git clone https://git.lucas.co/go_mono.git
vector: tweak the asm register assignment.
This makes the fixed point and floating point code more similar.
Benchmarks don't show any significant change.
Change-Id: I723fa1605eaa248b40e784201b680c16cc3d26a2
Reviewed-on: https://go-review.googlesource.com/32134
Reviewed-by: David Crawshaw <[email protected]>
vector/acc_amd64.s | 60 +++++++++++++++++++++++++++---------------------------
vector/gen.go | 16 +++++++--------
2 files changed, 38 insertions(+), 38 deletions(-)
diff --git a/vector/acc_amd64.s b/vector/acc_amd64.s
index 299e052..6a424bc 100644
--- a/vector/acc_amd64.s
+++ b/vector/acc_amd64.s
@@ -577,9 +577,9 @@ fxAccMaskEnd:
// xmm0 scratch
// xmm1 x
// xmm2 y, z
-// xmm3 flAlmost65536
+// xmm3 flSignMask
// xmm4 flOne
-// xmm5 flSignMask
+// xmm5 flAlmost65536
// xmm6 gather
// xmm7 offset
// xmm8 scatterAndMulBy0x101
@@ -608,12 +608,12 @@ TEXT ·floatingAccumulateOpOverSIMD(SB), NOSPLIT, $8-48
ORL $0x6000, AX
MOVL AX, mxcsrNew-4(SP)
- // flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
- // flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
// flSignMask := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
- MOVOU flAlmost65536<>(SB), X3
+ // flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
+ // flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
+ MOVOU flSignMask<>(SB), X3
MOVOU flOne<>(SB), X4
- MOVOU flSignMask<>(SB), X5
+ MOVOU flAlmost65536<>(SB), X5
// gather := XMM(see above) // PSHUFB shuffle mask.
// scatterAndMulBy0x101 := XMM(see above) // PSHUFB shuffle mask.
@@ -659,10 +659,10 @@ flAccOpOverLoop4:
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
- MOVOU X5, X2
+ MOVOU X3, X2
ANDPS X1, X2
MINPS X4, X2
- MULPS X3, X2
+ MULPS X5, X2
// z = convertToInt32(y)
LDMXCSR mxcsrNew-4(SP)
@@ -738,10 +738,10 @@ flAccOpOverLoop1:
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
- MOVOU X5, X2
+ MOVOU X3, X2
ANDPS X1, X2
MINPS X4, X2
- MULPS X3, X2
+ MULPS X5, X2
// z = convertToInt32(y)
LDMXCSR mxcsrNew-4(SP)
@@ -791,9 +791,9 @@ flAccOpOverEnd:
// xmm0 scratch
// xmm1 x
// xmm2 y, z
-// xmm3 flAlmost65536
+// xmm3 flSignMask
// xmm4 flOne
-// xmm5 flSignMask
+// xmm5 flAlmost65536
// xmm6 gather
// xmm7 offset
// xmm8 -
@@ -822,12 +822,12 @@ TEXT ·floatingAccumulateOpSrcSIMD(SB), NOSPLIT, $8-48
ORL $0x6000, AX
MOVL AX, mxcsrNew-4(SP)
- // flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
- // flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
// flSignMask := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
- MOVOU flAlmost65536<>(SB), X3
+ // flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
+ // flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
+ MOVOU flSignMask<>(SB), X3
MOVOU flOne<>(SB), X4
- MOVOU flSignMask<>(SB), X5
+ MOVOU flAlmost65536<>(SB), X5
// gather := XMM(see above) // PSHUFB shuffle mask.
MOVOU gather<>(SB), X6
@@ -867,10 +867,10 @@ flAccOpSrcLoop4:
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
- MOVOU X5, X2
+ MOVOU X3, X2
ANDPS X1, X2
MINPS X4, X2
- MULPS X3, X2
+ MULPS X5, X2
// z = convertToInt32(y)
LDMXCSR mxcsrNew-4(SP)
@@ -906,10 +906,10 @@ flAccOpSrcLoop1:
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
- MOVOU X5, X2
+ MOVOU X3, X2
ANDPS X1, X2
MINPS X4, X2
- MULPS X3, X2
+ MULPS X5, X2
// z = convertToInt32(y)
LDMXCSR mxcsrNew-4(SP)
@@ -945,9 +945,9 @@ flAccOpSrcEnd:
// xmm0 scratch
// xmm1 x
// xmm2 y, z
-// xmm3 flAlmost65536
+// xmm3 flSignMask
// xmm4 flOne
-// xmm5 flSignMask
+// xmm5 flAlmost65536
// xmm6 -
// xmm7 offset
// xmm8 -
@@ -976,12 +976,12 @@ TEXT ·floatingAccumulateMaskSIMD(SB), NOSPLIT, $8-48
ORL $0x6000, AX
MOVL AX, mxcsrNew-4(SP)
- // flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
- // flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
// flSignMask := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
- MOVOU flAlmost65536<>(SB), X3
+ // flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
+ // flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
+ MOVOU flSignMask<>(SB), X3
MOVOU flOne<>(SB), X4
- MOVOU flSignMask<>(SB), X5
+ MOVOU flAlmost65536<>(SB), X5
// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
XORPS X7, X7
@@ -1018,10 +1018,10 @@ flAccMaskLoop4:
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
- MOVOU X5, X2
+ MOVOU X3, X2
ANDPS X1, X2
MINPS X4, X2
- MULPS X3, X2
+ MULPS X5, X2
// z = convertToInt32(y)
LDMXCSR mxcsrNew-4(SP)
@@ -1055,10 +1055,10 @@ flAccMaskLoop1:
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
- MOVOU X5, X2
+ MOVOU X3, X2
ANDPS X1, X2
MINPS X4, X2
- MULPS X3, X2
+ MULPS X5, X2
// z = convertToInt32(y)
LDMXCSR mxcsrNew-4(SP)
diff --git a/vector/gen.go b/vector/gen.go
index cafd860..28b298b 100644
--- a/vector/gen.go
+++ b/vector/gen.go
@@ -242,13 +242,13 @@ const (
sizeOfUint32 = 4
fxXMM3 = `-`
- flXMM3 = `flAlmost65536`
+ flXMM3 = `flSignMask`
fxXMM4 = `-`
flXMM4 = `flOne`
fxXMM5 = `fxAlmost65536`
- flXMM5 = `flSignMask`
+ flXMM5 = `flAlmost65536`
oneArgLoadArgs = `
MOVQ buf_base+0(FP), DI
@@ -281,12 +281,12 @@ const (
MOVOU fxAlmost65536<>(SB), X5
`
flLoadXMMRegs = `
- // flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
- // flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
// flSignMask := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
- MOVOU flAlmost65536<>(SB), X3
+ // flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
+ // flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
+ MOVOU flSignMask<>(SB), X3
MOVOU flOne<>(SB), X4
- MOVOU flSignMask<>(SB), X5
+ MOVOU flAlmost65536<>(SB), X5
`
fxAdd = `PADDD`
@@ -312,10 +312,10 @@ const (
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
- MOVOU X5, X2
+ MOVOU X3, X2
ANDPS X1, X2
MINPS X4, X2
- MULPS X3, X2
+ MULPS X5, X2
`
fxConvertToInt32 = `