreg(i->ARM64in.VTriS.arg3);
+ return;
case ARM64in_VCmpD:
vex_printf("fcmp ");
ppHRegARM64(i->ARM64in.VCmpD.argL);
@@ -2197,6 +2247,18 @@ void getRegUsage_ARM64Instr ( HRegUsage* u, const ARM64Instr* i, Bool mode64 )
addHRegUse(u, HRmRead, i->ARM64in.VBinS.argL);
addHRegUse(u, HRmRead, i->ARM64in.VBinS.argR);
return;
+ case ARM64in_VTriD:
+ addHRegUse(u, HRmWrite, i->ARM64in.VTriD.dst);
+ addHRegUse(u, HRmRead, i->ARM64in.VTriD.arg1);
+ addHRegUse(u, HRmRead, i->ARM64in.VTriD.arg2);
+ addHRegUse(u, HRmRead, i->ARM64in.VTriD.arg3);
+ return;
+ case ARM64in_VTriS:
+ addHRegUse(u, HRmWrite, i->ARM64in.VTriS.dst);
+ addHRegUse(u, HRmRead, i->ARM64in.VTriS.arg1);
+ addHRegUse(u, HRmRead, i->ARM64in.VTriS.arg2);
+ addHRegUse(u, HRmRead, i->ARM64in.VTriS.arg3);
+ return;
case ARM64in_VCmpD:
addHRegUse(u, HRmRead, i->ARM64in.VCmpD.argL);
addHRegUse(u, HRmRead, i->ARM64in.VCmpD.argR);
@@ -2454,6 +2516,18 @@ void mapRegs_ARM64Instr ( HRegRemap* m, ARM64Instr* i, Bool mode64 )
i->ARM64in.VBinS.argL = lookupHRegRemap(m, i->ARM64in.VBinS.argL);
i->ARM64in.VBinS.argR = lookupHRegRemap(m, i->ARM64in.VBinS.argR);
return;
+ case ARM64in_VTriD:
+ i->ARM64in.VTriD.dst = lookupHRegRemap(m, i->ARM64in.VTriD.dst);
+ i->ARM64in.VTriD.arg1 = lookupHRegRemap(m, i->ARM64in.VTriD.arg1);
+ i->ARM64in.VTriD.arg2 = lookupHRegRemap(m, i->ARM64in.VTriD.arg2);
+ i->ARM64in.VTriD.arg3 = lookupHRegRemap(m, i->ARM64in.VTriD.arg3);
+ return;
+ case ARM64in_VTriS:
+ i->ARM64in.VTriS.dst = lookupHRegRemap(m, i->ARM64in.VTriS.dst);
+ i->ARM64in.VTriS.arg1 = lookupHRegRemap(m, i->ARM64in.VTriS.arg1);
+ i->ARM64in.VTriS.arg2 = lookupHRegRemap(m, i->ARM64in.VTriS.arg2);
+ i->ARM64in.VTriS.arg3 = lookupHRegRemap(m, i->ARM64in.VTriS.arg3);
+ return;
case ARM64in_VCmpD:
i->ARM64in.VCmpD.argL = lookupHRegRemap(m, i->ARM64in.VCmpD.argL);
i->ARM64in.VCmpD.argR = lookupHRegRemap(m, i->ARM64in.VCmpD.argR);
@@ -2812,7 +2886,8 @@ static inline UInt qregEnc ( HReg r )
#define X11110011 BITS8(1,1,1,1,0,0,1,1)
#define X11110101 BITS8(1,1,1,1,0,1,0,1)
#define X11110111 BITS8(1,1,1,1,0,1,1,1)
-
+#define X11111000 BITS8(1,1,1,1,1,0,0,0)
+#define X11111010 BITS8(1,1,1,1,1,0,1,0)
/* --- 4 fields --- */
@@ -2972,6 +3047,27 @@ static inline UInt X_3_6_1_6_6_5_5 ( UInt f1, UInt f2, UInt f3,
}
+static inline UInt X_3_8_5_1_5_5_5 ( UInt f1, UInt f2, UInt f3, UInt f4,
+ UInt f5, UInt f6, UInt f7 ) {
+ vassert(3+8+5+1+5+5+5 == 32);
+ vassert(f1 < (1<<3));
+ vassert(f2 < (1<<8));
+ vassert(f3 < (1<<5));
+ vassert(f4 < (1<<1));
+ vassert(f5 < (1<<5));
+ vassert(f6 < (1<<5));
+ vassert(f7 < (1<<5));
+ UInt w = 0;
+ w = (w << 3) | f1;
+ w = (w << 8) | f2;
+ w = (w << 5) | f3;
+ w = (w << 1) | f4;
+ w = (w << 5) | f5;
+ w = (w << 5) | f6;
+ w = (w << 5) | f7;
+ return w;
+}
+
//ZZ #define X0000 BITS4(0,0,0,0)
//ZZ #define X0001 BITS4(0,0,0,1)
//ZZ #define X0010 BITS4(0,0,1,0)
@@ -4339,6 +4435,44 @@ Int emit_ARM64Instr ( /*MB_MOD*/Bool* is_profInc,
= X_3_8_5_6_5_5(X000, X11110001, sM, (b1512 << 2) | X10, sN, sD);
goto done;
}
+ case ARM64in_VTriD: {
+ /* 31 20 15 14 9 4
+ 000 11111 010 m 0 a n d FMADD Dd,Dn,Dm,Da
+ ---------------- 1 ------ FMSUB -----------
+ */
+ UInt dD = dregEnc(i->ARM64in.VTriD.dst);
+ UInt dN = dregEnc(i->ARM64in.VTriD.arg1);
+ UInt dM = dregEnc(i->ARM64in.VTriD.arg2);
+ UInt dA = dregEnc(i->ARM64in.VTriD.arg3);
+ UInt b15 = 2; /* impossible */
+ switch (i->ARM64in.VTriD.op) {
+ case ARM64fpt_FMADD: b15 = 0; break;
+ case ARM64fpt_FMSUB: b15 = 1; break;
+ default: goto bad;
+ }
+ vassert(b15 < 2);
+ *p++ = X_3_8_5_1_5_5_5(X000, X11111010, dM, b15, dA, dN, dD);
+ goto done;
+ }
+ case ARM64in_VTriS: {
+ /* 31 20 15 14 9 4
+ 000 11111 000 m 0 a n d FMADD Dd,Dn,Dm,Da
+ ---------------- 1 ------ FMSUB -----------
+ */
+ UInt dD = dregEnc(i->ARM64in.VTriD.dst);
+ UInt dN = dregEnc(i->ARM64in.VTriD.arg1);
+ UInt dM = dregEnc(i->ARM64in.VTriD.arg2);
+ UInt dA = dregEnc(i->ARM64in.VTriD.arg3);
+ UInt b15 = 2; /* impossible */
+ switch (i->ARM64in.VTriD.op) {
+ case ARM64fpt_FMADD: b15 = 0; break;
+ case ARM64fpt_FMSUB: b15 = 1; break;
+ default: goto bad;
+ }
+ vassert(b15 < 2);
+ *p++ = X_3_8_5_1_5_5_5(X000, X11111000, dM, b15, dA, dN, dD);
+ goto done;
+ }
case ARM64in_VCmpD: {
/* 000 11110 01 1 m 00 1000 n 00 000 FCMP Dn, Dm */
UInt dN = dregEnc(i->ARM64in.VCmpD.argL);
diff --git a/VEX/priv/host_arm64_defs.h b/VEX/priv/host_arm64_defs.h
index 05dba7ab8..5a82564ce 100644
--- a/VEX/priv/host_arm64_defs.h
+++ b/VEX/priv/host_arm64_defs.h
@@ -289,6 +289,14 @@ typedef
}
ARM64FpBinOp;
+typedef
+ enum {
+ ARM64fpt_FMADD=105,
+ ARM64fpt_FMSUB,
+ ARM64fpt_INVALID
+ }
+ ARM64FpTriOp;
+
typedef
enum {
ARM64fpu_NEG=110,
@@ -498,6 +506,8 @@ typedef
ARM64in_VUnaryS,
ARM64in_VBinD,
ARM64in_VBinS,
+ ARM64in_VTriD,
+ ARM64in_VTriS,
ARM64in_VCmpD,
ARM64in_VCmpS,
ARM64in_VFCSel,
@@ -799,6 +809,22 @@ typedef
HReg argL;
HReg argR;
} VBinS;
+ /* 64-bit FP ternary arithmetic */
+ struct {
+ ARM64FpTriOp op;
+ HReg dst;
+ HReg arg1;
+ HReg arg2;
+ HReg arg3;
+ } VTriD;
+ /* 32-bit FP ternary arithmetic */
+ struct {
+ ARM64FpTriOp op;
+ HReg dst;
+ HReg arg1;
+ HReg arg2;
+ HReg arg3;
+ } VTriS;
/* 64-bit FP compare */
struct {
HReg argL;
@@ -970,6 +996,10 @@ extern ARM64Instr* ARM64Instr_VUnaryD ( ARM64FpUnaryOp op, HReg dst, HReg src );
extern ARM64Instr* ARM64Instr_VUnaryS ( ARM64FpUnaryOp op, HReg dst, HReg src );
extern ARM64Instr* ARM64Instr_VBinD ( ARM64FpBinOp op, HReg, HReg, HReg );
extern ARM64Instr* ARM64Instr_VBinS ( ARM64FpBinOp op, HReg, HReg, HReg );
+extern ARM64Instr* ARM64Instr_VTriD ( ARM64FpTriOp op, HReg dst,
+ HReg, HReg, HReg );
