[llvm] [Gisel] smin/umin optimization rewrites (PR #203939)

[Luisa Cicolini via llvm-commits]

================
@@ -156,3 +156,158 @@ body:             |
     %mul:_(s64) = G_MUL %max, %sub2
     $x0 = COPY %mul
     RET_ReallyLR implicit $x0
+
+...
+---
+name:   ASubSMin
+body:             |
+  bb.0:
+    liveins: $x0
+
+
+    ; CHECK-LABEL: name: ASubSMin
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %a:_(s64) = COPY $x0
+    ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+    ; CHECK-NEXT: [[SUB:%[0-9]+]]:_(s64) = G_SUB %zero, %a
+    ; CHECK-NEXT: %sub2:_(s64) = G_SMAX %a, [[SUB]]
+    ; CHECK-NEXT: $x0 = COPY %sub2(s64)
+    ; CHECK-NEXT: RET_ReallyLR implicit $x0
+    %a:_(s64) = COPY $x0
+    %zero:_(s64) = G_CONSTANT i64 0
+    %sub1:_(s64) = G_SUB %zero, %a
+    %max:_(s64) = G_SMIN %a, %sub1
+    %sub2:_(s64) = G_SUB %zero, %max
+    $x0 = COPY %sub2
+    RET_ReallyLR implicit $x0
+
+...
+---
+name:   ASubSMinVec
+body:             |
+  bb.0:
+    liveins: $q0
+
+    ; CHECK-LABEL: name: ASubSMinVec
+    ; CHECK: liveins: $q0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %a:_(<4 x s32>) = COPY $q0
+    ; CHECK-NEXT: %c:_(s32) = G_CONSTANT i32 0
+    ; CHECK-NEXT: %c1:_(<4 x s32>) = G_BUILD_VECTOR %c(s32), %c(s32), %c(s32), %c(s32)
+    ; CHECK-NEXT: [[SUB:%[0-9]+]]:_(<4 x s32>) = G_SUB %c1, %a
+    ; CHECK-NEXT: %sub2:_(<4 x s32>) = G_SMAX %a, [[SUB]]
+    ; CHECK-NEXT: $q0 = COPY %sub2(<4 x s32>)
+    ; CHECK-NEXT: RET_ReallyLR implicit $q0
+    %a:_(<4 x s32>) = COPY $q0
+    %c:_(s32) = G_CONSTANT i32 0
+    %c1:_(<4 x s32>) = G_BUILD_VECTOR %c(s32), %c(s32), %c(s32), %c(s32)
+    %sub1:_(<4 x s32>) = G_SUB %c1, %a
+    %max:_(<4 x s32>) = G_SMIN %a, %sub1
+    %sub2:_(<4 x s32>) = G_SUB %c1, %max
+    $q0 = COPY %sub2
+    RET_ReallyLR implicit $q0
+
+...
+---
+name:   ASubSMinOneUse
+body:             |
+  bb.0:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: ASubSMinOneUse
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %a:_(s64) = COPY $x0
+    ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+    ; CHECK-NEXT: %sub1:_(s64) = G_SUB %zero, %a
+    ; CHECK-NEXT: %max:_(s64) = G_SMIN %a, %sub1
+    ; CHECK-NEXT: %sub2:_(s64) = G_SUB %zero, %max
+    ; CHECK-NEXT: %mul:_(s64) = G_MUL %max, %sub2
+    ; CHECK-NEXT: $x0 = COPY %mul(s64)
+    ; CHECK-NEXT: RET_ReallyLR implicit $x0
+    %a:_(s64) = COPY $x0
+    %zero:_(s64) = G_CONSTANT i64 0
+    %sub1:_(s64) = G_SUB %zero, %a
+    %max:_(s64) = G_SMIN %a, %sub1
+    %sub2:_(s64) = G_SUB %zero, %max
+    %mul:_(s64) = G_MUL %max, %sub2
+    $x0 = COPY %mul
+    RET_ReallyLR implicit $x0
+
+...
+---
+name:   ASubUMin
+body:             |
+  bb.0:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: ASubUMin
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %a:_(s64) = COPY $x0
+    ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+    ; CHECK-NEXT: [[SUB:%[0-9]+]]:_(s64) = G_SUB %zero, %a
+    ; CHECK-NEXT: %sub2:_(s64) = G_UMAX %a, [[SUB]]
+    ; CHECK-NEXT: $x0 = COPY %sub2(s64)
+    ; CHECK-NEXT: RET_ReallyLR implicit $x0
+    %a:_(s64) = COPY $x0
+    %zero:_(s64) = G_CONSTANT i64 0
+    %sub1:_(s64) = G_SUB %zero, %a
+    %max:_(s64) = G_UMIN %a, %sub1
+    %sub2:_(s64) = G_SUB %zero, %max
+    $x0 = COPY %sub2
+    RET_ReallyLR implicit $x0
+
+...
+---
+name:   ASubUMinVec
+body:             |
+  bb.0:
+    liveins: $q0
+
+    ; CHECK-LABEL: name: ASubUMinVec
+    ; CHECK: liveins: $q0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %a:_(<4 x s32>) = COPY $q0
+    ; CHECK-NEXT: %c:_(s32) = G_CONSTANT i32 0
+    ; CHECK-NEXT: %c1:_(<4 x s32>) = G_BUILD_VECTOR %c(s32), %c(s32), %c(s32), %c(s32)
+    ; CHECK-NEXT: [[SUB:%[0-9]+]]:_(<4 x s32>) = G_SUB %c1, %a
+    ; CHECK-NEXT: %sub2:_(<4 x s32>) = G_UMAX %a, [[SUB]]
+    ; CHECK-NEXT: $q0 = COPY %sub2(<4 x s32>)
+    ; CHECK-NEXT: RET_ReallyLR implicit $q0
+    %a:_(<4 x s32>) = COPY $q0
+    %c:_(s32) = G_CONSTANT i32 0
+    %c1:_(<4 x s32>) = G_BUILD_VECTOR %c(s32), %c(s32), %c(s32), %c(s32)
+    %sub1:_(<4 x s32>) = G_SUB %c1, %a
+    %max:_(<4 x s32>) = G_UMIN %a, %sub1
+    %sub2:_(<4 x s32>) = G_SUB %c1, %max
+    $q0 = COPY %sub2
+    RET_ReallyLR implicit $q0
+
+...
+---
+name:   ASubUMinOneUse
+body:             |
+  bb.0:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: ASubUMinOneUse
+    ; CHECK: liveins: $x0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %a:_(s64) = COPY $x0
+    ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+    ; CHECK-NEXT: %sub1:_(s64) = G_SUB %zero, %a
+    ; CHECK-NEXT: %max:_(s64) = G_UMIN %a, %sub1
+    ; CHECK-NEXT: %sub2:_(s64) = G_SUB %zero, %max
+    ; CHECK-NEXT: %mul:_(s64) = G_MUL %max, %sub2
+    ; CHECK-NEXT: $x0 = COPY %mul(s64)
+    ; CHECK-NEXT: RET_ReallyLR implicit $x0
+    %a:_(s64) = COPY $x0
+    %zero:_(s64) = G_CONSTANT i64 0
+    %sub1:_(s64) = G_SUB %zero, %a
+    %max:_(s64) = G_UMIN %a, %sub1
+    %sub2:_(s64) = G_SUB %zero, %max
+    %mul:_(s64) = G_MUL %max, %sub2
+    $x0 = COPY %mul
+    RET_ReallyLR implicit $x0
----------------
luisacicolini wrote:

added

https://github.com/llvm/llvm-project/pull/203939