[ARM] Extract shifts out of multiply-by-constant

Turning (op x (mul y k)) into (op x (lsl (mul y k>>n) n)) is beneficial when
we can do the lsl as a shifted operand and the resulting multiply constant is
simpler to generate.

Do this by doing the transformation when trying to select a shifted operand,
as that ensures that it actually turns out better (the alternative would be to
do it in PreprocessISelDAG, but we don't know for sure there if extracting the
shift would allow a shifted operand to be used).

Differential Revision: http://reviews.llvm.org/D12196

llvm-svn: 247569

1 files changed, 182 insertions, 31 deletions

diff --git a/llvm/test/CodeGen/ARM/shifter_operand.ll b/llvm/test/CodeGen/ARM/shifter_operand.ll
index 6f5c0e8279a..11a8ee6ef7d 100644
--- a/llvm/test/CodeGen/ARM/shifter_operand.ll
+++ b/llvm/test/CodeGen/ARM/shifter_operand.ll
@@ -1,14 +1,14 @@
-; RUN: llc < %s -mtriple=armv7-apple-darwin -mcpu=cortex-a8 | FileCheck %s -check-prefix=A8
-; RUN: llc < %s -mtriple=armv7-apple-darwin -mcpu=cortex-a9 | FileCheck %s -check-prefix=A9
+; RUN: llc < %s -mtriple=armv7-apple-darwin -mcpu=cortex-a8 | FileCheck %s -check-prefix=CHECK -check-prefix=CHECK-ARM
+; RUN: llc < %s -mtriple=armv7-apple-darwin -mcpu=cortex-a9 | FileCheck %s -check-prefix=CHECK -check-prefix=CHECK-ARM
+; RUN: llc < %s -mtriple=thumbv7m-none-eabi | FileCheck %s -check-prefix=CHECK -check-prefix=CHECK-THUMB
 ; rdar://8576755
 
 
 define i32 @test1(i32 %X, i32 %Y, i8 %sh) {
-; A8-LABEL: test1:
-; A8: add r0, r0, r1, lsl r2
-
-; A9-LABEL: test1:
-; A9: add r0, r0, r1, lsl r2
+; CHECK-LABEL: test1:
+; CHECK-ARM: add r0, r0, r1, lsl r2
+; CHECK-THUMB: lsls r1, r2
+; CHECK-THUMB: add r0, r1
         %shift.upgrd.1 = zext i8 %sh to i32
         %A = shl i32 %Y, %shift.upgrd.1
         %B = add i32 %X, %A
@@ -16,11 +16,10 @@ define i32 @test1(i32 %X, i32 %Y, i8 %sh) {
 }
 
 define i32 @test2(i32 %X, i32 %Y, i8 %sh) {
-; A8-LABEL: test2:
-; A8: bic r0, r0, r1, asr r2
-
-; A9-LABEL: test2:
-; A9: bic r0, r0, r1, asr r2
+; CHECK-LABEL: test2:
+; CHECK-ARM: bic r0, r0, r1, asr r2
+; CHECK-THUMB: asrs r1, r2
+; CHECK-THUMB: bics r0, r1
         %shift.upgrd.2 = zext i8 %sh to i32
         %A = ashr i32 %Y, %shift.upgrd.2
         %B = xor i32 %A, -1
@@ -30,14 +29,9 @@ define i32 @test2(i32 %X, i32 %Y, i8 %sh) {
 
 define i32 @test3(i32 %base, i32 %base2, i32 %offset) {
 entry:
-; A8-LABEL: test3:
-; A8: ldr r0, [r0, r2, lsl #2]
-; A8: ldr r1, [r1, r2, lsl #2]
-
-; lsl #2 is free
-; A9-LABEL: test3:
-; A9: ldr r0, [r0, r2, lsl #2]
-; A9: ldr r1, [r1, r2, lsl #2]
+; CHECK-LABEL: test3:
+; CHECK: ldr{{(.w)?}} r0, [r0, r2, lsl #2]
+; CHECK: ldr{{(.w)?}} r1, [r1, r2, lsl #2]
         %tmp1 = shl i32 %offset, 2
         %tmp2 = add i32 %base, %tmp1
         %tmp3 = inttoptr i32 %tmp2 to i32*
@@ -53,17 +47,11 @@ declare i8* @malloc(...)
 
