[Thumb1] Any imm8 should have cost of 1

A simple MOVS rd, imm8 can materialize [-128, 127] in signed i8 type or
[0, 255] in unsigned i8 type on Thumb1.

Differential Revision: https://reviews.llvm.org/D52257

llvm-svn: 342898

diff --git a/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp b/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp
index 94f9cef..1b0d162 100644 (file)
--- a/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp
+++ b/llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp
@@ -77,8 +77,8 @@ int ARMTTIImpl::getIntImmCost(const APInt &Imm, Type *Ty) {
       return 1;
     return ST->hasV6T2Ops() ? 2 : 3;
   }
-  // Thumb1.
-  if (SImmVal >= 0 && SImmVal < 256)
+  // Thumb1, any i8 imm cost 1.
+  if (Bits == 8 || (SImmVal >= 0 && SImmVal < 256))
     return 1;
   if ((~SImmVal < 256) || ARM_AM::isThumbImmShiftedVal(ZImmVal))
     return 2;

diff --git a/llvm/test/CodeGen/Thumb/consthoist-imm8-costs-1.ll b/llvm/test/CodeGen/Thumb/consthoist-imm8-costs-1.ll
new file mode 100644 (file)
index 0000000..33c2e30
--- /dev/null
+++ b/llvm/test/CodeGen/Thumb/consthoist-imm8-costs-1.ll
@@ -0,0 +1,39 @@
+; RUN: llc %s -o - | FileCheck %s
+
+target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
+target triple = "thumbv6m-none-unknown-musleabi"
+
+@a = global i8 undef, align 4
+
+; Check that store-merging generates a single str i32 rather than strb+strb+strh,
+; i.e., -1 is not moved by constant-hoisting.
+; CHECK: movs [[R1:r[0-9]+]], #255
+; CHECK: lsls [[R2:r[0-9]+]], [[R1]], #16
+; CHECK: str  [[R2]]
+; CHECK: movs [[R3:r[0-9]+]], #255
+; CHECK: lsls [[R4:r[0-9]+]], [[R3]], #16
+; CHECK: str  [[R4]]
+; CHECK-NOT: strh
+; CHECK-NOT: strb
+
+define void @ham() {
+bb:
+  br i1 undef, label %bb1, label %bb2
+
+bb1:
+  store i8 0, i8* getelementptr inbounds (i8, i8* @a, i32 1), align 1
+  store i8 0, i8* getelementptr inbounds (i8, i8* @a, i32 0), align 4
+  store i8 -1, i8* getelementptr inbounds (i8, i8* @a, i32 2), align 2
+  store i8 0, i8* getelementptr inbounds (i8, i8* @a, i32 3), align 1
+  br label %bb3
+
+bb2:
+  store i8 0, i8* getelementptr inbounds (i8, i8* @a, i32 9), align 1
+  store i8 0, i8* getelementptr inbounds (i8, i8* @a, i32 8), align 4
+  store i8 -1, i8* getelementptr inbounds (i8, i8* @a, i32 10), align 2
+  store i8 0, i8* getelementptr inbounds (i8, i8* @a, i32 11), align 1
+  br label %bb3
+
+bb3:
+  ret void
+}