+extern ARM64Instr* ARM64Instr_VTriS ( ARM64FpTriOp op, HReg dst,
+ HReg, HReg, HReg );
extern ARM64Instr* ARM64Instr_VCmpD ( HReg argL, HReg argR );
extern ARM64Instr* ARM64Instr_VCmpS ( HReg argL, HReg argR );
extern ARM64Instr* ARM64Instr_VFCSel ( HReg dst, HReg argL, HReg argR,
diff --git a/VEX/priv/host_arm64_isel.c b/VEX/priv/host_arm64_isel.c
index 2f19eab81..da1218715 100644
--- a/VEX/priv/host_arm64_isel.c
+++ b/VEX/priv/host_arm64_isel.c
@@ -3255,6 +3255,25 @@ static HReg iselDblExpr_wrk ( ISelEnv* env, IRExpr* e )
}
}
+ if (e->tag == Iex_Qop) {
+ IRQop* qop = e->Iex.Qop.details;
+ ARM64FpTriOp triop = ARM64fpt_INVALID;
+ switch (qop->op) {
+ case Iop_MAddF64: triop = ARM64fpt_FMADD; break;
+ case Iop_MSubF64: triop = ARM64fpt_FMSUB; break;
+ default: break;
+ }
+ if (triop != ARM64fpt_INVALID) {
+ HReg N = iselDblExpr(env, qop->arg2);
+ HReg M = iselDblExpr(env, qop->arg3);
+ HReg A = iselDblExpr(env, qop->arg4);
+ HReg dst = newVRegD(env);
+ set_FPCR_rounding_mode(env, qop->arg1);
+ addInstr(env, ARM64Instr_VTriD(triop, dst, N, M, A));
+ return dst;
+ }
+ }
+
if (e->tag == Iex_ITE) {
/* ITE(ccexpr, iftrue, iffalse) */
ARM64CondCode cc;
@@ -3450,6 +3469,26 @@ static HReg iselFltExpr_wrk ( ISelEnv* env, IRExpr* e )
return dst;
}
+ if (e->tag == Iex_Qop) {
+ IRQop* qop = e->Iex.Qop.details;
+ ARM64FpTriOp triop = ARM64fpt_INVALID;
+ switch (qop->op) {
+ case Iop_MAddF32: triop = ARM64fpt_FMADD; break;
+ case Iop_MSubF32: triop = ARM64fpt_FMSUB; break;
+ default: break;
+ }
+
+ if (triop != ARM64fpt_INVALID) {
+ HReg N = iselFltExpr(env, qop->arg2);
+ HReg M = iselFltExpr(env, qop->arg3);
+ HReg A = iselFltExpr(env, qop->arg4);
+ HReg dst = newVRegD(env);
+ set_FPCR_rounding_mode(env, qop->arg1);
+ addInstr(env, ARM64Instr_VTriS(triop, dst, N, M, A));
+ return dst;
+ }
+ }
+
ppIRExpr(e);
vpanic("iselFltExpr_wrk");
}
diff --git a/none/tests/arm64/Makefile.am b/none/tests/arm64/Makefile.am
index 7b3ebbdca..4ecab36ad 100644
--- a/none/tests/arm64/Makefile.am
+++ b/none/tests/arm64/Makefile.am
@@ -10,14 +10,16 @@ EXTRA_DIST = \
integer.stdout.exp integer.stderr.exp integer.vgtest \
memory.stdout.exp memory.stderr.exp memory.vgtest \
atomics_v81.stdout.exp atomics_v81.stderr.exp atomics_v81.vgtest \
- simd_v81.stdout.exp simd_v81.stderr.exp simd_v81.vgtest
+ simd_v81.stdout.exp simd_v81.stderr.exp simd_v81.vgtest \
+ fmadd_sub.stdout.exp fmadd_sub.stderr.exp fmadd_sub.vgtest
check_PROGRAMS = \
allexec \
cvtf_imm \
fp_and_simd \
integer \
- memory
+ memory \
+ fmadd_sub
if BUILD_ARMV8_CRC_TESTS
check_PROGRAMS += crc32
diff --git a/none/tests/arm64/fmadd_sub.c b/none/tests/arm64/fmadd_sub.c
new file mode 100644
index 000000000..dcab22d1b
--- /dev/null
+++ b/none/tests/arm64/fmadd_sub.c
@@ -0,0 +1,98 @@
+#include <math.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#define COUNT 5
+
+static void
+print_float(const char *ident, float x)
+{
+ union
+ {
+ float f;
+ uint32_t i;
+ } u;
+
+ u.f = x;
+ printf("%s = %08x = %.17g\n", ident, u.i, x);
+}
+
+static void
+print_double(const char *ident, double x)
+{
+ union
+ {
+ double f;
+ uint64_t i;
+ } u;
+
+ u.f = x;
+ printf("%s = %016lx = %.17g\n", ident, u.i, x);
+}
+
+int
+main(int argc, char **argv)
+{
+ float x[] = { 55, 0.98076171874999996, 0, 1, 0xFFFFFFFF } ;
+ float y[] = { 0.69314718055994529, 1.015625, 0, 1, 0xFFFFFFFF };
+ float z[] = { 38.123094930796988, 1, 0, 1, 0xFFFFFFFF };
+ float dst = -5;
+
+ double dx[] = { 55, 0.98076171874999996, 0, 1, 0xFFFFFFFF } ;
+ double dy[] = { 0.69314718055994529, 1.015625, 0, 1, 0xFFFFFFFF };
+ double dz[] = { 38.123094930796988, 1, 0, 1, 0xFFFFFFFF };
+ double ddst= -5;
+
+ int i;
+
+ for (i = 0; i < COUNT; i++) {
+ //32bit variant
+ asm("fmadd %s0, %s1, %s2, %s3\n;" : "=w"(dst) : "w"(x[i]), "w"(y[i]), "w"(z[i]));
+ printf("FMADD 32bit: dst = z + x * y\n");
+ printf("%f = %f + %f * %f\n", dst, z[i], x[i], y[i]);
+ print_float("dst", dst);
+
+ // Floating-point negated fused multiply-add
+ asm("fnmadd %s0, %s1, %s2, %s3\n;" : "=w"(dst) : "w"(x[i]), "w"(y[i]), "w"(z[i]));
+ printf("FNMADD 32bit: dst = -z + (-x) * y\n");
+ printf("%f = -%f + (-%f) * %f\n", dst, z[i], x[i], y[i]);
+ print_float("dst", dst);
+
+ asm("fmsub %s0, %s1, %s2, %s3\n;" : "=w"(dst) : "w"(x[i]), "w"(y[i]), "w"(z[i]));
+ printf("FMSUB 32bit: dst = z + (-x) * y\n");
+ printf("%f = %f + (-%f) * %f\n", dst, z[i], x[i], y[i]);
+ print_float("dst", dst);
+
+ asm("fnmsub %s0, %s1, %s2, %s3\n;" : "=w"(dst) : "w"(x[i]), "w"(y[i]), "w"(z[i]));
+ printf("FNMSUB 32bit: dst = -z + x * y\n");
+ printf("%f = -%f + %f * %f\n", dst, z[i], x[i], y[i]);
+ print_float("dst", dst);
+
+ //64bit variant
+ asm("fmadd %d0, %d1, %d2, %d3\n;" : "=w"(ddst) : "w"(dx[i]), "w"(dy[i]), "w"(dz[i]));
+ printf("FMADD 64bit: dst = z + x * y\n");
+ printf("%f = %f + %f * %f\n", ddst, dz[i], dx[i], dy[i]);
+ print_double("dst", ddst);
+
+ asm("fnmadd %d0, %d1, %d2, %d3\n;" : "=w"(ddst) : "w"(dx[i]), "w"(dy[i]), "w"(dz[i]));
+ printf("FNMADD 64bit: dst = -z + (-x) * y\n");
+ printf("%f = -%f - %f * %f\n", ddst, dz[i], dx[i], dy[i]);
+ print_double("dst", ddst);
+
+ asm("fmsub %d0, %d1, %d2, %d3\n;" : "=w"(ddst) : "w"(dx[i]), "w"(dy[i]), "w"(dz[i]));
+ printf("FMSUB 64bit: dst = z + (-x) * y\n");
+ printf("%f = %f + (-%f) * %f\n", ddst, dz[i], dx[i], dy[i]);
+ print_double("dst", ddst);
+
+ asm("fnmsub %d0, %d1, %d2, %d3\n;" : "=w"(ddst) : "w"(dx[i]), "w"(dy[i]), "w"(dz[i]));
+ printf("FNMSUB 64bit: dst = -z + x * y\n");
+ printf("%f = -%f + %f * %f\n", ddst, dz[i], dx[i], dy[i]);
+ print_double("dst", ddst);
+
+ printf("\n");
+ }
+
