cmd/compile/internal/ssa: remove redundant zeroextensions on amd64

Some instructions operating on <= 32 bits also zero out upper 32bits.
Remove zeroextensions of such values. Triggers a few times during
all.bash. Also removes ugly code like:
MOVL CX,CX

Change-Id: I66a46c190dd6929b7e3c52f3fe6b967768d00638
Reviewed-on: https://go-review.googlesource.com/58090
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>

diff --git a/src/cmd/compile/internal/ssa/gen/AMD64.rules b/src/cmd/compile/internal/ssa/gen/AMD64.rules
index 0f1249a1d68b3020d0908ac507605a54d3b40c28..0bbe14dd8c1ebbdc50b3cf3636b5fad56bb054ee 100644 (file)
--- a/src/cmd/compile/internal/ssa/gen/AMD64.rules
+++ b/src/cmd/compile/internal/ssa/gen/AMD64.rules
@@ -995,6 +995,8 @@
 (MOVLQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVLload <v.Type> [off] {sym} ptr mem)
 (MOVLQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVLload <v.Type> [off] {sym} ptr mem)
 
+(MOVLQZX x) && zeroUpper32Bits(x,3) -> x
+
 (MOVBQZX x:(MOVBloadidx1 [off] {sym} ptr idx mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVBloadidx1 <v.Type> [off] {sym} ptr idx mem)
 (MOVWQZX x:(MOVWloadidx1 [off] {sym} ptr idx mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVWloadidx1 <v.Type> [off] {sym} ptr idx mem)
 (MOVWQZX x:(MOVWloadidx2 [off] {sym} ptr idx mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVWloadidx2 <v.Type> [off] {sym} ptr idx mem)

diff --git a/src/cmd/compile/internal/ssa/rewrite.go b/src/cmd/compile/internal/ssa/rewrite.go
index 561946f640df439bdc3a830a802c70935d5c3e0d..2002a1ab590dc17511ed022f888ce4d6b65e2932 100644 (file)
--- a/src/cmd/compile/internal/ssa/rewrite.go
+++ b/src/cmd/compile/internal/ssa/rewrite.go
@@ -648,3 +648,35 @@ func overlap(offset1, size1, offset2, size2 int64) bool {
        }
        return false
 }
+
+// check if value zeroes out upper 32-bit of 64-bit register.
+// depth limits recursion depth. In AMD64.rules 3 is used as limit,
+// because it catches same amount of cases as 4.
+func zeroUpper32Bits(x *Value, depth int) bool {
+       switch x.Op {
+       case OpAMD64MOVLconst, OpAMD64MOVLload, OpAMD64MOVLQZX, OpAMD64MOVLloadidx1,
+               OpAMD64MOVWload, OpAMD64MOVWloadidx1, OpAMD64MOVBload, OpAMD64MOVBloadidx1,
+               OpAMD64MOVLloadidx4, OpAMD64ADDLmem, OpAMD64SUBLmem, OpAMD64ANDLmem,
+               OpAMD64ORLmem, OpAMD64XORLmem, OpAMD64CVTTSD2SL,
+               OpAMD64ADDL, OpAMD64ADDLconst, OpAMD64SUBL, OpAMD64SUBLconst,
+               OpAMD64ANDL, OpAMD64ANDLconst, OpAMD64ORL, OpAMD64ORLconst,
+               OpAMD64XORL, OpAMD64XORLconst, OpAMD64NEGL, OpAMD64NOTL:
+               return true
+       case OpArg, OpSelect0, OpSelect1:
+               return x.Type.Width == 4
+       case OpPhi:
+               // Phis can use each-other as an arguments, instead of tracking visited values,
+               // just limit recursion depth.
+               if depth <= 0 {
+                       return false
+               }
+               for i := range x.Args {
+                       if !zeroUpper32Bits(x.Args[i], depth-1) {
+                               return false
+                       }
+               }
+               return true
+
+       }
+       return false
+}

diff --git a/src/cmd/compile/internal/ssa/rewriteAMD64.go b/src/cmd/compile/internal/ssa/rewriteAMD64.go
index 24eda50f40717499e1b1c882bdd9c9b1d2f90c38..e707fcd519c2438c9bf1432a680a843de4d1e142 100644 (file)
--- a/src/cmd/compile/internal/ssa/rewriteAMD64.go
+++ b/src/cmd/compile/internal/ssa/rewriteAMD64.go
@@ -6957,6 +6957,19 @@ func rewriteValueAMD64_OpAMD64MOVLQZX_0(v *Value) bool {
                v0.AddArg(mem)
                return true
        }
+       // match: (MOVLQZX x)
+       // cond: zeroUpper32Bits(x,3)
+       // result: x
+       for {
+               x := v.Args[0]
+               if !(zeroUpper32Bits(x, 3)) {
+                       break
+               }
+               v.reset(OpCopy)
+               v.Type = x.Type
+               v.AddArg(x)
+               return true
+       }
        // match: (MOVLQZX x:(MOVLloadidx1 [off] {sym} ptr idx mem))
        // cond: x.Uses == 1 && clobber(x)
        // result: @x.Block (MOVLloadidx1 <v.Type> [off] {sym} ptr idx mem)