arm64: Optimise GT/GE/LT/LE comparisons to constant zero

[jonathandavies-arm]

No description provided.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[jonathandavies-arm]

Please can I have a review @EgorBo?
There was an old PR with this work (#112863) and this is mostly the same except that I fixed the issue with the tests.

@dotnet/arm64-contrib

[xtqqczze]

Diffs show a small number of regressions, e.g.:

; System.IO.Tests.File_NotifyFilter_Tests:FileSystemWatcher_File_NotifyFilter_Size(int):this (FullOpts)
@@ -160,11 +160,13 @@ G_M49668_IG03:        ; bbWeight=0.50, gcVars=0000000000008000 {V03}, gcrefRegs=
             movk    x0, #1 LSL #32
             str     x0, [x19, #0x18]
             mov     w23, wzr
-            and     w0, w20, #124
             cmp     w20, #8
             ccmp    w20, #16, z, ne
-            ccmp    w0, #0, 0, ne
-            bgt     G_M49668_IG04
+            cset    x0, eq
+            ands    w1, w20, #124
+            cset    x1, gt
+            orr     w0, w0, w1
+            cbnz    w0, G_M49668_IG04
             movz    x0, #0xD1FFAB1E      // code for System.PlatformDetection:get_IsWindows7():bool
             movk    x0, #0xD1FFAB1E LSL #16
             movk    x0, #1 LSL #32
@@ -174,7 +176,7 @@ G_M49668_IG03:        ; bbWeight=0.50, gcVars=0000000000008000 {V03}, gcrefRegs=
             cbz     w0, G_M49668_IG05
             cmp     w20, #4
             bne     G_M49668_IG05
-						;; size=132 bbWeight=0.50 PerfScore 12.75
+						;; size=140 bbWeight=0.50 PerfScore 13.25

[EgorBo]

Thanks for adding tests!

[AndyAyersMS]

@EgorBo if this is ready can we merge?

[xtqqczze]

@MihuBot -arm

[xtqqczze]

Diffs show a small number of regressions, e.g.:

; System.IO.Tests.File_NotifyFilter_Tests:FileSystemWatcher_File_NotifyFilter_Size(int):this (FullOpts)
@@ -160,11 +160,13 @@ G_M49668_IG03:        ; bbWeight=0.50, gcVars=0000000000008000 {V03}, gcrefRegs=
             movk    x0, #1 LSL #32
             str     x0, [x19, #0x18]
             mov     w23, wzr
-            and     w0, w20, #124
             cmp     w20, #8
             ccmp    w20, #16, z, ne
-            ccmp    w0, #0, 0, ne
-            bgt     G_M49668_IG04
+            cset    x0, eq
+            ands    w1, w20, #124
+            cset    x1, gt
+            orr     w0, w0, w1
+            cbnz    w0, G_M49668_IG04
             movz    x0, #0xD1FFAB1E      // code for System.PlatformDetection:get_IsWindows7():bool
             movk    x0, #0xD1FFAB1E LSL #16
             movk    x0, #1 LSL #32
@@ -174,7 +176,7 @@ G_M49668_IG03:        ; bbWeight=0.50, gcVars=0000000000008000 {V03}, gcrefRegs=
             cbz     w0, G_M49668_IG05
             cmp     w20, #4
             bne     G_M49668_IG05
-						;; size=132 bbWeight=0.50 PerfScore 12.75
+						;; size=140 bbWeight=0.50 PerfScore 13.25

@EgorBo Which optimization is being missed here, TryLowerAndOrToCCMP?

[EgorBo]

Diffs show a small number of regressions, e.g.:

; System.IO.Tests.File_NotifyFilter_Tests:FileSystemWatcher_File_NotifyFilter_Size(int):this (FullOpts)
@@ -160,11 +160,13 @@ G_M49668_IG03:        ; bbWeight=0.50, gcVars=0000000000008000 {V03}, gcrefRegs=
             movk    x0, #1 LSL #32
             str     x0, [x19, #0x18]
             mov     w23, wzr
-            and     w0, w20, #124
             cmp     w20, #8
             ccmp    w20, #16, z, ne
-            ccmp    w0, #0, 0, ne
-            bgt     G_M49668_IG04
+            cset    x0, eq
+            ands    w1, w20, #124
+            cset    x1, gt
+            orr     w0, w0, w1
+            cbnz    w0, G_M49668_IG04
             movz    x0, #0xD1FFAB1E      // code for System.PlatformDetection:get_IsWindows7():bool
             movk    x0, #0xD1FFAB1E LSL #16
             movk    x0, #1 LSL #32
@@ -174,7 +176,7 @@ G_M49668_IG03:        ; bbWeight=0.50, gcVars=0000000000008000 {V03}, gcrefRegs=
             cbz     w0, G_M49668_IG05
             cmp     w20, #4
             bne     G_M49668_IG05
-						;; size=132 bbWeight=0.50 PerfScore 12.75
+						;; size=140 bbWeight=0.50 PerfScore 13.25

@EgorBo Which optimization is being missed here, TryLowerAndOrToCCMP?

hm.. i do not recall this, last round of SPMI was a clean improvement if my memory is not failing me?

[xtqqczze]

hm.. i do not recall this, last round of SPMI was a clean improvement if my memory is not failing me?

I think there were 3-4 similar regressions, the build has since been deleted.