[llvm] acbc0c5 - [AArch64][GlobalISel] Widen non-pow-2 types for shifts before clamping.

[Amara Emerson via llvm-commits]

Author: Amara Emerson
Date: 2021-07-24T15:50:43-07:00
New Revision: acbc0c5f0ebd8b7ebfa2eb3ae77428eb83c428c5

URL: https://github.com/llvm/llvm-project/commit/acbc0c5f0ebd8b7ebfa2eb3ae77428eb83c428c5
DIFF: https://github.com/llvm/llvm-project/commit/acbc0c5f0ebd8b7ebfa2eb3ae77428eb83c428c5.diff

LOG: [AArch64][GlobalISel] Widen non-pow-2 types for shifts before clamping.

For types like s96, we don't want to clamp to s64, we want to first widen to
s128 and then narrow it. Otherwise we end up with impossible to legalize types.

Added: 
    

Modified: 
    llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
    llvm/test/CodeGen/AArch64/GlobalISel/legalize-shift.mir

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
index e2788276a9361..08e4a119127c9 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
@@ -139,9 +139,9 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
           {v4s32, v4s32},
           {v2s64, v2s64},
       })
+      .widenScalarToNextPow2(0)
       .clampScalar(1, s32, s64)
       .clampScalar(0, s32, s64)
-      .widenScalarToNextPow2(0)
       .clampNumElements(0, v2s32, v4s32)
       .clampNumElements(0, v2s64, v2s64)
       .moreElementsToNextPow2(0)

diff  --git a/llvm/test/CodeGen/AArch64/GlobalISel/legalize-shift.mir b/llvm/test/CodeGen/AArch64/GlobalISel/legalize-shift.mir
index bf0a95776d5de..ef97250744378 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/legalize-shift.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/legalize-shift.mir
@@ -1,5 +1,5 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -O0 -march=aarch64 -run-pass=legalizer %s -o - | FileCheck %s
+# RUN: llc -O0 -march=aarch64 -run-pass=legalizer -global-isel-abort=1 %s -o - | FileCheck %s
 # RUN: llc -O0 -debugify-and-strip-all-safe -march=aarch64 -run-pass=legalizer -verify-machineinstrs %s -o - | FileCheck %s --check-prefix=CHECK
 ---
 name:            test_shift
@@ -474,3 +474,39 @@ body:             |
     %2:_(<8 x s8>) = G_LSHR %0, %1
     $d0 = COPY %2
 ...
+---
+name: test_lshr_nonpow2
+body:             |
+  bb.0:
+    ; CHECK-LABEL: name: test_lshr_nonpow2
+    ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $x0
+    ; CHECK: %a:_(s32) = COPY $w1
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES %a(s32), %a(s32)
+    ; CHECK: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES %a(s32), [[C]](s32)
+    ; CHECK: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 64
+    ; CHECK: [[SUB:%[0-9]+]]:_(s64) = G_SUB [[COPY]], [[C1]]
+    ; CHECK: [[SUB1:%[0-9]+]]:_(s64) = G_SUB [[C1]], [[COPY]]
+    ; CHECK: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+    ; CHECK: [[ICMP:%[0-9]+]]:_(s32) = G_ICMP intpred(ult), [[COPY]](s64), [[C1]]
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s1) = G_TRUNC [[ICMP]](s32)
+    ; CHECK: [[ICMP1:%[0-9]+]]:_(s32) = G_ICMP intpred(eq), [[COPY]](s64), [[C2]]
+    ; CHECK: [[TRUNC1:%[0-9]+]]:_(s1) = G_TRUNC [[ICMP1]](s32)
+    ; CHECK: [[LSHR:%[0-9]+]]:_(s64) = G_LSHR [[MV1]], [[COPY]](s64)
+    ; CHECK: [[LSHR1:%[0-9]+]]:_(s64) = G_LSHR [[MV]], [[COPY]](s64)
+    ; CHECK: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[MV1]], [[SUB1]](s64)
+    ; CHECK: [[OR:%[0-9]+]]:_(s64) = G_OR [[LSHR1]], [[SHL]]
+    ; CHECK: [[LSHR2:%[0-9]+]]:_(s64) = G_LSHR [[MV1]], [[SUB]](s64)
+    ; CHECK: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[TRUNC]](s1), [[OR]], [[LSHR2]]
+    ; CHECK: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[TRUNC1]](s1), [[MV]], [[SELECT]]
+    ; CHECK: [[SELECT2:%[0-9]+]]:_(s64) = G_SELECT [[TRUNC]](s1), [[LSHR]], [[C2]]
+    ; CHECK: %d1:_(s32), %d2:_(s32) = G_UNMERGE_VALUES [[SELECT1]](s64)
+    ; CHECK: $w0 = COPY %d2(s32)
+    %0:_(s64) = COPY $x0
+    %a:_(s32) = COPY $w1
+    %merge:_(s96) = G_MERGE_VALUES %a, %a, %a
+    %lshr:_(s96) = G_LSHR %merge, %0
+    %d1:_(s32), %d2:_(s32), %d3:_(s32) = G_UNMERGE_VALUES %lshr
+    $w0 = COPY %d2(s32)
+
+...