[llvm] [RISCV] Add Zilsd/Zclsd support to RISCVMakeCompressible. (PR #140136)

[Craig Topper via llvm-commits]

================
@@ -0,0 +1,299 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -o - %s -mtriple=riscv32 -mattr=+zilsd,+zclsd,+zdinx -simplify-mir \
+# RUN:   -run-pass=riscv-make-compressible | FileCheck --check-prefixes=RV32 %s
+--- |
+  define void @store_common_value_double(ptr %a, ptr %b, ptr %c, i32 %d, double %e, double %f) #0 {
+  entry:
+    store double %f, ptr %a, align 8
+    store double %f, ptr %b, align 8
+    store double %f, ptr %c, align 8
+    ret void
+  }
+
+  define void @store_common_ptr_double(double %a, double %b, double %d, ptr %p) #0 {
+  entry:
+    store volatile double %a, ptr %p, align 8
+    store volatile double %b, ptr %p, align 8
+    store volatile double %b, ptr %p, align 8
+    ret void
+  }
+
+  define void @load_common_ptr_double(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f, ptr %g) #0 {
+  entry:
+    %0 = load double, ptr %g, align 8
+    %arrayidx1 = getelementptr inbounds { double, double, i32 }, ptr %g, i32 0, i32 1
+    %1 = load double, ptr %arrayidx1, align 8
+    %arrayidx2 = getelementptr inbounds { double, double, i32 }, ptr %g, i32 0, i32 2
+    %2 = load i32, ptr %arrayidx2, align 8
+    tail call void @load_common_ptr_double_1(double %0, double %1, i32 %2)
+    ret void
+  }
+
+  declare void @load_common_ptr_double_1(double, double, double) #0
+
+  define void @store_large_offset_double(ptr %p, i32 %dummy, double %a, double %b, double %c) #0 {
+  entry:
+    %0 = getelementptr inbounds double, ptr %p, i32 100
+    store volatile double %a, ptr %0, align 8
+    %1 = getelementptr inbounds double, ptr %p, i32 101
+    store volatile double %b, ptr %1, align 8
+    %2 = getelementptr inbounds double, ptr %p, i32 102
+    store volatile double %b, ptr %2, align 8
+    ret void
+  }
+
+  define void @load_large_offset_double(i32 %a, i32 %b, i32 %c, i32 %d, ptr %p) #0 {
+  entry:
+    %arrayidx = getelementptr inbounds { [102 x double], i32 }, ptr %p, i32 0, i32 0, i32 100
+    %0 = load double, ptr %arrayidx, align 8
+    %arrayidx1 = getelementptr inbounds { [102 x double], i32 }, ptr %p, i32 0, i32 0, i32 101
+    %1 = load double, ptr %arrayidx1, align 8
+    %arrayidx2 = getelementptr inbounds { [102 x double], i32 }, ptr %p, i32 0, i32 1
+    %2 = load i32, ptr %arrayidx2, align 8
+    tail call void @load_large_offset_double_1(double %0, double %1, i32 %2)
+    ret void
+  }
+
+  declare void @load_large_offset_double_1(double, double) #0
+
+  define void @store_common_value_double_no_opt(ptr %a, i32 %b, double %c, double %d, double %e) #0 {
+  entry:
+    store double %e, ptr %a, align 8
+    ret void
+  }
+
+  define void @store_common_ptr_double_no_opt(double %a, i32 %b, i32 %c, i32 %d, i32 %e, ptr %p) #0 {
+  entry:
+    store volatile double %a, ptr %p, align 8
+    ret void
+  }
+
+  define double @load_common_ptr_double_no_opt(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f, ptr %g) #0 {
+  entry:
+    %0 = load double, ptr %g, align 8
+    ret double %0
+  }
+
+  define void @store_large_offset_double_no_opt(ptr %p, double %a, double %b) #0 {
+  entry:
+    %0 = getelementptr inbounds double, ptr %p, i32 100
+    store volatile double %a, ptr %0, align 8
+    %1 = getelementptr inbounds double, ptr %p, i32 101
+    store volatile double %b, ptr %1, align 8
+    ret void
+  }
+
+  define { double, double } @load_large_offset_double_no_opt(ptr %p) #0 {
+  entry:
+    %arrayidx = getelementptr inbounds double, ptr %p, i32 100
+    %0 = load double, ptr %arrayidx, align 8
+    %arrayidx1 = getelementptr inbounds double, ptr %p, i32 101
+    %1 = load double, ptr %arrayidx1, align 8
+    %2 = insertvalue { double, double } undef, double %0, 0
+    %3 = insertvalue { double, double } %2, double %1, 1
+    ret { double, double } %3
+  }
+
+  attributes #0 = { minsize "target-features"="+zilsd,+zdinx" }
+...
+---
+name:            store_common_value_double
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10, $x11, $x12, $x16, $x17
+
+    ; RV32-LABEL: name: store_common_value_double
+    ; RV32: liveins: $x10, $x11, $x12, $x16, $x17
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: $x14 = ADDI $x16, 0
+    ; RV32-NEXT: $x15 = ADDI $x17, 0
+    ; RV32-NEXT: SD_RV32 $x14_x15, killed renamable $x10, 0 :: (store (s64) into %ir.a)
+    ; RV32-NEXT: SD_RV32 $x14_x15, killed renamable $x11, 0 :: (store (s64) into %ir.b)
+    ; RV32-NEXT: SD_RV32 killed $x14_x15, killed renamable $x12, 0 :: (store (s64) into %ir.c)
+    ; RV32-NEXT: PseudoRET
+    SD_RV32 renamable $x16_x17, killed renamable $x10, 0 :: (store (s64) into %ir.a)
+    SD_RV32 renamable $x16_x17, killed renamable $x11, 0 :: (store (s64) into %ir.b)
+    SD_RV32 killed renamable $x16_x17, killed renamable $x12, 0 :: (store (s64) into %ir.c)
+    PseudoRET
+
+...
+---
+name:            store_common_ptr_double
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x10, $x11, $x12, $x13, $x14, $x15, $x16
+
+    ; RV32-LABEL: name: store_common_ptr_double
+    ; RV32: liveins: $x10, $x11, $x12, $x13, $x14, $x15, $x16
+    ; RV32-NEXT: {{  $}}
+    ; RV32-NEXT: $x14 = ADDI $x16, 0
+    ; RV32-NEXT: SD_RV32 killed renamable $x10_x11, $x14, 0 :: (volatile store (s64) into %ir.p)
+    ; RV32-NEXT: SD_RV32 renamable $x12_x13, $x14, 0 :: (volatile store (s64) into %ir.p)
+    ; RV32-NEXT: SD_RV32 killed renamable $x12_x13, killed $x14, 0 :: (volatile store (s64) into %ir.p)
----------------
topperc wrote:

I stored x12_x13 twice to make sure we had a free register in the 6 compressible caller saved registers.

https://github.com/llvm/llvm-project/pull/140136