[llvm] e628092 - [X86][MMX] Optimize MMX shift intrinsics.

[via llvm-commits]

Author: Wang, Pengfei
Date: 2020-07-11T11:16:23+08:00
New Revision: e6280925249c87c11568a305a074581cc073bd45

URL: https://github.com/llvm/llvm-project/commit/e6280925249c87c11568a305a074581cc073bd45
DIFF: https://github.com/llvm/llvm-project/commit/e6280925249c87c11568a305a074581cc073bd45.diff

LOG: [X86][MMX] Optimize MMX shift intrinsics.

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D83534

Added: 
    

Modified: 
    llvm/lib/Target/X86/X86ISelLowering.cpp
    llvm/test/CodeGen/X86/mmx-intrinsics.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 695b6ef35f11..721b262aa433 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -25236,6 +25236,9 @@ SDValue X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
       // Clamp out of bounds shift amounts since they will otherwise be masked
       // to 8-bits which may make it no longer out of bounds.
       unsigned ShiftAmount = C->getAPIntValue().getLimitedValue(255);
+      if (ShiftAmount == 0)
+        return Op.getOperand(1);
+
       return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op.getValueType(),
                          Op.getOperand(0), Op.getOperand(1),
                          DAG.getTargetConstant(ShiftAmount, DL, MVT::i32));

diff  --git a/llvm/test/CodeGen/X86/mmx-intrinsics.ll b/llvm/test/CodeGen/X86/mmx-intrinsics.ll
index 48d4ad0490f7..abc8fea58014 100644
--- a/llvm/test/CodeGen/X86/mmx-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/mmx-intrinsics.ll
@@ -311,6 +311,19 @@ entry:
   ret i64 %4
 }
 
+define i64 @test72_2(<1 x i64> %a) nounwind readnone optsize ssp {
+; ALL-LABEL: @test72_2
+; ALL-NOT: psraw
+entry:
+  %0 = bitcast <1 x i64> %a to <4 x i16>
+  %mmx_var.i = bitcast <4 x i16> %0 to x86_mmx
+  %1 = tail call x86_mmx @llvm.x86.mmx.psrai.w(x86_mmx %mmx_var.i, i32 0) nounwind
+  %2 = bitcast x86_mmx %1 to <4 x i16>
+  %3 = bitcast <4 x i16> %2 to <1 x i64>
+  %4 = extractelement <1 x i64> %3, i32 0
+  ret i64 %4
+}
+
 declare x86_mmx @llvm.x86.mmx.psrli.q(x86_mmx, i32) nounwind readnone
 
 define i64 @test71(<1 x i64> %a) nounwind readnone optsize ssp {
@@ -339,6 +352,19 @@ entry:
   ret i64 %4
 }
 
+define i64 @test70_2(<1 x i64> %a) nounwind readnone optsize ssp {
+; ALL-LABEL: @test70_2
+; ALL-NOT: psrld
+entry:
+  %0 = bitcast <1 x i64> %a to <2 x i32>
+  %mmx_var.i = bitcast <2 x i32> %0 to x86_mmx
+  %1 = tail call x86_mmx @llvm.x86.mmx.psrli.d(x86_mmx %mmx_var.i, i32 0) nounwind
+  %2 = bitcast x86_mmx %1 to <2 x i32>
+  %3 = bitcast <2 x i32> %2 to <1 x i64>
+  %4 = extractelement <1 x i64> %3, i32 0
+  ret i64 %4
+}
+
 declare x86_mmx @llvm.x86.mmx.psrli.w(x86_mmx, i32) nounwind readnone
 
 define i64 @test69(<1 x i64> %a) nounwind readnone optsize ssp {
@@ -397,6 +423,19 @@ entry:
   ret i64 %4
 }
 
+define i64 @test66_2(<1 x i64> %a) nounwind readnone optsize ssp {
+; ALL-LABEL: @test66_2
+; ALL-NOT: psllw
+entry:
+  %0 = bitcast <1 x i64> %a to <4 x i16>
+  %mmx_var.i = bitcast <4 x i16> %0 to x86_mmx
+  %1 = tail call x86_mmx @llvm.x86.mmx.pslli.w(x86_mmx %mmx_var.i, i32 0) nounwind
+  %2 = bitcast x86_mmx %1 to <4 x i16>
+  %3 = bitcast <4 x i16> %2 to <1 x i64>
+  %4 = extractelement <1 x i64> %3, i32 0
+  ret i64 %4
+}
+
 declare x86_mmx @llvm.x86.mmx.psra.d(x86_mmx, x86_mmx) nounwind readnone
 
 define i64 @test65(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp {