[ARM] Do not emit ldrexd/strexd on Cortex-M chips

The ldrexd/strexd instructions are not supported on M-class chips, see
for example
https://developer.arm.com/documentation/dui0489/e/arm-and-thumb-instructions/memory-access-instructions/ldrex-and-strex
which says:

> All these 32-bit Thumb instructions are available in ARMv6T2 and
> above, except that LDREXD and STREXD are not available in the ARMv7-M
> architecture.

Looking at the ARMv8-M architecture, it appears that these instructions
aren't supported either. The Architecture Reference Manual lists
ldrex/strex but not ldrexd/strexd:
https://developer.arm.com/documentation/ddi0553/bn/

Godbolt example on LLVM 11.0.0, which incorrectly emits ldrexd/strexd
instructions: https://llvm.godbolt.org/z/5qqPnE

Differential Revision: https://reviews.llvm.org/D95891

Author: aykevl

llvm/lib/Target/ARM/ARMISelLowering.cpp

@@ -18752,16 +18752,20 @@ ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
              : AtomicExpansionKind::None;
 }
 
+// Similar to shouldExpandAtomicRMWInIR, ldrex/strex can be used  up to 32
+// bits, and up to 64 bits on the non-M profiles.
 TargetLowering::AtomicExpansionKind
 ARMTargetLowering::shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const {
   // At -O0, fast-regalloc cannot cope with the live vregs necessary to
   // implement cmpxchg without spilling. If the address being exchanged is also
   // on the stack and close enough to the spill slot, this can lead to a
   // situation where the monitor always gets cleared and the atomic operation
   // can never succeed. So at -O0 we need a late-expanded pseudo-inst instead.
+  unsigned Size = AI->getOperand(1)->getType()->getPrimitiveSizeInBits();
   bool HasAtomicCmpXchg =
       !Subtarget->isThumb() || Subtarget->hasV8MBaselineOps();
-  if (getTargetMachine().getOptLevel() != 0 && HasAtomicCmpXchg)
+  if (getTargetMachine().getOptLevel() != 0 && HasAtomicCmpXchg &&
+      Size <= (Subtarget->isMClass() ? 32U : 64U))
     return AtomicExpansionKind::LLSC;
   return AtomicExpansionKind::None;
 }

llvm/test/CodeGen/ARM/atomic-64bit.ll

@@ -2,6 +2,8 @@
 ; RUN: llc < %s -mtriple=thumbv7-none-linux-gnueabihf | FileCheck %s --check-prefix=CHECK-THUMB --check-prefix=CHECK-THUMB-LE
 ; RUN: llc < %s -mtriple=armebv7 -target-abi apcs | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-BE
 ; RUN: llc < %s -mtriple=thumbebv7-none-linux-gnueabihf | FileCheck %s --check-prefix=CHECK-THUMB --check-prefix=CHECK-THUMB-BE
+; RUN: llc < %s -mtriple=armv7m--none-eabi | FileCheck %s --check-prefix=CHECK-M
+; RUN: llc < %s -mtriple=armv8m--none-eabi | FileCheck %s --check-prefix=CHECK-M
 