 define fastcc void @test4(i16 %addr) nounwind {
 entry:
-; A8-LABEL: test4:
-; A8: ldr [[REG:r[0-9]+]], [r0, r1, lsl #2]
-; A8-NOT: ldr [[REG:r[0-9]+]], [r0, r1, lsl #2]!
-; A8: str [[REG]], [r0, r1, lsl #2]
-; A8-NOT: str [[REG]], [r0]
-
-; A9-LABEL: test4:
-; A9: ldr [[REG:r[0-9]+]], [r0, r1, lsl #2]
-; A9-NOT: ldr [[REG:r[0-9]+]], [r0, r1, lsl #2]!
-; A9: str [[REG]], [r0, r1, lsl #2]
-; A9-NOT: str [[REG]], [r0]
+; CHECK-LABEL: test4:
+; CHECK: ldr{{(.w)?}} [[REG:r[0-9]+]], [r0, r1, lsl #2]
+; CHECK-NOT: ldr{{(.w)?}} [[REG:r[0-9]+]], [r0, r1, lsl #2]!
+; CHECK: str{{(.w)?}} [[REG]], [r0, r1, lsl #2]
+; CHECK-NOT: str{{(.w)?}} [[REG]], [r0]
   %0 = tail call i8* (...) @malloc(i32 undef) nounwind
   %1 = bitcast i8* %0 to i32*
   %2 = sext i16 %addr to i32
@@ -73,3 +61,166 @@ entry:
   store i32 %5, i32* %3, align 4
   ret void
 }
+
+define i32 @test_orr_extract_from_mul_1(i32 %x, i32 %y) {
+entry:
+; CHECK-LABEL: test_orr_extract_from_mul_1
+; CHECK: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-ARM: orr r0, r1, r0
+; CHECK-THUMB: muls r1, r2, r1
+; CHECk-THUMB: orrs r0, r1
+  %mul = mul i32 %y, 63767
+  %or = or i32 %mul, %x
+  ret i32 %or
+}
+
+define i32 @test_orr_extract_from_mul_2(i32 %x, i32 %y) {
+; CHECK-LABEL: test_orr_extract_from_mul_2
+; CHECK: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-THUMB: muls r1, r2, r1
+; CHECK: orr{{(.w)?}} r0, r0, r1, lsl #1
+entry:
+  %mul1 = mul i32 %y, 127534
+  %or = or i32 %mul1, %x
+  ret i32 %or
+}
+
+define i32 @test_orr_extract_from_mul_3(i32 %x, i32 %y) {
+; CHECK-LABEL: test_orr_extract_from_mul_3
+; CHECK: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-THUMB: muls r1, r2, r1
+; CHECK: orr{{(.w)?}} r0, r0, r1, lsl #2
+entry:
+  %mul1 = mul i32 %y, 255068
+  %or = or i32 %mul1, %x
+  ret i32 %or
+}
+
+define i32 @test_orr_extract_from_mul_4(i32 %x, i32 %y) {
+; CHECK-LABEL: test_orr_extract_from_mul_4
+; CHECK: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-THUMB: muls r1, r2, r1
+; CHECK: orr{{(.w)?}} r0, r0, r1, lsl #3
+entry:
+  %mul1 = mul i32 %y, 510136
+  %or = or i32 %mul1, %x
+  ret i32 %or
+}
+
+define i32 @test_orr_extract_from_mul_5(i32 %x, i32 %y) {
+; CHECK-LABEL: test_orr_extract_from_mul_5
+; CHECK: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-THUMB: muls r1, r2, r1
+; CHECK: orr{{(.w)?}} r0, r0, r1, lsl #4
+entry:
+  %mul1 = mul i32 %y, 1020272
+  %or = or i32 %mul1, %x
+  ret i32 %or
+}
+
+define i32 @test_orr_extract_from_mul_6(i32 %x, i32 %y) {
+; CHECK-LABEL: test_orr_extract_from_mul_6
+; CHECK: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-THUMB: muls r1, r2, r1
+; CHECK: orr{{(.w)?}} r0, r0, r1, lsl #16
+entry:
+  %mul = mul i32 %y, -115933184
+  %or = or i32 %mul, %x
