<div dir="ltr"><br><div class="gmail_extra"><br><div class="gmail_quote">On Tue, Jun 2, 2015 at 11:31 PM, Tobias Grosser <span dir="ltr"><<a href="mailto:tobias@grosser.es" target="_blank">tobias@grosser.es</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Author: grosser<br>
Date: Wed Jun 3 01:31:30 2015<br>
New Revision: 238905<br>
<br>
URL: <a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__llvm.org_viewvc_llvm-2Dproject-3Frev-3D238905-26view-3Drev&d=AwMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=mQ4LZ2PUj9hpadE3cDHZnIdEwhEBrbAstXeMaFoB9tg&m=8l8_2G9-tM9fri5rMAdckHNGDJY2ztWGp7kzol1Ox08&s=Db-TWpE0o7nGewaRzLuw_8NEzBnDEm5eatG_uw4Ni0M&e=" target="_blank">http://llvm.org/viewvc/llvm-project?rev=238905&view=rev</a><br>
Log:<br>
Translate power-of-two floor-division into ashr<br>
<br>
Power-of-two floor divisions can be translated into an arithmetic shift<br>
operation. This allows us to replace a complex lowering that requires division<br>
operations:<br></blockquote><div><br></div><div>This only holds if the numerator is non-negative. -1/INT_MIN should give 0 but calculating it with a right-shift would give -1.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
%pexp.fdiv_q.0 = sub i64 %21, 128<br>
%pexp.fdiv_q.1 = add i64 %pexp.fdiv_q.0, 1<br>
%pexp.fdiv_q.2 = icmp slt i64 %21, 0<br>
%pexp.fdiv_q.3 = select i1 %pexp.fdiv_q.2, i64 %pexp.fdiv_q.1, i64 %21<br>
%pexp.fdiv_q.4 = sdiv i64 %pexp.fdiv_q.3, 128<br>
<br>
with a simple ashr:<br>
<br>
%polly.fdiv_q.shr = ashr i64 %21, 7<br>
<br>
Added:<br>
polly/trunk/test/Isl/CodeGen/exprModDiv___%for.cond---%for.end.jscop.pow2<br>
- copied, changed from r238645, polly/trunk/test/Isl/CodeGen/exprModDiv___%for.cond---%for.end.jscop<br>
Modified:<br>
polly/trunk/lib/CodeGen/IslExprBuilder.cpp<br>
polly/trunk/test/Isl/CodeGen/exprModDiv.ll<br>
polly/trunk/test/Isl/CodeGen/exprModDiv___%for.cond---%for.end.jscop<br>
<br>
Modified: polly/trunk/lib/CodeGen/IslExprBuilder.cpp<br>
URL: <a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__llvm.org_viewvc_llvm-2Dproject_polly_trunk_lib_CodeGen_IslExprBuilder.cpp-3Frev-3D238905-26r1-3D238904-26r2-3D238905-26view-3Ddiff&d=AwMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=mQ4LZ2PUj9hpadE3cDHZnIdEwhEBrbAstXeMaFoB9tg&m=8l8_2G9-tM9fri5rMAdckHNGDJY2ztWGp7kzol1Ox08&s=-XaiIb1coppUCe2Dhl0nw94WmE7hTjSBon3frtebApI&e=" target="_blank">http://llvm.org/viewvc/llvm-project/polly/trunk/lib/CodeGen/IslExprBuilder.cpp?rev=238905&r1=238904&r2=238905&view=diff</a><br>
==============================================================================<br>
--- polly/trunk/lib/CodeGen/IslExprBuilder.cpp (original)<br>
+++ polly/trunk/lib/CodeGen/IslExprBuilder.cpp Wed Jun 3 01:31:30 2015<br>