+ return 0;
+}
+
diff --git a/none/tests/arm64/fmadd_sub.stderr.exp b/none/tests/arm64/fmadd_sub.stderr.exp
new file mode 100644
index 000000000..e69de29bb
diff --git a/none/tests/arm64/fmadd_sub.stdout.exp b/none/tests/arm64/fmadd_sub.stdout.exp
new file mode 100644
index 000000000..f1824b12b
--- /dev/null
+++ b/none/tests/arm64/fmadd_sub.stdout.exp
@@ -0,0 +1,125 @@
+FMADD 32bit: dst = z + x * y
+76.246193 = 38.123096 + 55.000000 * 0.693147
+dst = 42987e0d = 76.246192932128906
+FNMADD 32bit: dst = -z + (-x) * y
+-76.246193 = -38.123096 + (-55.000000) * 0.693147
+dst = c2987e0d = -76.246192932128906
+FMSUB 32bit: dst = z + (-x) * y
+0.000001 = 38.123096 + (-55.000000) * 0.693147
+dst = 35c00000 = 1.430511474609375e-06
+FNMSUB 32bit: dst = -z + x * y
+-0.000001 = -38.123096 + 55.000000 * 0.693147
+dst = b5c00000 = -1.430511474609375e-06
+FMADD 64bit: dst = z + x * y
+76.246190 = 38.123095 + 55.000000 * 0.693147
+dst = 40530fc1931f09c9 = 76.246189861593976
+FNMADD 64bit: dst = -z + (-x) * y
+-76.246190 = -38.123095 - 55.000000 * 0.693147
+dst = c0530fc1931f09c9 = -76.246189861593976
+FMSUB 64bit: dst = z + (-x) * y
+-0.000000 = 38.123095 + (-55.000000) * 0.693147
+dst = bce9000000000000 = -2.7755575615628914e-15
+FNMSUB 64bit: dst = -z + x * y
+0.000000 = -38.123095 + 55.000000 * 0.693147
+dst = 3ce9000000000000 = 2.7755575615628914e-15
+
+FMADD 32bit: dst = z + x * y
+1.996086 = 1.000000 + 0.980762 * 1.015625
+dst = 3fff7fc0 = 1.9960861206054688
+FNMADD 32bit: dst = -z + (-x) * y
+-1.996086 = -1.000000 + (-0.980762) * 1.015625
+dst = bfff7fc0 = -1.9960861206054688
+FMSUB 32bit: dst = z + (-x) * y
+0.003914 = 1.000000 + (-0.980762) * 1.015625
+dst = 3b80401a = 0.00391389150172472
+FNMSUB 32bit: dst = -z + x * y
+-0.003914 = -1.000000 + 0.980762 * 1.015625
+dst = bb80401a = -0.00391389150172472
+FMADD 64bit: dst = z + x * y
+1.996086 = 1.000000 + 0.980762 * 1.015625
+dst = 3fffeff800000000 = 1.9960861206054688
+FNMADD 64bit: dst = -z + (-x) * y
+-1.996086 = -1.000000 - 0.980762 * 1.015625
+dst = bfffeff800000000 = -1.9960861206054688
+FMSUB 64bit: dst = z + (-x) * y
+0.003914 = 1.000000 + (-0.980762) * 1.015625
+dst = 3f70080000000034 = 0.0039138793945312951
+FNMSUB 64bit: dst = -z + x * y
+-0.003914 = -1.000000 + 0.980762 * 1.015625
+dst = bf70080000000034 = -0.0039138793945312951
+
+FMADD 32bit: dst = z + x * y
+0.000000 = 0.000000 + 0.000000 * 0.000000
+dst = 00000000 = 0
+FNMADD 32bit: dst = -z + (-x) * y
+-0.000000 = -0.000000 + (-0.000000) * 0.000000
+dst = 80000000 = -0
+FMSUB 32bit: dst = z + (-x) * y
+0.000000 = 0.000000 + (-0.000000) * 0.000000
+dst = 00000000 = 0
+FNMSUB 32bit: dst = -z + x * y
+0.000000 = -0.000000 + 0.000000 * 0.000000
+dst = 00000000 = 0
+FMADD 64bit: dst = z + x * y
+0.000000 = 0.000000 + 0.000000 * 0.000000
