cmd/compile: change riscv64 Eq32/Neq32 to zero extend before subtraction

As done with other equality tests, zero extend before subtraction rather than
after (or in this case, at the same time). While at face value this appears to
require more instructions, in reality it allows for most sign extensions to
be completely eliminated due to correctly typed loads. Existing optimisations
(such as subtraction of zero) then become more effective.

This removes more than 10,000 instructions from the Go binary and in particular,
a writeBarrier check only requires three instructions (AUIPC, LWU, BNEZ) instead
of the current four (AUIPC, LWU, NEGW, BNEZ).

Change-Id: I7afdc1921c4916ddbd414c3b3f5c2089107ec016
Reviewed-on: https://go-review.googlesource.com/c/go/+/274066
Trust: Joel Sing <joel@sing.id.au>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>

diff --git a/src/cmd/compile/internal/ssa/gen/RISCV64.rules b/src/cmd/compile/internal/ssa/gen/RISCV64.rules
index 4380a5efef719d52601d4858586b90cf57c86932..15361fd37a925a82b50323e76d5920d91c79dee0 100644 (file)
--- a/src/cmd/compile/internal/ssa/gen/RISCV64.rules
+++ b/src/cmd/compile/internal/ssa/gen/RISCV64.rules
@@ -251,7 +251,7 @@
 
 (EqPtr x y) => (SEQZ (SUB <x.Type> x y))
 (Eq64  x y) => (SEQZ (SUB <x.Type> x y))
-(Eq32  x y) => (SEQZ (SUBW <x.Type> x y))
+(Eq32  x y) => (SEQZ (SUB <x.Type> (ZeroExt32to64 x) (ZeroExt32to64 y)))
 (Eq16  x y) => (SEQZ (SUB <x.Type> (ZeroExt16to64 x) (ZeroExt16to64 y)))
 (Eq8   x y) => (SEQZ (SUB <x.Type> (ZeroExt8to64  x) (ZeroExt8to64  y)))
 (Eq64F ...) => (FEQD ...)
@@ -259,7 +259,7 @@
 
 (NeqPtr x y) => (SNEZ (SUB <x.Type> x y))
 (Neq64  x y) => (SNEZ (SUB <x.Type> x y))
-(Neq32  x y) => (SNEZ (SUBW <x.Type> x y))
+(Neq32  x y) => (SNEZ (SUB <x.Type> (ZeroExt32to64 x) (ZeroExt32to64 y)))
 (Neq16  x y) => (SNEZ (SUB <x.Type> (ZeroExt16to64 x) (ZeroExt16to64 y)))
 (Neq8   x y) => (SNEZ (SUB <x.Type> (ZeroExt8to64  x) (ZeroExt8to64  y)))
 (Neq64F ...) => (FNED ...)

diff --git a/src/cmd/compile/internal/ssa/rewriteRISCV64.go b/src/cmd/compile/internal/ssa/rewriteRISCV64.go
index fb507b65c4ce9a5e9be8d4b17c772d66535fecf6..1abdf1aa0639d91880232ca93c3e8d0d1c46631e 100644 (file)
--- a/src/cmd/compile/internal/ssa/rewriteRISCV64.go
+++ b/src/cmd/compile/internal/ssa/rewriteRISCV64.go
@@ -890,14 +890,19 @@ func rewriteValueRISCV64_OpEq32(v *Value) bool {
        v_1 := v.Args[1]
        v_0 := v.Args[0]
        b := v.Block
+       typ := &b.Func.Config.Types
        // match: (Eq32 x y)
-       // result: (SEQZ (SUBW <x.Type> x y))
+       // result: (SEQZ (SUB <x.Type> (ZeroExt32to64 x) (ZeroExt32to64 y)))
        for {
                x := v_0
                y := v_1
                v.reset(OpRISCV64SEQZ)
-               v0 := b.NewValue0(v.Pos, OpRISCV64SUBW, x.Type)
-               v0.AddArg2(x, y)
+               v0 := b.NewValue0(v.Pos, OpRISCV64SUB, x.Type)
+               v1 := b.NewValue0(v.Pos, OpZeroExt32to64, typ.UInt64)
+               v1.AddArg(x)
+               v2 := b.NewValue0(v.Pos, OpZeroExt32to64, typ.UInt64)
+               v2.AddArg(y)
+               v0.AddArg2(v1, v2)
                v.AddArg(v0)
                return true
        }
@@ -2466,14 +2471,19 @@ func rewriteValueRISCV64_OpNeq32(v *Value) bool {
        v_1 := v.Args[1]
        v_0 := v.Args[0]
        b := v.Block
+       typ := &b.Func.Config.Types
        // match: (Neq32 x y)
-       // result: (SNEZ (SUBW <x.Type> x y))
+       // result: (SNEZ (SUB <x.Type> (ZeroExt32to64 x) (ZeroExt32to64 y)))
        for {
                x := v_0
                y := v_1
                v.reset(OpRISCV64SNEZ)
-               v0 := b.NewValue0(v.Pos, OpRISCV64SUBW, x.Type)
-               v0.AddArg2(x, y)
+               v0 := b.NewValue0(v.Pos, OpRISCV64SUB, x.Type)
+               v1 := b.NewValue0(v.Pos, OpZeroExt32to64, typ.UInt64)
+               v1.AddArg(x)
+               v2 := b.NewValue0(v.Pos, OpZeroExt32to64, typ.UInt64)
+               v2.AddArg(y)
+               v0.AddArg2(v1, v2)
                v.AddArg(v0)
                return true
        }