@@ -301,6 +301,11 @@ Value *IslExprBuilder::createOpBin(__isl<br>
Res = Builder.CreateUDiv(LHS, RHS, "pexp.p_div_q");<br>
break;<br>
case isl_ast_op_fdiv_q: { // Round towards -infty<br>
+ auto &Int = dyn_cast<ConstantInt>(RHS)->getValue();<br></blockquote><div><br></div><div>The dyn_cast seems superfluous if you immediately dereference it. If the dyn_cast cannot fail, please use cast.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
+ if (Int.isPowerOf2()) {<br>
+ Res = Builder.CreateAShr(LHS, Int.ceilLogBase2(), "polly.fdiv_q.shr");<br>
+ break;<br>
+ }<br>
// TODO: Review code and check that this calculation does not yield<br>
// incorrect overflow in some bordercases.<br>
//<br>
<br>
Modified: polly/trunk/test/Isl/CodeGen/exprModDiv.ll<br>
URL: <a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__llvm.org_viewvc_llvm-2Dproject_polly_trunk_test_Isl_CodeGen_exprModDiv.ll-3Frev-3D238905-26r1-3D238904-26r2-3D238905-26view-3Ddiff&d=AwMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=mQ4LZ2PUj9hpadE3cDHZnIdEwhEBrbAstXeMaFoB9tg&m=8l8_2G9-tM9fri5rMAdckHNGDJY2ztWGp7kzol1Ox08&s=wWGJt73555XQdjnnrzk_CJHL6zm7Lt2bk4Nb5PF5QO8&e=" target="_blank">http://llvm.org/viewvc/llvm-project/polly/trunk/test/Isl/CodeGen/exprModDiv.ll?rev=238905&r1=238904&r2=238905&view=diff</a><br>
==============================================================================<br>
--- polly/trunk/test/Isl/CodeGen/exprModDiv.ll (original)<br>
+++ polly/trunk/test/Isl/CodeGen/exprModDiv.ll Wed Jun 3 01:31:30 2015<br>
@@ -1,4 +1,5 @@<br>
; RUN: opt %loadPolly -polly-import-jscop -polly-import-jscop-dir=%S -polly-codegen -S < %s | FileCheck %s<br>
+; RUN: opt %loadPolly -polly-import-jscop -polly-import-jscop-dir=%S -polly-codegen -polly-import-jscop-postfix=pow2 -S < %s | FileCheck %s -check-prefix=POW2<br>
;<br>
; void exprModDiv(float *A, float *B, float *C, long N, long p) {<br>
; for (long i = 0; i < N; i++)<br>
@@ -12,34 +13,58 @@<br>
; useful as LLVM will translate urem and udiv operations with power-of-two<br>
; denominators to fast bitwise and or shift operations.<br>
<br>
-; A[i % 128]<br>
-; CHECK: %pexp.pdiv_r = urem i64 %polly.indvar, 128<br>
+; A[i % 127]<br>
+; CHECK: %pexp.pdiv_r = urem i64 %polly.indvar, 127<br>
; CHECK: %polly.access.A6 = getelementptr float, float* %A, i64 %pexp.pdiv_r<br>
<br>
-; A[i / 128]<br>
-; CHECK: %pexp.div = sdiv i64 %polly.indvar, 128<br>
+; A[i / 127]<br>
+; CHECK: %pexp.div = sdiv i64 %polly.indvar, 127<br>
; CHECK: %polly.access.B8 = getelementptr float, float* %B, i64 %pexp.div<br>
;<br>
; FIXME: Make isl mark this as an udiv expression.<br>
<br>
; #define floord(n,d) ((n < 0) ? (n - d + 1) : n) / d<br>
-; A[p + 128 * floord(-p - 1, 128) + 128]<br>
+; A[p + 127 * floord(-p - 1, 127) + 127]<br>