+dst = 0000000000000000 = 0
+FNMADD 64bit: dst = -z + (-x) * y
+-0.000000 = -0.000000 - 0.000000 * 0.000000
+dst = 8000000000000000 = -0
+FMSUB 64bit: dst = z + (-x) * y
+0.000000 = 0.000000 + (-0.000000) * 0.000000
+dst = 0000000000000000 = 0
+FNMSUB 64bit: dst = -z + x * y
+0.000000 = -0.000000 + 0.000000 * 0.000000
+dst = 0000000000000000 = 0
+
+FMADD 32bit: dst = z + x * y
+2.000000 = 1.000000 + 1.000000 * 1.000000
+dst = 40000000 = 2
+FNMADD 32bit: dst = -z + (-x) * y
+-2.000000 = -1.000000 + (-1.000000) * 1.000000
+dst = c0000000 = -2
+FMSUB 32bit: dst = z + (-x) * y
+0.000000 = 1.000000 + (-1.000000) * 1.000000
+dst = 00000000 = 0
+FNMSUB 32bit: dst = -z + x * y
+0.000000 = -1.000000 + 1.000000 * 1.000000
+dst = 00000000 = 0
+FMADD 64bit: dst = z + x * y
+2.000000 = 1.000000 + 1.000000 * 1.000000
+dst = 4000000000000000 = 2
+FNMADD 64bit: dst = -z + (-x) * y
+-2.000000 = -1.000000 - 1.000000 * 1.000000
+dst = c000000000000000 = -2
+FMSUB 64bit: dst = z + (-x) * y
+0.000000 = 1.000000 + (-1.000000) * 1.000000
+dst = 0000000000000000 = 0
+FNMSUB 64bit: dst = -z + x * y
+0.000000 = -1.000000 + 1.000000 * 1.000000
+dst = 0000000000000000 = 0
+
+FMADD 32bit: dst = z + x * y
+18446744073709551616.000000 = 4294967296.000000 + 4294967296.000000 * 4294967296.000000
+dst = 5f800000 = 1.8446744073709552e+19
+FNMADD 32bit: dst = -z + (-x) * y
+-18446744073709551616.000000 = -4294967296.000000 + (-4294967296.000000) * 4294967296.000000
+dst = df800000 = -1.8446744073709552e+19
+FMSUB 32bit: dst = z + (-x) * y
+-18446744073709551616.000000 = 4294967296.000000 + (-4294967296.000000) * 4294967296.000000
+dst = df800000 = -1.8446744073709552e+19
+FNMSUB 32bit: dst = -z + x * y
+18446744073709551616.000000 = -4294967296.000000 + 4294967296.000000 * 4294967296.000000
+dst = 5f800000 = 1.8446744073709552e+19
+FMADD 64bit: dst = z + x * y
+18446744069414584320.000000 = 4294967295.000000 + 4294967295.000000 * 4294967295.000000
+dst = 43efffffffe00000 = 1.8446744069414584e+19
+FNMADD 64bit: dst = -z + (-x) * y
+-18446744069414584320.000000 = -4294967295.000000 - 4294967295.000000 * 4294967295.000000
+dst = c3efffffffe00000 = -1.8446744069414584e+19
+FMSUB 64bit: dst = z + (-x) * y
+-18446744060824649728.000000 = 4294967295.000000 + (-4294967295.000000) * 4294967295.000000
+dst = c3efffffffa00000 = -1.844674406082465e+19
+FNMSUB 64bit: dst = -z + x * y
+18446744060824649728.000000 = -4294967295.000000 + 4294967295.000000 * 4294967295.000000
+dst = 43efffffffa00000 = 1.844674406082465e+19
+
diff --git a/none/tests/arm64/fmadd_sub.vgtest b/none/tests/arm64/fmadd_sub.vgtest
new file mode 100644
index 000000000..b4c53eea4
--- /dev/null
+++ b/none/tests/arm64/fmadd_sub.vgtest
@@ -0,0 +1,3 @@
+prog: fmadd_sub
+prereq: test -x fmadd_sub
+vgopts: -q
--
2.18.4