Trying this CL again, with a fixed test that allows platforms
to disagree on the exact behavior of converting NaNs.
We store 32-bit floating point constants in a 64-bit field, by
converting that 32-bit float to 64-bit float to store it, and convert
it back to use it.
That works for *almost* all floating-point constants. The exception is
signaling NaNs. The round trip described above means we can't represent
a 32-bit signaling NaN, because conversions strip the signaling bit.
To fix this issue, just forbid NaNs as floating-point constants in SSA
form. This shouldn't affect any real-world code, as people seldom
constant-propagate NaNs (except in test code).
Additionally, NaNs are somewhat underspecified (which of the many NaNs
do you get when dividing 0/0?), so when cross-compiling there's a
danger of using the compiler machine's NaN regime for some math, and
the target machine's NaN regime for other math. Better to use the
target machine's NaN regime always.
Update #36400
Change-Id: Idf203b688a15abceabbd66ba290d4e9f63619ecb
Reviewed-on: https://go-review.googlesource.com/c/go/+/221790
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
This reverts CL 213477.
Reason for revert: tests are failing on linux-mips*-rtrk builders.
Change-Id: I8168f7450890233f1bd7e53930b73693c26d4dc0
Reviewed-on: https://go-review.googlesource.com/c/go/+/220897
Run-TryBot: Bryan C. Mills <bcmills@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
We store 32-bit floating point constants in a 64-bit field, by
converting that 32-bit float to 64-bit float to store it, and convert
it back to use it.
That works for *almost* all floating-point constants. The exception is
signaling NaNs. The round trip described above means we can't represent
a 32-bit signaling NaN, because conversions strip the signaling bit.
To fix this issue, just forbid NaNs as floating-point constants in SSA
form. This shouldn't affect any real-world code, as people seldom
constant-propagate NaNs (except in test code).
Additionally, NaNs are somewhat underspecified (which of the many NaNs
do you get when dividing 0/0?), so when cross-compiling there's a
danger of using the compiler machine's NaN regime for some math, and
the target machine's NaN regime for other math. Better to use the
target machine's NaN regime always.
This has been a bug since 1.10, and there's an easy workaround
(declare a global varaible containing the signaling NaN pattern, and
use that as the argument to math.Float32frombits) so we'll fix it in
1.15.
Fixes#36400
Update #36399
Change-Id: Icf155e743281560eda2eed953d19a829552ccfda
Reviewed-on: https://go-review.googlesource.com/c/go/+/213477
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
This API was added for #25819, where it was discussed as math.FMA.
The commit adding it used math.Fma, presumably for consistency
with the rest of the unusual names in package math
(Sincos, Acosh, Erfcinv, Float32bits, etc).
I believe that using an idiomatic Go name is more important here
than consistency with these other names, most of which are historical
baggage from C's standard library.
Early additions like Float32frombits happened before "uppercase for export"
(so they were originally like "float32frombits") and they were not properly
reconsidered when we uppercased the symbols to export them.
That's a mistake we live with.
The names of functions we have added since then, and even a few
that were legacy, are more properly Go-cased, such as IsNaN, IsInf,
and RoundToEven, rather than Isnan, Isinf, and Roundtoeven.
And also constants like MaxFloat32.
For new API, we should keep using proper Go-cased symbols
instead of minimally-upper-cased-C symbols.
So math.FMA, not math.Fma.
This API has not yet been released, so this change does not break
the compatibility promise.
This CL also modifies cmd/compile, since the compiler knows
the name of the function. I could have stopped at changing the
string constants, but it seemed to make more sense to use a
consistent casing everywhere.
Change-Id: I0f6f3407f41e99bfa8239467345c33945088896e
Reviewed-on: https://go-review.googlesource.com/c/go/+/205317
Run-TryBot: Russ Cox <rsc@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
This change introduces an arm intrinsic that generates the FMULAD
instruction for the fused-multiply-add operation on systems that
support it. System support is detected via cpu.ARM.HasVFPv4. A rewrite
rule translates the generic intrinsic to FMULAD.
Updates #25819.
Change-Id: I8459e5dd1cdbdca35f88a78dbeb7d387f1e20efa
Reviewed-on: https://go-review.googlesource.com/c/go/+/142117
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
To permit ssa-level optimization, this change introduces an amd64 intrinsic
that generates the VFMADD231SD instruction for the fused-multiply-add
operation on systems that support it. System support is detected via
cpu.X86.HasFMA. A rewrite rule can then translate the generic ssa intrinsic
("Fma") to VFMADD231SD.
The benchmark compares the software implementation (old) with the intrinsic
(new).
name old time/op new time/op delta
Fma-4 27.2ns ± 1% 1.0ns ± 9% -96.48% (p=0.008 n=5+5)
Updates #25819.