; CHECK: %20 = sub nsw i64 0, %p<br>
; CHECK: %21 = sub nsw i64 %20, 1<br>
-; CHECK: %pexp.fdiv_q.0 = sub i64 %21, 128<br>
+; CHECK: %pexp.fdiv_q.0 = sub i64 %21, 127<br>
; CHECK: %pexp.fdiv_q.1 = add i64 %pexp.fdiv_q.0, 1<br>
; CHECK: %pexp.fdiv_q.2 = icmp slt i64 %21, 0<br>
; CHECK: %pexp.fdiv_q.3 = select i1 %pexp.fdiv_q.2, i64 %pexp.fdiv_q.1, i64 %21<br>
-; CHECK: %pexp.fdiv_q.4 = sdiv i64 %pexp.fdiv_q.3, 128<br>
-; CHECK: %22 = mul nsw i64 128, %pexp.fdiv_q.4<br>
+; CHECK: %pexp.fdiv_q.4 = sdiv i64 %pexp.fdiv_q.3, 127<br>
+; CHECK: %22 = mul nsw i64 127, %pexp.fdiv_q.4<br>
; CHECK: %23 = add nsw i64 %p, %22<br>
-; CHECK: %24 = add nsw i64 %23, 128<br>
+; CHECK: %24 = add nsw i64 %23, 127<br>
; CHECK: %polly.access.A10 = getelementptr float, float* %A, i64 %24<br>
<br>
-; A[p / 128]<br>
-; CHECK: %pexp.div12 = sdiv i64 %p, 128<br>
+; A[p / 127]<br>
+; CHECK: %pexp.div12 = sdiv i64 %p, 127<br>
; CHECK: %polly.access.B13 = getelementptr float, float* %B, i64 %pexp.div12<br>
<br>
+; A[i % 128]<br>
+; POW2: %pexp.pdiv_r = urem i64 %polly.indvar, 128<br>
+; POW2: %polly.access.A6 = getelementptr float, float* %A, i64 %pexp.pdiv_r<br>
+<br>
+; A[i / 128]<br>
+; POW2: %pexp.div = sdiv i64 %polly.indvar, 128<br>
+; POW2: %polly.access.B8 = getelementptr float, float* %B, i64 %pexp.div<br>
+;<br>
+; FIXME: Make isl mark this as an udiv expression.<br>
+<br>
+; #define floord(n,d) ((n < 0) ? (n - d + 1) : n) / d<br>
+; A[p + 128 * floord(-p - 1, 128) + 128]<br>
+; POW2: %20 = sub nsw i64 0, %p<br>
+; POW2: %21 = sub nsw i64 %20, 1<br>
+; POW2: %polly.fdiv_q.shr = ashr i64 %21, 7<br>
+; POW2: %22 = mul nsw i64 128, %polly.fdiv_q.shr<br>
+; POW2: %23 = add nsw i64 %p, %22<br>
+; POW2: %24 = add nsw i64 %23, 128<br>
+; POW2: %polly.access.A10 = getelementptr float, float* %A, i64 %24<br>
+<br>
+; A[p / 128]<br>
+; POW2: %pexp.div12 = sdiv i64 %p, 128<br>
+; POW2: %polly.access.B13 = getelementptr float, float* %B, i64 %pexp.div12<br>
+<br>
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"<br>
<br>
define void @exprModDiv(float* %A, float* %B, float* %C, i64 %N, i64 %p) {<br>
<br>
Modified: polly/trunk/test/Isl/CodeGen/exprModDiv___%for.cond---%for.end.jscop<br>
URL: <a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__llvm.org_viewvc_llvm-2Dproject_polly_trunk_test_Isl_CodeGen_exprModDiv-5F-5F-5F-2525for.cond-2D-2D-2D-2525for.end.jscop-3Frev-3D238905-26r1-3D238904-26r2-3D238905-26view-3Ddiff&d=AwMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=mQ4LZ2PUj9hpadE3cDHZnIdEwhEBrbAstXeMaFoB9tg&m=8l8_2G9-tM9fri5rMAdckHNGDJY2ztWGp7kzol1Ox08&s=9UQ80N6-BOel0K5oafNX0APyL64d60rMOgxmlOdAkPk&e=" target="_blank">http://llvm.org/viewvc/llvm-project/polly/trunk/test/Isl/CodeGen/exprModDiv___%25for.cond---%25for.end.jscop?rev=238905&r1=238904&r2=238905&view=diff</a><br>