 define i64 @test1(i64* %ptr, i64 %val) {
 ; CHECK-LABEL: test1:
@@ -28,6 +30,8 @@ define i64 @test1(i64* %ptr, i64 %val) {
 ; CHECK-THUMB: bne
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_fetch_and_add_8
+
   %r = atomicrmw add i64* %ptr, i64 %val seq_cst
   ret i64 %r
 }
@@ -57,6 +61,8 @@ define i64 @test2(i64* %ptr, i64 %val) {
 ; CHECK-THUMB: bne
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_fetch_and_sub_8
+
   %r = atomicrmw sub i64* %ptr, i64 %val seq_cst
   ret i64 %r
 }
@@ -86,6 +92,8 @@ define i64 @test3(i64* %ptr, i64 %val) {
 ; CHECK-THUMB: bne
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_fetch_and_and_8
+
   %r = atomicrmw and i64* %ptr, i64 %val seq_cst
   ret i64 %r
 }
@@ -115,6 +123,8 @@ define i64 @test4(i64* %ptr, i64 %val) {
 ; CHECK-THUMB: bne
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_fetch_and_or_8
+
   %r = atomicrmw or i64* %ptr, i64 %val seq_cst
   ret i64 %r
 }
@@ -144,6 +154,8 @@ define i64 @test5(i64* %ptr, i64 %val) {
 ; CHECK-THUMB: bne
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_fetch_and_xor_8
+
   %r = atomicrmw xor i64* %ptr, i64 %val seq_cst
   ret i64 %r
 }
@@ -165,6 +177,8 @@ define i64 @test6(i64* %ptr, i64 %val) {
 ; CHECK-THUMB: bne
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_lock_test_and_set_8
+
   %r = atomicrmw xchg i64* %ptr, i64 %val seq_cst
   ret i64 %r
 }
@@ -199,12 +213,15 @@ define i64 @test7(i64* %ptr, i64 %val1, i64 %val2) {
 ; CHECK-THUMB: beq
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_val_compare_and_swap_8
+
   %pair = cmpxchg i64* %ptr, i64 %val1, i64 %val2 seq_cst seq_cst
   %r = extractvalue { i64, i1 } %pair, 0
   ret i64 %r
 }
 
-; Compiles down to a single ldrexd
+; Compiles down to a single ldrexd, except on M class devices where ldrexd
+; isn't supported.
 define i64 @test8(i64* %ptr) {
 ; CHECK-LABEL: test8:
 ; CHECK: ldrexd [[REG1:(r[0-9]?[02468])]], [[REG2:(r[0-9]?[13579])]]
@@ -220,12 +237,15 @@ define i64 @test8(i64* %ptr) {
 ; CHECK-THUMB-NOT: strexd
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_val_compare_and_swap_8
+
   %r = load atomic i64, i64* %ptr seq_cst, align 8
   ret i64 %r
 }
 
 ; Compiles down to atomicrmw xchg; there really isn't any more efficient
-; way to write it.
+; way to write it. Except on M class devices, where ldrexd/strexd aren't
+; supported.
 define void @test9(i64* %ptr, i64 %val) {
 ; CHECK-LABEL: test9:
 ; CHECK: dmb {{ish$}}
@@ -243,6 +263,8 @@ define void @test9(i64* %ptr, i64 %val) {
 ; CHECK-THUMB: bne
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_lock_test_and_set_8
+
   store atomic i64 %val, i64* %ptr seq_cst, align 8
   ret void
 }
@@ -286,6 +308,8 @@ define i64 @test10(i64* %ptr, i64 %val) {
 ; CHECK-THUMB: bne
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_fetch_and_min_8
+
   %r = atomicrmw min i64* %ptr, i64 %val seq_cst
   ret i64 %r
 }
@@ -329,6 +353,8 @@ define i64 @test11(i64* %ptr, i64 %val) {
 ; CHECK-THUMB: bne
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_fetch_and_umin_8
+
   %r = atomicrmw umin i64* %ptr, i64 %val seq_cst
   ret i64 %r
 }
@@ -372,6 +398,8 @@ define i64 @test12(i64* %ptr, i64 %val) {
 ; CHECK-THUMB: bne
 ; CHECK-THUMB: dmb {{ish$}}
 
+; CHECK-M: __sync_fetch_and_max_8
+
   %r = atomicrmw max i64* %ptr, i64 %val seq_cst
   ret i64 %r
 }
@@ -414,6 +442,9 @@ define i64 @test13(i64* %ptr, i64 %val) {
 ; CHECK-THUMB: cmp
 ; CHECK-THUMB: bne
 ; CHECK-THUMB: dmb {{ish$}}
+
+; CHECK-M: __sync_fetch_and_umax_8
+
   %r = atomicrmw umax i64* %ptr, i64 %val seq_cst
   ret i64 %r
 }