[ARM] Don't shrink STM if it would cause an unknown base register store

[john.brawn]

Summary

If a 16-bit thumb STM with writeback stores the base register but it isn't the first register in the list, then an unknown value is stored. The load/store optimizer knows this and generates a 32-bit STM without writeback instead, but thumb2 size reduction converts it into a 16-bit STM. Fix this by having thumb2 size reduction notice such STMs and leave them as they are.

[efriedma]

llvm/lib/Target/ARM/Thumb2SizeReduction.cpp

525

It's probably worth reorganizing the comments here to clarify: tSTMIA_UPD the only 16-bit STM. So the Thumb2 instruction is fine, but the rewritten one would be unpredictable.

[john.brawn]

Adjust comments. Also, after double-checking the 8-A and 8-M ARMARMs it's OK if the base register being stored is the lowest one so adjusted to allow that.

[efriedma]

LGTM

The ARMv7 ARM claims the form with writeback of the lowest register is deprecated, but it looks like ARMv8 un-deprecated it, so we probably don't care about that.

llvm/lib/Target/ARM/Thumb2SizeReduction.cpp

case ARM::t2LDMIA: {
 
     if (!isOK)
       return false;
 
     OpNum = 0;
     isLdStMul = true;
     break;
   }
-  case ARM::t2STMIA:
+  case ARM::t2STMIA: {
     // If the base register is killed, we don't care what its value is after the
     // instruction, so we can use an updating STMIA.
     if (!MI->getOperand(0).isKill())
       return false;
 
+    // STM with writeback that stores the base register is unpredictable.
+    Register BaseReg = MI->getOperand(0).getReg();
+    for (unsigned i = 3; i < MI->getNumOperands(); ++i)
+      if (MI->getOperand(i).getReg() == BaseReg)
+        return false;
+
     break;
+  }
   case ARM::t2LDMIA_RET: {
     Register BaseReg = MI->getOperand(1).getReg();
     if (BaseReg != ARM::SP)
       return false;
     Opc = Entry.NarrowOpc2; // tPOP_RET
     OpNum = 2;
     isLdStMul = true;
     break;

llvm/test/CodeGen/Thumb/stm-deprecated.ll

-; RUN: llc -mtriple=thumbv6m-eabi -verify-machineinstrs %s -o - | FileCheck %s
-; RUN: llc -mtriple=thumbv5e-linux-gnueabi -verify-machineinstrs %s -o - | FileCheck %s
+; RUN: llc -mtriple=thumbv6m-eabi -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-T1
+; RUN: llc -mtriple=thumbv5e-linux-gnueabi -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-T1
+; RUN: llc -mtriple=thumbv7m -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-T2
+; RUN: llc -mtriple=thumbv7a -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-T2
 
-%0 = type { %0*, %0*, i32 }
+%struct1 = type { %struct1*, %struct1*, i32 }
+%struct2 = type { i32, i32, i32* }
 
-@x1 = external global %0, align 4
-@x2 = external global %0, align 4
+@x1 = external global %struct1, align 4
+@x2 = external global %struct1, align 4
 
-; CHECK: str r0, [r1]
-; CHECK-NEXT: str r1, [r1, #4]
+declare void @fn1(i32, i32)
+declare void @fn2(%struct2*)
+
+; CHECK-LABEL: test1:
+; CHECK-T1: str r0, [r1]
+; CHECK-T1-NEXT: str r1, [r1, #4]
+; CHECK-T2: strd r0, r1, [r1]
 ; CHECK-NOT: stm
+define void @test1(i32 %unused, %struct1* %x) {
+  %first = getelementptr inbounds %struct1, %struct1* %x, i32 0, i32 0
+  %second = getelementptr inbounds %struct1, %struct1* %x, i32 0, i32 1
+  store %struct1* @x1, %struct1** %first
+  store %struct1* %x, %struct1** %second
+  ret void
+}
+
+; CHECK-LABEL: test2:
+; CHECK-T1: str r0, [r2]
+; CHECK-T1-NEXT: str r1, [r2, #4]
+; CHECK-T1-NEXT: str r2, [r2, #8]
+; CHECK-T2: stm.w r2, {r0, r1, r2}
+; CHECK-NOT: stm r2!, {r0, r1, r2}
+define i32 @test2(i32 %a, i32 %b, %struct2* %p) {
+entry:
+  %p1 = getelementptr inbounds %struct2, %struct2* %p, i32 0, i32 0
+  %p2 = getelementptr inbounds %struct2, %struct2* %p, i32 0, i32 1
+  %p3 = getelementptr inbounds %struct2, %struct2* %p, i32 0, i32 2
+  store i32 %a, i32* %p1, align 4
+  store i32 %b, i32* %p2, align 4
+  store i32* %p1, i32** %p3, align 4
+  call void @fn1(i32 %a, i32 %b)
+  ret i32 0
+}
 
-define void @foo(i32 %unused, %0* %x) {
-  %first = getelementptr inbounds %0, %0* %x, i32 0, i32 0
-  %second = getelementptr inbounds %0, %0* %x, i32 0, i32 1
-  store %0* @x1, %0** %first
-  store %0* %x, %0** %second
-  unreachable
+; CHECK-LABEL: test3:
+; CHECK-T1: str r0, [r2]
+; CHECK-T1-NEXT: str r1, [r2, #4]
+; CHECK-T1-NEXT: str r2, [r2, #8]
+; CHECK-T2: stm.w r2, {r0, r1, r2}
+; CHECK-NOT: stm r2!, {r0, r1, r2}
+define i32 @test3(i32 %a, i32 %b, %struct2* %p) {
+entry:
+  %p1 = getelementptr inbounds %struct2, %struct2* %p, i32 0, i32 0
+  %p2 = getelementptr inbounds %struct2, %struct2* %p, i32 0, i32 1
+  %p3 = getelementptr inbounds %struct2, %struct2* %p, i32 0, i32 2
+  store i32 %a, i32* %p1, align 4
+  store i32 %b, i32* %p2, align 4
+  store i32* %p1, i32** %p3, align 4
+  %p4 = getelementptr inbounds %struct2, %struct2* %p, i32 1
+  call void @fn2(%struct2* %p4)
+  ret i32 0
 }