[release-branch.go1.18] cmd/compile: fix conditional move rule on PPC64

Similar to CL 456556 but for ppc64 instead of arm64.

Change docs about how booleans are stored in registers for ppc64.
We now don't promise to keep the upper bits zeroed; they might be junk.

To test, I changed the boolean generation instructions (MOVBZload* and ISEL*
with boolean type) to OR in 0x100 to the result. all.bash still passed,
so I think nothing else is depending on the upper bits of booleans.

Update #57211

Change-Id: Ie66f8934a0dafa34d0a8c2a37324868d959a852c
Reviewed-on: https://go-review.googlesource.com/c/go/+/456437
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: Keith Randall <khr@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: KAMPANAT THUMWONG (KONG PC) <1992kongpc.kth@gmail.com>
Run-TryBot: Archana Ravindar <aravind5@in.ibm.com>
Reviewed-on: https://go-review.googlesource.com/c/go/+/456735
Reviewed-by: Than McIntosh <thanm@google.com>

diff --git a/src/cmd/compile/internal/ssa/gen/PPC64.rules b/src/cmd/compile/internal/ssa/gen/PPC64.rules
index a90a3d0937dd793a6544f0b61b9e9e85c7a5b2b8..f83ed78bab3a210c6f7769f09e5e2578ab7a98ce 100644 (file)
--- a/src/cmd/compile/internal/ssa/gen/PPC64.rules
+++ b/src/cmd/compile/internal/ssa/gen/PPC64.rules
@@ -562,9 +562,9 @@
 ((EQ|NE|LT|LE|GT|GE) (CMPconst [0] z:(XOR x y)) yes no) && z.Uses == 1 => ((EQ|NE|LT|LE|GT|GE) (XORCC x y) yes no)
 
 // Only lower after bool is lowered. It should always lower. This helps ensure the folding below happens reliably.
-(CondSelect x y bool) && flagArg(bool) == nil => (ISEL [6] x y (CMPWconst [0] bool))
+(CondSelect x y bool) && flagArg(bool) == nil => (ISEL [6] x y (Select1 <types.TypeFlags> (ANDCCconst [1] bool)))
 // Fold any CR -> GPR -> CR transfers when applying the above rule.
-(ISEL [6] x y (CMPWconst [0] (ISELB [c] one cmp))) => (ISEL [c] x y cmp)
+(ISEL [6] x y (Select1 (ANDCCconst [1] (ISELB [c] one cmp)))) => (ISEL [c] x y cmp)
 
 // Lowering loads
 (Load <t> ptr mem) && (is64BitInt(t) || isPtr(t)) => (MOVDload ptr mem)

diff --git a/src/cmd/compile/internal/ssa/gen/PPC64Ops.go b/src/cmd/compile/internal/ssa/gen/PPC64Ops.go
index 59d8af1a9d3952fb1bfeee88af01200fc618cf6a..d18cbcc787140fdb9c7940da8e684e720e263239 100644 (file)
--- a/src/cmd/compile/internal/ssa/gen/PPC64Ops.go
+++ b/src/cmd/compile/internal/ssa/gen/PPC64Ops.go
@@ -11,8 +11,8 @@ import "strings"
 
 // Notes:
 //  - Less-than-64-bit integer types live in the low portion of registers.
-//    For now, the upper portion is junk; sign/zero-extension might be optimized in the future, but not yet.
-//  - Boolean types are zero or 1; stored in a byte, but loaded with AMOVBZ so the upper bytes of a register are zero.
+//    The upper portion is junk.
+//  - Boolean types are zero or 1; stored in a byte, with upper bytes of the register containing junk.
 //  - *const instructions may use a constant larger than the instruction can encode.
 //    In this case the assembler expands to multiple instructions and uses tmp
 //    register (R31).

diff --git a/src/cmd/compile/internal/ssa/rewritePPC64.go b/src/cmd/compile/internal/ssa/rewritePPC64.go
index 5a28b9d4f7687f26d7ed8751869f7223fd663c85..c7bcc248fc3984fd91d85ff489b7cffeddec2a96 100644 (file)
--- a/src/cmd/compile/internal/ssa/rewritePPC64.go
+++ b/src/cmd/compile/internal/ssa/rewritePPC64.go
@@ -1167,9 +1167,10 @@ func rewriteValuePPC64_OpCondSelect(v *Value) bool {
        v_1 := v.Args[1]
        v_0 := v.Args[0]
        b := v.Block
+       typ := &b.Func.Config.Types
        // match: (CondSelect x y bool)
        // cond: flagArg(bool) == nil
-       // result: (ISEL [6] x y (CMPWconst [0] bool))
+       // result: (ISEL [6] x y (Select1 <types.TypeFlags> (ANDCCconst [1] bool)))
        for {
                x := v_0
                y := v_1
@@ -1179,9 +1180,11 @@ func rewriteValuePPC64_OpCondSelect(v *Value) bool {
                }
                v.reset(OpPPC64ISEL)
                v.AuxInt = int32ToAuxInt(6)
-               v0 := b.NewValue0(v.Pos, OpPPC64CMPWconst, types.TypeFlags)
-               v0.AuxInt = int32ToAuxInt(0)
-               v0.AddArg(bool)
+               v0 := b.NewValue0(v.Pos, OpSelect1, types.TypeFlags)
+               v1 := b.NewValue0(v.Pos, OpPPC64ANDCCconst, types.NewTuple(typ.Int, types.TypeFlags))
+               v1.AuxInt = int64ToAuxInt(1)
+               v1.AddArg(bool)
+               v0.AddArg(v1)
                v.AddArg3(x, y, v0)
                return true
        }
@@ -5895,7 +5898,7 @@ func rewriteValuePPC64_OpPPC64ISEL(v *Value) bool {
                v.AddArg(y)
                return true
        }
-       // match: (ISEL [6] x y (CMPWconst [0] (ISELB [c] one cmp)))
+       // match: (ISEL [6] x y (Select1 (ANDCCconst [1] (ISELB [c] one cmp))))
        // result: (ISEL [c] x y cmp)
        for {
                if auxIntToInt32(v.AuxInt) != 6 {
@@ -5903,15 +5906,19 @@ func rewriteValuePPC64_OpPPC64ISEL(v *Value) bool {
                }
                x := v_0
                y := v_1
-               if v_2.Op != OpPPC64CMPWconst || auxIntToInt32(v_2.AuxInt) != 0 {
+               if v_2.Op != OpSelect1 {
                        break
                }
                v_2_0 := v_2.Args[0]
-               if v_2_0.Op != OpPPC64ISELB {
+               if v_2_0.Op != OpPPC64ANDCCconst || auxIntToInt64(v_2_0.AuxInt) != 1 {
+                       break
+               }
+               v_2_0_0 := v_2_0.Args[0]
+               if v_2_0_0.Op != OpPPC64ISELB {
                        break
                }
-               c := auxIntToInt32(v_2_0.AuxInt)
-               cmp := v_2_0.Args[1]
+               c := auxIntToInt32(v_2_0_0.AuxInt)
+               cmp := v_2_0_0.Args[1]
                v.reset(OpPPC64ISEL)
                v.AuxInt = int32ToAuxInt(c)
                v.AddArg3(x, y, cmp)