Change-Id: I966655e5f96817a5d06dff5942418a3915b09584
Reviewed-on: https://go-review.googlesource.com/c/go/+/137156
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
In order to make math.FMA a compiler intrinsic for ISAs like ARM64,
PPC64[le], and S390X, a generic 3-argument opcode "Fma" is provided and
rewritten as
ARM64: (Fma x y z) -> (FMADDD z x y)
PPC64: (Fma x y z) -> (FMADD x y z)
S390X: (Fma x y z) -> (FMADD z x y)
Updates #25819.
Change-Id: Ie5bc628311e6feeb28ddf9adaa6e702c8c291efa
Reviewed-on: https://go-review.googlesource.com/c/go/+/131959
Run-TryBot: Akhil Indurti <aindurti@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
On modern 64bit CPUs a SHR, SHL or AND instruction take 1 cycle to execute.
A pair of shifts that operate on the same register will take 2 cycles
and needs to wait for the input register value to be available.
Large constants used to mask the high bits of a register with an AND
instruction can not be encoded as an immediate in the AND instruction
on amd64 and therefore need to be loaded into a register with a MOV
instruction.
However that MOV instruction is not dependent on the output register and
on many CPUs does not compete with the AND or shift instructions for
execution ports.
Using a pair of shifts to mask high bits instead of an AND to mask high
bits of a register has a shorter encoding and uses one less general
purpose register but is slower due to taking one clock cycle longer
if there is no register pressure that would make the AND variant need to
generate a spill.
For example the instructions emitted for (x & 1 << 63) before this CL are:
48c1ea3f SHRQ $0x3f, DX
48c1e23f SHLQ $0x3f, DX
after this CL the instructions are the same as GCC and LLVM use:
48b80000000000000080 MOVQ $0x8000000000000000, AX
4821d0 ANDQ DX, AX
Some platforms such as arm64 already have SSA optimization rules to fuse
two shift instructions back into an AND.
Removing the general rule to rewrite AND to SHR+SHL speeds up this benchmark:
var GlobalU uint
func BenchmarkAndHighBits(b *testing.B) {
x := uint(0)
for i := 0; i < b.N; i++ {
x &= 1 << 63
}
GlobalU = x
}
amd64/darwin on Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz:
name old time/op new time/op delta
AndHighBits-4 0.61ns ± 6% 0.42ns ± 6% -31.42% (p=0.000 n=25+25):
'go run run.go -all_codegen -v codegen' passes with following adjustments:
ARM64: The BFXIL pattern ((x << lc) >> rc | y & ac) needed adjustment
since ORshiftRL generation fusing '>> rc' and '|' interferes
with matching ((x << lc) >> rc) to generate UBFX. Previously
ORshiftLL was created first using the shifts generated for (y & ac).
S390X: Add rules for abs and copysign to match use of AND instead of SHIFTs.
Updates #33826
Updates #32781
Change-Id: I5a59f6239660d53c029cd22dfb44ddf39f93a56c
Reviewed-on: https://go-review.googlesource.com/c/go/+/196810
Run-TryBot: Martin Möhrmann <moehrmann@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
On modern 64bit CPUs a SHR, SHL or AND instruction take 1 cycle to execute.
A pair of shifts that operate on the same register will take 2 cycles
and needs to wait for the input register value to be available.
Large constants used to mask the high bits of a register with an AND
instruction can not be encoded as an immediate in the AND instruction
on amd64 and therefore need to be loaded into a register with a MOV
instruction.
However that MOV instruction is not dependent on the output register and
on many CPUs does not compete with the AND or shift instructions for
execution ports.
Using a pair of shifts to mask high bits instead of an AND to mask high
bits of a register has a shorter encoding and uses one less general
purpose register but is slower due to taking one clock cycle longer
if there is no register pressure that would make the AND variant need to
generate a spill.
For example the instructions emitted for (x & 1 << 63) before this CL are:
48c1ea3f SHRQ $0x3f, DX
48c1e23f SHLQ $0x3f, DX
after this CL the instructions are the same as GCC and LLVM use:
48b80000000000000080 MOVQ $0x8000000000000000, AX
4821d0 ANDQ DX, AX
Some platforms such as arm64 already have SSA optimization rules to fuse
two shift instructions back into an AND.
Removing the general rule to rewrite AND to SHR+SHL speeds up this benchmark:
var GlobalU uint
func BenchmarkAndHighBits(b *testing.B) {
x := uint(0)
for i := 0; i < b.N; i++ {
x &= 1 << 63
}
GlobalU = x
}
amd64/darwin on Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz:
name old time/op new time/op delta
AndHighBits-4 0.61ns ± 6% 0.42ns ± 6% -31.42% (p=0.000 n=25+25):
'go run run.go -all_codegen -v codegen' passes with following adjustments:
ARM64: The BFXIL pattern ((x << lc) >> rc | y & ac) needed adjustment
since ORshiftRL generation fusing '>> rc' and '|' interferes
with matching ((x << lc) >> rc) to generate UBFX. Previously
ORshiftLL was created first using the shifts generated for (y & ac).
