[llvm] [GISel][RISCV]Implement indirect parameter passing (PR #95429)

[Sergei Barannikov via llvm-commits]

================
@@ -16,6 +16,580 @@
 ; Check that on RV32, i64 is passed in a pair of registers. Unlike
 ; the convention for varargs, this need not be an aligned pair.
 
+define i64 @callee_128i_in_regs_stack_fst(i64 %x1, i64 %x2, i64 %x3, i64 %x4, i128 %y, i127 %y2) {
+  ; RV32I-LABEL: name: callee_128i_in_regs_stack_fst
+  ; RV32I: bb.1 (%ir-block.0):
+  ; RV32I-NEXT:   liveins: $x10, $x11, $x12, $x13, $x14, $x15, $x16, $x17
+  ; RV32I-NEXT: {{  $}}
+  ; RV32I-NEXT:   [[COPY:%[0-9]+]]:_(s32) = COPY $x10
+  ; RV32I-NEXT:   [[COPY1:%[0-9]+]]:_(s32) = COPY $x11
+  ; RV32I-NEXT:   [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[COPY]](s32), [[COPY1]](s32)
+  ; RV32I-NEXT:   [[COPY2:%[0-9]+]]:_(s32) = COPY $x12
+  ; RV32I-NEXT:   [[COPY3:%[0-9]+]]:_(s32) = COPY $x13
+  ; RV32I-NEXT:   [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[COPY2]](s32), [[COPY3]](s32)
+  ; RV32I-NEXT:   [[COPY4:%[0-9]+]]:_(s32) = COPY $x14
+  ; RV32I-NEXT:   [[COPY5:%[0-9]+]]:_(s32) = COPY $x15
+  ; RV32I-NEXT:   [[MV2:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[COPY4]](s32), [[COPY5]](s32)
+  ; RV32I-NEXT:   [[COPY6:%[0-9]+]]:_(s32) = COPY $x16
+  ; RV32I-NEXT:   [[COPY7:%[0-9]+]]:_(s32) = COPY $x17
+  ; RV32I-NEXT:   [[MV3:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[COPY6]](s32), [[COPY7]](s32)
+  ; RV32I-NEXT:   [[FRAME_INDEX:%[0-9]+]]:_(p0) = G_FRAME_INDEX %fixed-stack.1
+  ; RV32I-NEXT:   [[LOAD:%[0-9]+]]:_(p0) = G_LOAD [[FRAME_INDEX]](p0) :: (load (s32) from %fixed-stack.1, align 16)
+  ; RV32I-NEXT:   [[LOAD1:%[0-9]+]]:_(s128) = G_LOAD [[LOAD]](p0) :: (load (s128) from stack, align 8)
+  ; RV32I-NEXT:   [[FRAME_INDEX1:%[0-9]+]]:_(p0) = G_FRAME_INDEX %fixed-stack.0
+  ; RV32I-NEXT:   [[LOAD2:%[0-9]+]]:_(p0) = G_LOAD [[FRAME_INDEX1]](p0) :: (load (s32) from %fixed-stack.0)
+  ; RV32I-NEXT:   [[LOAD3:%[0-9]+]]:_(s127) = G_LOAD [[LOAD2]](p0) :: (load (s127) from stack + 4, align 4, basealign 8)
----------------
s-barannikov wrote:

Note that memory operands are effectively metadata, they don't encode any operations. I.e. you can't fold G_PTR_ADD into G_LOAD by adding some offset to memory operand. This metadata can be used to quickly determine the effective location of the access without analyzing the pointer operand of the instruction. This is primarily used by alias analysis.


https://github.com/llvm/llvm-project/pull/95429