==============================================================================<br>
--- polly/trunk/test/Isl/CodeGen/exprModDiv___%for.cond---%for.end.jscop (original)<br>
+++ polly/trunk/test/Isl/CodeGen/exprModDiv___%for.cond---%for.end.jscop Wed Jun 3 01:31:30 2015<br>
@@ -6,19 +6,19 @@<br>
"accesses" : [<br>
{<br>
"kind" : "read",<br>
- "relation" : "[N, p] -> { Stmt_for_body[i0] -> MemRef_A[i0 % 128] }"<br>
+ "relation" : "[N, p] -> { Stmt_for_body[i0] -> MemRef_A[i0 % 127] }"<br>
},<br>
{<br>
"kind" : "read",<br>
- "relation" : "[N, p] -> { Stmt_for_body[i0] -> MemRef_B[i0 / 128] }"<br>
+ "relation" : "[N, p] -> { Stmt_for_body[i0] -> MemRef_B[i0 / 127] }"<br>
},<br>
{<br>
"kind" : "read",<br>
- "relation" : "[N, p] -> { Stmt_for_body[i0] -> MemRef_A[p % 128] }"<br>
+ "relation" : "[N, p] -> { Stmt_for_body[i0] -> MemRef_A[p % 127] }"<br>
},<br>
{<br>
"kind" : "read",<br>
- "relation" : "[N, p] -> { Stmt_for_body[i0] -> MemRef_B[p / 128] }"<br>
+ "relation" : "[N, p] -> { Stmt_for_body[i0] -> MemRef_B[p / 127] }"<br>
},<br>
{<br>
"kind" : "read",<br>
<br>
Copied: polly/trunk/test/Isl/CodeGen/exprModDiv___%for.cond---%for.end.jscop.pow2 (from r238645, polly/trunk/test/Isl/CodeGen/exprModDiv___%for.cond---%for.end.jscop)<br>
URL: <a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__llvm.org_viewvc_llvm-2Dproject_polly_trunk_test_Isl_CodeGen_exprModDiv-5F-5F-5F-2525for.cond-2D-2D-2D-2525for.end.jscop.pow2-3Fp2-3Dpolly_trunk_test_Isl_CodeGen_exprModDiv-5F-5F-5F-2525for.cond-2D-2D-2D-2525for.end.jscop.pow2-26p1-3Dpolly_trunk_test_Isl_CodeGen_exprModDiv-5F-5F-5F-2525for.cond-2D-2D-2D-2525for.end.jscop-26r1-3D238645-26r2-3D238905-26rev-3D238905-26view-3Ddiff&d=AwMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=mQ4LZ2PUj9hpadE3cDHZnIdEwhEBrbAstXeMaFoB9tg&m=8l8_2G9-tM9fri5rMAdckHNGDJY2ztWGp7kzol1Ox08&s=caGQn8BE070H_ltPLAF0PZaNy2a-bt--37SR6idq5L4&e=" target="_blank">http://llvm.org/viewvc/llvm-project/polly/trunk/test/Isl/CodeGen/exprModDiv___%25for.cond---%25for.end.jscop.pow2?p2=polly/trunk/test/Isl/CodeGen/exprModDiv___%25for.cond---%25for.end.jscop.pow2&p1=polly/trunk/test/Isl/CodeGen/exprModDiv___%25for.cond---%25for.end.jscop&r1=238645&r2=238905&rev=238905&view=diff</a><br>
==============================================================================<br>
(empty)<br>
<br>
<br>
_______________________________________________<br>
llvm-commits mailing list<br>
<a href="mailto:llvm-commits@cs.uiuc.edu">llvm-commits@cs.uiuc.edu</a><br>
<a href="http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits" target="_blank">http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits</a><br>
</blockquote></div><br></div></div>