Subzero: Fix a mul lowering error.

tests_lit/llvm2ice_tests/multidef_kill.ll

+; This tests against a lowering error in a multiply instruction that produces
+; results in a low and high register.  This is usually lowered as a mul
+; instruction whose dest contains the low portion, and a FakeDef of the high
+; portion.  The problem is that if the high portion is unused (e.g. the multiply
+; is followed by a truncation), the FakeDef may be eliminated, and the register
+; allocator may assign the high register to a variable that is live across the
+; mul instruction.  This is incorrect because the mul instruction smashes the
+; register.
+
+; REQUIRES: allow_dump
+
+; RUN: %p2i --target x8632 -i %s --filetype=asm --args -O2 -asm-verbose \
+; RUN:   --reg-use=eax,edx -reg-reserve | FileCheck --check-prefix=X8632 %s
+; RUN: %p2i --target arm32 -i %s --filetype=asm --args -O2 -asm-verbose \
+; RUN:   | FileCheck --check-prefix=ARM32 %s
+
+define internal i32 @mul(i64 %a, i64 %b, i32 %c) {
+  ; Force an early use of %c.
+  store i32 %c, i32* undef, align 1
+  %m = mul i64 %a, %b
+  %t = trunc i64 %m to i32
+  ; Make many uses of %c to give it high weight.
+  %t1 = add i32 %t, %c
+  %t2 = add i32 %t1, %c
+  %t3 = add i32 %t2, %c
+  ret i32 %t3
+}
+
+; For x8632, we want asm-verbose to print the stack offset assignment for lv$c
+; ("local variable 'c'") in the prolog, and then have at least one use of lv$c
+; in the body, i.e. don't register-allocate edx to %c.
+
+; X8632-LABEL: mul
+; X8632: lv$c =
+; X8632: lv$c
+
+; For arm32, the failure would manifest as a translation error - no register
+; being allocated to the high operand, so we just check for successful
+; translation.
+
+; ARM32-LABEL: mul