[llvm] f269bb7 - [ARM] Fix crash trying to generate i1 immediates

[David Green via llvm-commits]

Author: David Green
Date: 2020-06-16T12:27:24+01:00
New Revision: f269bb7da0162f7171395071a59ec7584318383e

URL: https://github.com/llvm/llvm-project/commit/f269bb7da0162f7171395071a59ec7584318383e
DIFF: https://github.com/llvm/llvm-project/commit/f269bb7da0162f7171395071a59ec7584318383e.diff

LOG: [ARM] Fix crash trying to generate i1 immediates

These code patterns attempt to call isVMOVModifiedImm on a splat of i1
values, leading to an unreachable being hit. I've guarded the call on a
more specific set of sizes, as i1 vectors are legal under MVE.

Differential Revision: https://reviews.llvm.org/D81860

Added: 
    

Modified: 
    llvm/lib/Target/ARM/ARMISelLowering.cpp
    llvm/test/CodeGen/Thumb2/mve-vmovimm.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index f163bd620f26..a1362a40dc31 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -12467,7 +12467,8 @@ static SDValue PerformANDCombine(SDNode *N,
   bool HasAnyUndefs;
   if (BVN && (Subtarget->hasNEON() || Subtarget->hasMVEIntegerOps()) &&
       BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
-    if (SplatBitSize <= 64) {
+    if (SplatBitSize == 8 || SplatBitSize == 16 || SplatBitSize == 32 ||
+        SplatBitSize == 64) {
       EVT VbicVT;
       SDValue Val = isVMOVModifiedImm((~SplatBits).getZExtValue(),
                                       SplatUndef.getZExtValue(), SplatBitSize,
@@ -12759,7 +12760,8 @@ static SDValue PerformORCombine(SDNode *N,
   bool HasAnyUndefs;
   if (BVN && (Subtarget->hasNEON() || Subtarget->hasMVEIntegerOps()) &&
       BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
-    if (SplatBitSize <= 64) {
+    if (SplatBitSize == 8 || SplatBitSize == 16 || SplatBitSize == 32 ||
+        SplatBitSize == 64) {
       EVT VorrVT;
       SDValue Val =
           isVMOVModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),

diff  --git a/llvm/test/CodeGen/Thumb2/mve-vmovimm.ll b/llvm/test/CodeGen/Thumb2/mve-vmovimm.ll
index f2101001673f..2ca8f93f5385 100644
--- a/llvm/test/CodeGen/Thumb2/mve-vmovimm.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-vmovimm.ll
@@ -514,3 +514,75 @@ entry:
   %o = or <8 x i16> %i, <i16 -1, i16 0, i16 -1, i16 -1, i16 -1, i16 0, i16 -1, i16 -1>
   ret <8 x i16> %o
 }
+
+define arm_aapcs_vfpcc <4 x i32> @i1and_vmov(<4 x i32> %a, <4 x i32> %b, i32 %c) {
+; CHECKLE-LABEL: i1and_vmov:
+; CHECKLE:       @ %bb.0: @ %entry
+; CHECKLE-NEXT:    cmp r0, #0
+; CHECKLE-NEXT:    mov.w r1, #15
+; CHECKLE-NEXT:    cset r0, eq
+; CHECKLE-NEXT:    and r0, r0, #1
+; CHECKLE-NEXT:    rsbs r0, r0, #0
+; CHECKLE-NEXT:    ands r0, r1
+; CHECKLE-NEXT:    vmsr p0, r0
+; CHECKLE-NEXT:    vpsel q0, q0, q1
+; CHECKLE-NEXT:    bx lr
+;
+; CHECKBE-LABEL: i1and_vmov:
+; CHECKBE:       @ %bb.0: @ %entry
+; CHECKBE-NEXT:    cmp r0, #0
+; CHECKBE-NEXT:    mov.w r1, #15
+; CHECKBE-NEXT:    cset r0, eq
+; CHECKBE-NEXT:    vrev64.32 q2, q1
+; CHECKBE-NEXT:    and r0, r0, #1
+; CHECKBE-NEXT:    vrev64.32 q1, q0
+; CHECKBE-NEXT:    rsbs r0, r0, #0
+; CHECKBE-NEXT:    ands r0, r1
+; CHECKBE-NEXT:    vmsr p0, r0
+; CHECKBE-NEXT:    vpsel q1, q1, q2
+; CHECKBE-NEXT:    vrev64.32 q0, q1
+; CHECKBE-NEXT:    bx lr
+entry:
+  %c1 = icmp eq i32 %c, zeroinitializer
+  %broadcast.splatinsert1967 = insertelement <4 x i1> undef, i1 %c1, i32 0
+  %broadcast.splat1968 = shufflevector <4 x i1> %broadcast.splatinsert1967, <4 x i1> undef, <4 x i32> zeroinitializer
+  %l699 = and <4 x i1> %broadcast.splat1968, <i1 true, i1 false, i1 false, i1 false>
+  %s = select <4 x i1> %l699, <4 x i32> %a, <4 x i32> %b
+  ret <4 x i32> %s
+}
+
+define arm_aapcs_vfpcc <4 x i32> @i1or_vmov(<4 x i32> %a, <4 x i32> %b, i32 %c) {
+; CHECKLE-LABEL: i1or_vmov:
+; CHECKLE:       @ %bb.0: @ %entry
+; CHECKLE-NEXT:    cmp r0, #0
+; CHECKLE-NEXT:    mov.w r1, #15
+; CHECKLE-NEXT:    cset r0, eq
+; CHECKLE-NEXT:    and r0, r0, #1
+; CHECKLE-NEXT:    rsbs r0, r0, #0
+; CHECKLE-NEXT:    orrs r0, r1
+; CHECKLE-NEXT:    vmsr p0, r0
+; CHECKLE-NEXT:    vpsel q0, q0, q1
+; CHECKLE-NEXT:    bx lr
+;
+; CHECKBE-LABEL: i1or_vmov:
+; CHECKBE:       @ %bb.0: @ %entry
+; CHECKBE-NEXT:    cmp r0, #0
+; CHECKBE-NEXT:    mov.w r1, #15
+; CHECKBE-NEXT:    cset r0, eq
+; CHECKBE-NEXT:    vrev64.32 q2, q1
+; CHECKBE-NEXT:    and r0, r0, #1
+; CHECKBE-NEXT:    vrev64.32 q1, q0
+; CHECKBE-NEXT:    rsbs r0, r0, #0
+; CHECKBE-NEXT:    orrs r0, r1
+; CHECKBE-NEXT:    vmsr p0, r0
+; CHECKBE-NEXT:    vpsel q1, q1, q2
+; CHECKBE-NEXT:    vrev64.32 q0, q1
+; CHECKBE-NEXT:    bx lr
+entry:
+  %c1 = icmp eq i32 %c, zeroinitializer
+  %broadcast.splatinsert1967 = insertelement <4 x i1> undef, i1 %c1, i32 0
+  %broadcast.splat1968 = shufflevector <4 x i1> %broadcast.splatinsert1967, <4 x i1> undef, <4 x i32> zeroinitializer
+  %l699 = or <4 x i1> %broadcast.splat1968, <i1 true, i1 false, i1 false, i1 false>
+  %s = select <4 x i1> %l699, <4 x i32> %a, <4 x i32> %b
+  ret <4 x i32> %s
+}