77 lines · plain
1; RUN: llc -mtriple=hexagon -mcpu=hexagonv5 < %s | FileCheck %s2;3; Remove the unnecessary 'add' instruction used for the hardware loop setup.4 5; CHECK: [[OP0:r[0-9]+]] = add([[OP1:r[0-9]+]],#-[[OP2:[0-9]+]]6; CHECK-NOT: add([[OP0]],#[[OP2]])7; CHECK: lsr([[OP1]],#{{[0-9]+}})8; CHECK: loop09 10define void @matrix_mul_matrix(i32 %N, ptr nocapture %C, ptr nocapture readnone %A, ptr nocapture readnone %B) #0 {11entry:12 %cmp4 = icmp eq i32 %N, 013 br i1 %cmp4, label %for.end, label %for.body.preheader14 15for.body.preheader:16 %maxval = add i32 %N, -717 %0 = icmp sgt i32 %maxval, 018 br i1 %0, label %for.body.preheader9, label %for.body.ur.preheader19 20for.body.preheader9:21 br label %for.body22 23for.body:24 %arrayidx.phi = phi ptr [ %arrayidx.inc.7, %for.body ], [ %C, %for.body.preheader9 ]25 %i.05 = phi i32 [ %inc.7, %for.body ], [ 0, %for.body.preheader9 ]26 store i32 %i.05, ptr %arrayidx.phi, align 427 %inc = add i32 %i.05, 128 %arrayidx.inc = getelementptr i32, ptr %arrayidx.phi, i32 129 store i32 %inc, ptr %arrayidx.inc, align 430 %inc.1 = add i32 %i.05, 231 %arrayidx.inc.1 = getelementptr i32, ptr %arrayidx.phi, i32 232 store i32 %inc.1, ptr %arrayidx.inc.1, align 433 %inc.2 = add i32 %i.05, 334 %arrayidx.inc.2 = getelementptr i32, ptr %arrayidx.phi, i32 335 store i32 %inc.2, ptr %arrayidx.inc.2, align 436 %inc.3 = add i32 %i.05, 437 %arrayidx.inc.3 = getelementptr i32, ptr %arrayidx.phi, i32 438 store i32 %inc.3, ptr %arrayidx.inc.3, align 439 %inc.4 = add i32 %i.05, 540 %arrayidx.inc.4 = getelementptr i32, ptr %arrayidx.phi, i32 541 store i32 %inc.4, ptr %arrayidx.inc.4, align 442 %inc.5 = add i32 %i.05, 643 %arrayidx.inc.5 = getelementptr i32, ptr %arrayidx.phi, i32 644 store i32 %inc.5, ptr %arrayidx.inc.5, align 445 %inc.6 = add i32 %i.05, 746 %arrayidx.inc.6 = getelementptr i32, ptr %arrayidx.phi, i32 747 store i32 %inc.6, ptr %arrayidx.inc.6, align 448 %inc.7 = add i32 %i.05, 849 %exitcond.7 = icmp slt i32 %inc.7, %maxval50 %arrayidx.inc.7 = getelementptr i32, ptr %arrayidx.phi, i32 851 br i1 %exitcond.7, label %for.body, label %for.end.loopexit.ur-lcssa52 53for.end.loopexit.ur-lcssa:54 %1 = icmp eq i32 %inc.7, %N55 br i1 %1, label %for.end, label %for.body.ur.preheader56 57for.body.ur.preheader:58 %arrayidx.phi.ur.ph = phi ptr [ %C, %for.body.preheader ], [ %arrayidx.inc.7, %for.end.loopexit.ur-lcssa ]59 %i.05.ur.ph = phi i32 [ 0, %for.body.preheader ], [ %inc.7, %for.end.loopexit.ur-lcssa ]60 br label %for.body.ur61 62for.body.ur:63 %arrayidx.phi.ur = phi ptr [ %arrayidx.inc.ur, %for.body.ur ], [ %arrayidx.phi.ur.ph, %for.body.ur.preheader ]64 %i.05.ur = phi i32 [ %inc.ur, %for.body.ur ], [ %i.05.ur.ph, %for.body.ur.preheader ]65 store i32 %i.05.ur, ptr %arrayidx.phi.ur, align 466 %inc.ur = add i32 %i.05.ur, 167 %exitcond.ur = icmp eq i32 %inc.ur, %N68 %arrayidx.inc.ur = getelementptr i32, ptr %arrayidx.phi.ur, i32 169 br i1 %exitcond.ur, label %for.end.loopexit, label %for.body.ur70 71for.end.loopexit:72 br label %for.end73 74for.end:75 ret void76}77