+  ret i32 %or
+}
+
+define i32 @test_load_extract_from_mul_1(i8* %x, i32 %y) {
+; CHECK-LABEL: test_load_extract_from_mul_1
+; CHECK: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-THUMB: muls r1, r2, r1
+; CHECK: ldrb r0, [r0, r1]
+entry:
+  %mul = mul i32 %y, 63767
+  %arrayidx = getelementptr inbounds i8, i8* %x, i32 %mul
+  %0 = load i8, i8* %arrayidx, align 1
+  %conv = zext i8 %0 to i32
+  ret i32 %conv
+}
+
+define i32 @test_load_extract_from_mul_2(i8* %x, i32 %y) {
+; CHECK-LABEL: test_load_extract_from_mul_2
+; CHECK: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-THUMB: muls r1, r2, r1
+; CHECK: ldrb{{(.w)?}} r0, [r0, r1, lsl #1]
+entry:
+  %mul1 = mul i32 %y, 127534
+  %arrayidx = getelementptr inbounds i8, i8* %x, i32 %mul1
+  %0 = load i8, i8* %arrayidx, align 1
+  %conv = zext i8 %0 to i32
+  ret i32 %conv
+}
+
+define i32 @test_load_extract_from_mul_3(i8* %x, i32 %y) {
+; CHECK-LABEL: test_load_extract_from_mul_3
+; CHECK: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-THUMB: muls r1, r2, r1
+; CHECK: ldrb{{(.w)?}} r0, [r0, r1, lsl #2]
+entry:
+  %mul1 = mul i32 %y, 255068
+  %arrayidx = getelementptr inbounds i8, i8* %x, i32 %mul1
+  %0 = load i8, i8* %arrayidx, align 1
+  %conv = zext i8 %0 to i32
+  ret i32 %conv
+}
+
+define i32 @test_load_extract_from_mul_4(i8* %x, i32 %y) {
+; CHECK-LABEL: test_load_extract_from_mul_4
+; CHECK: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-THUMB: muls r1, r2, r1
+; CHECK: ldrb{{(.w)?}} r0, [r0, r1, lsl #3]
+entry:
+  %mul1 = mul i32 %y, 510136
+  %arrayidx = getelementptr inbounds i8, i8* %x, i32 %mul1
+  %0 = load i8, i8* %arrayidx, align 1
+  %conv = zext i8 %0 to i32
+  ret i32 %conv
+}
+
+define i32 @test_load_extract_from_mul_5(i8* %x, i32 %y) {
+; CHECK-LABEL: test_load_extract_from_mul_5
+; CHECK-ARM: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-ARM: ldrb r0, [r0, r1, lsl #4]
+; CHECK-THUMB: movw r2, #37232
+; CHECK-THUMB: movt r2, #15
+; CHECK-THUMB: muls r1, r2, r1
+; CHECK-THUMB: ldrb r0, [r0, r1]
+entry:
+  %mul1 = mul i32 %y, 1020272
+  %arrayidx = getelementptr inbounds i8, i8* %x, i32 %mul1
+  %0 = load i8, i8* %arrayidx, align 1
+  %conv = zext i8 %0 to i32
+  ret i32 %conv
+}
+
+define i32 @test_load_extract_from_mul_6(i8* %x, i32 %y) {
+; CHECK-LABEL: test_load_extract_from_mul_6
+; CHECK-ARM: movw r2, #63767
+; CHECK-ARM: mul r1, r1, r2
+; CHECK-ARM: ldrb r0, [r0, r1, lsl #16]
+; CHECK-THUMB: movs r2, #0
+; CHECK-THUMB: movt r2, #63767
+; CHECK-THUMB: muls r1, r2, r1
+; CHECK-THUMB: ldrb r0, [r0, r1]
+entry:
+  %mul = mul i32 %y, -115933184
+  %arrayidx = getelementptr inbounds i8, i8* %x, i32 %mul
+  %0 = load i8, i8* %arrayidx, align 1
+  %conv = zext i8 %0 to i32
+  ret i32 %conv
+}