[llvm] 59588f0 - [SVE][ISel] Ensure explicit gather/scatter offset extension isn't lost.

[Paul Walker via llvm-commits]

Author: Paul Walker
Date: 2022-04-29T14:20:13+01:00
New Revision: 59588f0a3d47e3e366d675b8f9724c10a6222c0e

URL: https://github.com/llvm/llvm-project/commit/59588f0a3d47e3e366d675b8f9724c10a6222c0e
DIFF: https://github.com/llvm/llvm-project/commit/59588f0a3d47e3e366d675b8f9724c10a6222c0e.diff

LOG: [SVE][ISel] Ensure explicit gather/scatter offset extension isn't lost.

getGatherScatterIndexIsExtended currently looks through all
SIGN_EXTEND_INREG operations regardless of their input type.  This
patch restricts the code to only look through i32->i64 extensions,
which are the ones supported implicitly by SVE addressing modes.

Differential Revision: https://reviews.llvm.org/D123318

Added: 
    

Modified: 
    llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
    llvm/test/CodeGen/AArch64/sve-gather-scatter-addr-opts.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 3f6a36fe49f4..89bbced4c738 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -4630,9 +4630,13 @@ unsigned getSignExtendedGatherOpcode(unsigned Opcode) {
 }
 
 bool getGatherScatterIndexIsExtended(SDValue Index) {
+  // Ignore non-pointer sized indices.
+  if (Index.getValueType() != MVT::nxv2i64)
+    return false;
+
   unsigned Opcode = Index.getOpcode();
   if (Opcode == ISD::SIGN_EXTEND_INREG)
-    return true;
+    return cast<VTSDNode>(Index.getOperand(1))->getVT() == MVT::nxv2i32;
 
   if (Opcode == ISD::AND) {
     SDValue Splat = Index.getOperand(1);

diff  --git a/llvm/test/CodeGen/AArch64/sve-gather-scatter-addr-opts.ll b/llvm/test/CodeGen/AArch64/sve-gather-scatter-addr-opts.ll
index 4fdf4a106dbc..f4b4a033c343 100644
--- a/llvm/test/CodeGen/AArch64/sve-gather-scatter-addr-opts.ll
+++ b/llvm/test/CodeGen/AArch64/sve-gather-scatter-addr-opts.ll
@@ -374,11 +374,11 @@ define <vscale x 2 x i64> @masked_gather_nxv2i64_null_with__vec_plus_imm_offsets
   ret <vscale x 2 x i64> %data
 }
 
-; TODO: The generated code is wrong because we've lost the sign extension which
-; defines bits offsets[8:31].
 define <vscale x 4 x i32> @masked_gather_nxv4i32_s8_offsets(i32* %base, <vscale x 4 x i8> %offsets, <vscale x 4 x i1> %mask) #0 {
 ; CHECK-LABEL: masked_gather_nxv4i32_s8_offsets:
 ; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p1.s
+; CHECK-NEXT:    sxtb z0.s, p1/m, z0.s
 ; CHECK-NEXT:    ld1w { z0.s }, p0/z, [x0, z0.s, sxtw #2]
 ; CHECK-NEXT:    ret
   %offsets.sext = sext <vscale x 4 x i8> %offsets to <vscale x 4 x i32>
@@ -399,12 +399,13 @@ define <vscale x 4 x i32> @masked_gather_nxv4i32_u8_offsets(i32* %base, <vscale
   ret <vscale x 4 x i32> %data
 }
 
-; TODO: The generated code is wrong because we've lost the sign extension which
-; defines bits offsets[8:31] and we're also replicating offset[31] across
+; TODO: The generated code is wrong because we're replicating offset[31] across
 ; offset[32:63] even though the IR has explicitly zero'd those bits.
 define <vscale x 4 x i32> @masked_gather_nxv4i32_u32s8_offsets(i32* %base, <vscale x 4 x i8> %offsets, <vscale x 4 x i1> %mask) #0 {
 ; CHECK-LABEL: masked_gather_nxv4i32_u32s8_offsets:
 ; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p1.s
+; CHECK-NEXT:    sxtb z0.s, p1/m, z0.s
 ; CHECK-NEXT:    ld1w { z0.s }, p0/z, [x0, z0.s, sxtw #2]
 ; CHECK-NEXT:    ret
   %offsets.sext = sext <vscale x 4 x i8> %offsets to <vscale x 4 x i32>
@@ -456,11 +457,11 @@ define void @masked_scatter_nxv2i64_null_with__vec_plus_imm_offsets(<vscale x 2
   ret void
 }
 
-; TODO: The generated code is wrong because we've lost the sign extension which
-; defines bits offsets[8:31].
 define void @masked_scatter_nxv4i32_s8_offsets(i32* %base, <vscale x 4 x i8> %offsets, <vscale x 4 x i1> %mask, <vscale x 4 x i32> %data) #0 {
 ; CHECK-LABEL: masked_scatter_nxv4i32_s8_offsets:
 ; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p1.s
+; CHECK-NEXT:    sxtb z0.s, p1/m, z0.s
 ; CHECK-NEXT:    st1w { z1.s }, p0, [x0, z0.s, sxtw #2]
 ; CHECK-NEXT:    ret
   %offsets.sext = sext <vscale x 4 x i8> %offsets to <vscale x 4 x i32>
@@ -481,12 +482,13 @@ define void @masked_scatter_nxv4i32_u8_offsets(i32* %base, <vscale x 4 x i8> %of
   ret void
 }
 
-; TODO: The generated code is wrong because we've lost the sign extension which
-; defines bits offsets[8:31] and we're also replicating offset[31] across
+; TODO: The generated code is wrong because we're replicating offset[31] across
 ; offset[32:63] even though the IR has explicitly zero'd those bits.
 define void @masked_scatter_nxv4i32_u32s8_offsets(i32* %base, <vscale x 4 x i8> %offsets, <vscale x 4 x i1> %mask, <vscale x 4 x i32> %data) #0 {
 ; CHECK-LABEL: masked_scatter_nxv4i32_u32s8_offsets:
 ; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p1.s
+; CHECK-NEXT:    sxtb z0.s, p1/m, z0.s
 ; CHECK-NEXT:    st1w { z1.s }, p0, [x0, z0.s, sxtw #2]
 ; CHECK-NEXT:    ret
   %offsets.sext = sext <vscale x 4 x i8> %offsets to <vscale x 4 x i32>