S390X: Add rules for abs and copysign to match use of AND instead of SHIFTs.
Updates #33826
Updates #32781
Change-Id: I43227da76b625de03fbc51117162b23b9c678cdb
Reviewed-on: https://go-review.googlesource.com/c/go/+/194297
Run-TryBot: Martin Möhrmann <martisch@uos.de>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
On modern 64bit CPUs a SHR, SHL or AND instruction take 1 cycle to execute.
A pair of shifts that operate on the same register will take 2 cycles
and needs to wait for the input register value to be available.
Large constants used to mask the high bits of a register with an AND
instruction can not be encoded as an immediate in the AND instruction
on amd64 and therefore need to be loaded into a register with a MOV
instruction.
However that MOV instruction is not dependent on the output register and
on many CPUs does not compete with the AND or shift instructions for
execution ports.
Using a pair of shifts to mask high bits instead of an AND to mask high
bits of a register has a shorter encoding and uses one less general
purpose register but is slower due to taking one clock cycle longer
if there is no register pressure that would make the AND variant need to
generate a spill.
For example the instructions emitted for (x & 1 << 63) before this CL are:
48c1ea3f SHRQ $0x3f, DX
48c1e23f SHLQ $0x3f, DX
after this CL the instructions are the same as GCC and LLVM use:
48b80000000000000080 MOVQ $0x8000000000000000, AX
4821d0 ANDQ DX, AX
Some platforms such as arm64 already have SSA optimization rules to fuse
two shift instructions back into an AND.
Removing the general rule to rewrite AND to SHR+SHL speeds up this benchmark:
var GlobalU uint
func BenchmarkAndHighBits(b *testing.B) {
x := uint(0)
for i := 0; i < b.N; i++ {
x &= 1 << 63
}
GlobalU = x
}
amd64/darwin on Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz:
name old time/op new time/op delta
AndHighBits-4 0.61ns ± 6% 0.42ns ± 6% -31.42% (p=0.000 n=25+25):
Updates #33826
Updates #32781
Change-Id: I862d3587446410c447b9a7265196b57f85358633
Reviewed-on: https://go-review.googlesource.com/c/go/+/191780
Run-TryBot: Martin Möhrmann <moehrmann@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
math.RoundToEven can be done by one arm64 instruction FRINTND, intrinsify it to improve performance.
The current pure Go implementation of the function Abs is translated into five instructions on arm64:
str, ldr, and, str, ldr. The intrinsic implementation requires only one instruction, so in terms of
performance, intrinsify it is worthwhile.
Benchmarks:
name old time/op new time/op delta
Abs-8 3.50ns ± 0% 1.50ns ± 0% -57.14% (p=0.000 n=10+10)
RoundToEven-8 9.26ns ± 0% 1.50ns ± 0% -83.80% (p=0.000 n=10+10)
Change-Id: I9456b26ab282b544dfac0154fc86f17aed96ac3d
Reviewed-on: https://go-review.googlesource.com/116535
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
The CL 132915 added the wrong codegen test for math.Copysign(c, -1),
it should test that AND is not emitted. This CL fixes this error.
Change-Id: Ida1d3d54ebfc7f238abccbc1f70f914e1b5bfd91
Reviewed-on: https://go-review.googlesource.com/134815
Reviewed-by: Giovanni Bajo <rasky@develer.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Giovanni Bajo <rasky@develer.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
mips64 softfloat support is based on mips implementation and introduces
new enviroment variable GOMIPS64.
GOMIPS64 is a GOARCH=mips64{,le} specific option, for a choice between
hard-float and soft-float. Valid values are 'hardfloat' (default) and
'softfloat'. It is passed to the assembler as
'GOMIPS64_{hardfloat,softfloat}'.
Change-Id: I7f73078627f7cb37c588a38fb5c997fe09c56134
Reviewed-on: https://go-review.googlesource.com/108475
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This change implements math.Round as an intrinsic on ppc64x so it can be
done using a single instruction.
benchmark old ns/op new ns/op delta
BenchmarkRound-16 2.60 0.69 -73.46%
Change-Id: I9408363e96201abdfc73ced7bcd5f0c29db006a8
Reviewed-on: https://go-review.googlesource.com/109395
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Lynn Boger <laboger@linux.vnet.ibm.com>
This CL makes the codegen testsuite automatically test all
architecture variants for architecture specified in tests. For
instance, if a test file specifies a "arm" test, it will be
automatically run on all GOARM variants (5,6,7), to increase
the coverage.
The CL also introduces a syntax to specify only a specific
variant (eg: "arm/7") in case the test makes sense only there.
The same syntax also allows to specify the operating system
in case it matters (eg: "plan9/386/sse2").
Fixes#24658
Change-Id: I2eba8b918f51bb6a77a8431a309f8b71af07ea22
Reviewed-on: https://go-review.googlesource.com/107315
Run-TryBot: Giovanni Bajo <rasky@develer.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
